Tag Archives: computing

Cloud Computing’s Coming Energy Crisis

Post Syndicated from Mark Pesce original https://spectrum.ieee.org/computing/hardware/cloud-computings-coming-energy-crisis

How much of our computing now happens in the cloud? A lot. Providers of public cloud services alone take in more than a quarter of a trillion U.S. dollars a year. That’s why Amazon, Google, and Microsoft maintain massive data centers all around the world. Apple and Facebook, too, run similar facilities, all stuffed with high-core-count CPUs, sporting terabytes of RAM and petabytes of storage.

These machines do the heavy lifting to support what’s been called “surveillance capitalism”: the endless tracking, user profiling, and algorithmic targeting used to distribute advertising. All that computing rakes in a lot of dollars, of course, but it also consumes a lot of watts: Bloomberg recently estimated that about 1 percent of the world’s electricity goes to cloud computing.

That figure is poised to grow exponentially over the next decade. Bloomberg reckons that, globally, we might exit the 2020s needing as much as 8 percent of all electricity to power the future cloud. That might seem like a massive jump, but it’s probably a conservative estimate. After all, by 2030, with hundreds of millions of augmented-reality spectacles streaming real-time video into the cloud, and with the widespread adoption of smart digital currencies seamlessly blending money with code, the cloud will provide the foundation for nearly every financial transaction and user interaction with data.

How much energy can we dedicate to all this computing? In an earlier time, we could have relied on Moore’s Law to keep the power budget in check as we scaled up our computing resources. But now, as we wring out the last bits of efficiency from the final few process nodes before we reach atomic-scale devices, those improvements will hit physical limits. It won’t be long until computing and power consumption will once again be strongly coupled—as they were 60 years ago, before integrated CPUs changed the game.

We seem to be hurtling toward a brick wall, as the rising demand for computing collides with decreasing efficiencies. We can’t devote the whole of the planet’s electricity generation to support the cloud. Something will have to give.

The most immediate solutions will involve processing more data at the edge, before it goes into the cloud. But that only shifts the burden, buying time for rethinking how to manage our computing in the face of limited power resources.

Software and hardware engineering will no doubt reorient their design practices around power efficiency. More code will find its way into custom silicon. And that code will find more reasons to run infrequently, asynchronously, and as minimally as possible. All of that will help, but as software progressively eats more of the world—to borrow a now-famous metaphor—we will confront this challenge in ever-wider realms.

We can already spy one face of this future in the nearly demonic coupling of energy consumption and private profit that provides the proof-of-work mechanism for cryptocurrencies like Bitcoin. Companies like Square have announced investments in solar energy for Bitcoin mining, hoping to deflect some of the bad press associated with this activity. But more than public relations is at stake.

Bitcoin asks us right now to pit the profit motive against the health of the planet. More and more computing activities will do the same in the future. Let’s hope we never get to a point where the fate of the Earth hinges on the fate of the transistor.

This article appears in the August 2021 print issue as “Cloud Computing’s Dark Cloud.”

Google’s Quantum Computer Exponentially Suppresses Errors

Post Syndicated from Charles Q. Choi original https://spectrum.ieee.org/tech-talk/computing/hardware/googles-quantum-computer-exponentially-suppress-errors

In order to develop a practical quantum computer, scientists will have to design ways to deal with any errors that will inevitably pop up in its performance. Now Google has demonstrated that exponential suppression of such errors is possible, experiments that may help pave the way for scalable, fault-tolerant quantum computers.

A quantum computer with enough components known as quantum bits or “qubits” could in theory achieve a “quantum advantage” allowing it to find the answers to problems no classical computer could ever solve.

However, a critical drawback of current quantum computers is the way in which their inner workings are prone to errors. Current state-of-the-art quantum platforms typically have error rates near 10^-3 (or one in a thousand), but many practical applications call for error rates as low as 10^-15.

GPUs Can Now Analyze a Billion Complex Vectors in Record Time

Post Syndicated from Michelle Hampson original https://spectrum.ieee.org/tech-talk/computing/software/gpus-can-now-analyze-millions-of-images-in-record-time

Journal Watch report logo, link to report landing page

The complexity of a digital photo cannot be understated.

Each pixel comprises many data points, and there can be millions of pixels in just a single photo. These many data points in relation to each other are referred to as “high-dimensional” data and can require immense computing power to analyze, say if you were searching for similar photos in a database. Computer programmers and AI experts refer to this as “the curse of high curse of high dimensionality.” 

In a study published July 1 in IEEE Transactions on Big Data, researchers at Facebook AI Research propose a novel solution that aims to ease the burden of this curse. But rather than the traditional means of a computer’s central processing units (CPUs) to analyze high-dimensional media, they’ve harnessed Graphical Processing Units (GPUs). The advancement allows 4 GPUs to analyze more than 95 million high-dimensional images in just 35 minutes. This speed is 8.5 times faster than previous techniques that used GPUs to analyze high-dimensional data.

“The most straightforward technique for searching and indexing [high-dimensional data] is by brute-force comparison, whereby you need to check [each image] against every other image in the database,” explains Jeff Johnson, a research engineer at Facebook AI Research who co-developed the new approach using GPUs. “This is impractical for collections containing billions of vectors.”

CPUs, which have high memory storage and thus can handle large volumes of data, are capable of such a task. However, it takes a substantial amount of time for CPUs to transfer data among the various other supercomputer components, which causes an overall lag in computing time.

In contrast, GPUs offer more raw processing power. Therefore, Johnson and his team developed an algorithm that allows GPUs to both host and analyze a library of vectors. In this way, the data is managed by a small handful of GPUs that do all the work. Notably, GPUs typically have less overall memory storage than CPUs, but Johnson and his colleagues were able to overcome this pitfall using a technique that compresses vector databases and makes them more manageable for the GPUs to analyze.

“By keeping computations purely on a GPU, we can take advantage of the much faster memory available on the accelerator, instead of dealing with the slower memories of CPU servers and even slower machine-to-machine network interconnects within a traditional supercomputer cluster,” explains Johnson.

The researchers tested their approach against a database with one billion vectors, comprising 384 gigabytes of raw data. Their approach reduced the number of vector combinations that need to be analyzed, which would normally be a quintillion (1018), by at least 4 orders of magnitude.

“Both the improvement in speed and the decrease in database size allow for solving problems that would otherwise take hundreds of CPU machines, in effect democratizing large-scale indexing and search techniques using a much smaller amount of hardware,” he says.

Their approach has been made freely available through the Facebook AI Similarity Search (Faiss) open source library. Johnson notes that the computing tech giant Nvidia has already begun building extensions using this approach, which were unveiled at the company’s 2021 GPU Technology Conference.

The Future of Deep Learning Is Photonic

Post Syndicated from Ryan Hamerly original https://spectrum.ieee.org/computing/hardware/the-future-of-deep-learning-is-photonic

Think of the many tasks to which computers are being applied that in the not-so-distant past required human intuition. Computers routinely identify objects in images, transcribe speech, translate between languages, diagnose medical conditions, play complex games, and drive cars.

The technique that has empowered these stunning developments is called deep learning, a term that refers to mathematical models known as artificial neural networks. Deep learning is a subfield of machine learning, a branch of computer science based on fitting complex models to data.

While machine learning has been around a long time, deep learning has taken on a life of its own lately. The reason for that has mostly to do with the increasing amounts of computing power that have become widely available—along with the burgeoning quantities of data that can be easily harvested and used to train neural networks.

The amount of computing power at people’s fingertips started growing in leaps and bounds at the turn of the millennium, when graphical processing units (GPUs) began to be harnessed for nongraphical calculations, a trend that has become increasingly pervasive over the past decade. But the computing demands of deep learning have been rising even faster. This dynamic has spurred engineers to develop electronic hardware accelerators specifically targeted to deep learning, Google’s Tensor Processing Unit (TPU) being a prime example.

Here, I will describe a very different approach to this problem—using optical processors to carry out neural-network calculations with photons instead of electrons. To understand how optics can serve here, you need to know a little bit about how computers currently carry out neural-network calculations. So bear with me as I outline what goes on under the hood.

Almost invariably, artificial neurons are constructed using special software running on digital electronic computers of some sort. That software provides a given neuron with multiple inputs and one output. The state of each neuron depends on the weighted sum of its inputs, to which a nonlinear function, called an activation function, is applied. The result, the output of this neuron, then becomes an input for various other neurons.

For computational efficiency, these neurons are grouped into layers, with neurons connected only to neurons in adjacent layers. The benefit of arranging things that way, as opposed to allowing connections between any two neurons, is that it allows certain mathematical tricks of linear algebra to be used to speed the calculations.

While they are not the whole story, these linear-algebra calculations are the most computationally demanding part of deep learning, particularly as the size of the network grows. This is true for both training (the process of determining what weights to apply to the inputs for each neuron) and for inference (when the neural network is providing the desired results).

What are these mysterious linear-algebra calculations? They aren’t so complicated really. They involve operations on matrices, which are just rectangular arrays of numbers—spreadsheets if you will, minus the descriptive column headers you might find in a typical Excel file.

This is great news because modern computer hardware has been very well optimized for matrix operations, which were the bread and butter of high-performance computing long before deep learning became popular. The relevant matrix calculations for deep learning boil down to a large number of multiply-and-accumulate operations, whereby pairs of numbers are multiplied together and their products are added up.

Over the years, deep learning has required an ever-growing number of these multiply-and-accumulate operations. Consider LeNet, a pioneering deep neural network, designed to do image classification. In 1998 it was shown to outperform other machine techniques for recognizing handwritten letters and numerals. But by 2012 AlexNet, a neural network that crunched through about 1,600 times as many multiply-and-accumulate operations as LeNet, was able to recognize thousands of different types of objects in images.

Advancing from LeNet’s initial success to AlexNet required almost 11 doublings of computing performance. During the 14 years that took, Moore’s law provided much of that increase. The challenge has been to keep this trend going now that Moore’s law is running out of steam. The usual solution is simply to throw more computing resources—along with time, money, and energy—at the problem.

As a result, training today’s large neural networks often has a significant environmental footprint. One 2019 study found, for example, that training a certain deep neural network for natural-language processing produced five times the CO2 emissions typically associated with driving an automobile over its lifetime.

Improvements in digital electronic computers allowed deep learning to blossom, to be sure. But that doesn’t mean that the only way to carry out neural-network calculations is with such machines. Decades ago, when digital computers were still relatively primitive, some engineers tackled difficult calculations using analog computers instead. As digital electronics improved, those analog computers fell by the wayside. But it may be time to pursue that strategy once again, in particular when the analog computations can be done optically.

It has long been known that optical fibers can support much higher data rates than electrical wires. That’s why all long-haul communication lines went optical, starting in the late 1970s. Since then, optical data links have replaced copper wires for shorter and shorter spans, all the way down to rack-to-rack communication in data centers. Optical data communication is faster and uses less power. Optical computing promises the same advantages.

But there is a big difference between communicating data and computing with it. And this is where analog optical approaches hit a roadblock. Conventional computers are based on transistors, which are highly nonlinear circuit elements—meaning that their outputs aren’t just proportional to their inputs, at least when used for computing. Nonlinearity is what lets transistors switch on and off, allowing them to be fashioned into logic gates. This switching is easy to accomplish with electronics, for which nonlinearities are a dime a dozen. But photons follow Maxwell’s equations, which are annoyingly linear, meaning that the output of an optical device is typically proportional to its inputs.

The trick is to use the linearity of optical devices to do the one thing that deep learning relies on most: linear algebra.

To illustrate how that can be done, I’ll describe here a photonic device that, when coupled to some simple analog electronics, can multiply two matrices together. Such multiplication combines the rows of one matrix with the columns of the other. More precisely, it multiplies pairs of numbers from these rows and columns and adds their products together—the multiply-and-accumulate operations I described earlier. My MIT colleagues and I published a paper about how this could be done in 2019. We’re working now to build such an optical matrix multiplier.

The basic computing unit in this device is an optical element called a beam splitter. Although its makeup is in fact more complicated, you can think of it as a half-silvered mirror set at a 45-degree angle. If you send a beam of light into it from the side, the beam splitter will allow half that light to pass straight through it, while the other half is reflected from the angled mirror, causing it to bounce off at 90 degrees from the incoming beam.

Now shine a second beam of light, perpendicular to the first, into this beam splitter so that it impinges on the other side of the angled mirror. Half of this second beam will similarly be transmitted and half reflected at 90 degrees. The two output beams will combine with the two outputs from the first beam. So this beam splitter has two inputs and two outputs.

To use this device for matrix multiplication, you generate two light beams with electric-field intensities that are proportional to the two numbers you want to multiply. Let’s call these field intensities x and y. Shine those two beams into the beam splitter, which will combine these two beams. This particular beam splitter does that in a way that will produce two outputs whose electric fields have values of (x + y)/√2 and (xy)/√2.

In addition to the beam splitter, this analog multiplier requires two simple electronic components—photodetectors—to measure the two output beams. They don’t measure the electric field intensity of those beams, though. They measure the power of a beam, which is proportional to the square of its electric-field intensity.

Why is that relation important? To understand that requires some algebra—but nothing beyond what you learned in high school. Recall that when you square (x + y)/√2 you get (x2 + 2xy + y2)/2. And when you square (xy)/√2, you get (x2 − 2xy + y2)/2. Subtracting the latter from the former gives 2xy.

Pause now to contemplate the significance of this simple bit of math. It means that if you encode a number as a beam of light of a certain intensity and another number as a beam of another intensity, send them through such a beam splitter, measure the two outputs with photodetectors, and negate one of the resulting electrical signals before summing them together, you will have a signal proportional to the product of your two numbers.

My description has made it sound as though each of these light beams must be held steady. In fact, you can briefly pulse the light in the two input beams and measure the output pulse. Better yet, you can feed the output signal into a capacitor, which will then accumulate charge for as long as the pulse lasts. Then you can pulse the inputs again for the same duration, this time encoding two new numbers to be multiplied together. Their product adds some more charge to the capacitor. You can repeat this process as many times as you like, each time carrying out another multiply-and-accumulate operation.

Using pulsed light in this way allows you to perform many such operations in rapid-fire sequence. The most energy-intensive part of all this is reading the voltage on that capacitor, which requires an analog-to-digital converter. But you don’t have to do that after each pulse—you can wait until the end of a sequence of, say, N pulses. That means that the device can perform N multiply-and-accumulate operations using the same amount of energy to read the answer whether N is small or large. Here, N corresponds to the number of neurons per layer in your neural network, which can easily number in the thousands. So this strategy uses very little energy.

Sometimes you can save energy on the input side of things, too. That’s because the same value is often used as an input to multiple neurons. Rather than that number being converted into light multiple times—consuming energy each time—it can be transformed just once, and the light beam that is created can be split into many channels. In this way, the energy cost of input conversion is amortized over many operations.

Splitting one beam into many channels requires nothing more complicated than a lens, but lenses can be tricky to put onto a chip. So the device we are developing to perform neural-network calculations optically may well end up being a hybrid that combines highly integrated photonic chips with separate optical elements.

I’ve outlined here the strategy my colleagues and I have been pursuing, but there are other ways to skin an optical cat. Another promising scheme is based on something called a Mach-Zehnder interferometer, which combines two beam splitters and two fully reflecting mirrors. It, too, can be used to carry out matrix multiplication optically. Two MIT-based startups, Lightmatter and Lightelligence, are developing optical neural-network accelerators based on this approach. Lightmatter has already built a prototype that uses an optical chip it has fabricated. And the company expects to begin selling an optical accelerator board that uses that chip later this year.

Another startup using optics for computing is Optalysis, which hopes to revive a rather old concept. One of the first uses of optical computing back in the 1960s was for the processing of synthetic-aperture radar data. A key part of the challenge was to apply to the measured data a mathematical operation called the Fourier transform. Digital computers of the time struggled with such things. Even now, applying the Fourier transform to large amounts of data can be computationally intensive. But a Fourier transform can be carried out optically with nothing more complicated than a lens, which for some years was how engineers processed synthetic-aperture data. Optalysis hopes to bring this approach up to date and apply it more widely.

There is also a company called Luminous, spun out of Princeton University, which is working to create spiking neural networks based on something it calls a laser neuron. Spiking neural networks more closely mimic how biological neural networks work and, like our own brains, are able to compute using very little energy. Luminous’s hardware is still in the early phase of development, but the promise of combining two energy-saving approaches—spiking and optics—is quite exciting.

There are, of course, still many technical challenges to be overcome. One is to improve the accuracy and dynamic range of the analog optical calculations, which are nowhere near as good as what can be achieved with digital electronics. That’s because these optical processors suffer from various sources of noise and because the digital-to-analog and analog-to-digital converters used to get the data in and out are of limited accuracy. Indeed, it’s difficult to imagine an optical neural network operating with more than 8 to 10 bits of precision. While 8-bit electronic deep-learning hardware exists (the Google TPU is a good example), this industry demands higher precision, especially for neural-network training.

There is also the difficulty integrating optical components onto a chip. Because those components are tens of micrometers in size, they can’t be packed nearly as tightly as transistors, so the required chip area adds up quickly. A 2017 demonstration of this approach by MIT researchers involved a chip that was 1.5 millimeters on a side. Even the biggest chips are no larger than several square centimeters, which places limits on the sizes of matrices that can be processed in parallel this way.

There are many additional questions on the computer-architecture side that photonics researchers tend to sweep under the rug. What’s clear though is that, at least theoretically, photonics has the potential to accelerate deep learning by several orders of magnitude.

Based on the technology that’s currently available for the various components (optical modulators, detectors, amplifiers, analog-to-digital converters), it’s reasonable to think that the energy efficiency of neural-network calculations could be made 1,000 times better than today’s electronic processors. Making more aggressive assumptions about emerging optical technology, that factor might be as large as a million. And because electronic processors are power-limited, these improvements in energy efficiency will likely translate into corresponding improvements in speed.

Many of the concepts in analog optical computing are decades old. Some even predate silicon computers. Schemes for optical matrix multiplication, and even for optical neural networks, were first demonstrated in the 1970s. But this approach didn’t catch on. Will this time be different? Possibly, for three reasons.

First, deep learning is genuinely useful now, not just an academic curiosity. Second, we can’t rely on Moore’s Law alone to continue improving electronics. And finally, we have a new technology that was not available to earlier generations: integrated photonics. These factors suggest that optical neural networks will arrive for real this time—and the future of such computations may indeed be photonic.

About the Author

Ryan Hamerly is a senior scientist at NTT Research and a visiting scientist at MIT’s Quantum Photonics Laboratory.

This Quantum Computer is Sized For Server Rooms

Post Syndicated from Charles Q. Choi original https://spectrum.ieee.org/tech-talk/computing/hardware/iontrap-server

Usually quantum computers are devices that can fill entire laboratories. Now a compact prototype quantum computer can fit in two 19-inch server racks like those found in data centers throughout the world, which could help improve battery designs or crack codes, a new study finds.

A quantum computer with enough components known as quantum bits or “qubits” could in theory achieve a “quantum advantage” enabling it to find the answers to problems no classical computer could ever solve.

However, given how the most advanced current quantum computers typically only have a few dozen qubits at most yet fill up whole rooms. As such, scaling up to hundreds of qubits for practical applications is often thought to prove a challenging task.

Raj Reddy Bets on Babel Fish, Gordon Bell Says No Way

Post Syndicated from Tekla S. Perry original https://spectrum.ieee.org/view-from-the-valley/computing/hardware/raj-reddy-bets-on-babel-fish-gordon-bell-says-no-way

Some three decades ago, a group of pioneering computer scientists—Gordon Bell, John Hennessy, Ed Lazowska, Raj Reddy, Andy Van Dam, and sometimes a few others—began a tradition of placing bets against each other. 

This week, these long-time friends and competitors gathered virtually at an event held by the Computer History Museum in honor of Reddy being named a museum Fellow. In keeping with tradition, they made tech bets; this time for $2000 (years ago, their wagers had been for as much as $1000), to be paid by the losers as a donation to the museum.

Reddy predicted: “We will have a digital Babel fish [a universal translator, as described in The Hitchhiker’s Guide to the Galaxy] that will hide in your ear and translate all the world’s languages—in ten years,” he said. “Anybody will be able to watch any movie and talk to anyone in any language.” 

Or at least, he conceded, it will work for the 100 most common languages.

The names in this gaggle of wager-making technology enthusiasts are no doubt familiar to you: Bell architected the PDP-4 and PDP-6 and oversaw the development of the VAX at Digital Equipment Corp.—and cofounded the Computer History Museum. Hennessy helped write the book on RISC architectures, founded MIPS, served as president of Stanford University, and is chairman of the board of Alphabet. Lazowska, a pioneer in object-oriented computing, helped write the classic text on quantifying computer system performance. AI pioneer Reddy’s focus on speech recognition systems started computing research down the path that led to Siri and today’s other intelligent agents. Van Dam co-designed the first hypertext system and coauthored the key textbook used in the study of computer graphics (the character “Andy” in the movie Toy Story is named after him).

For many years, these five, sometimes with one or two others, sometime minus one or two, assembled at meetings of Microsoft’s Technical Advisory Board. When the talk turned to future technologies, their competitive instincts flared up. Reddy would make a prediction—something concrete, but always a reach in terms of what was currently possible. The others would jump in to bet for and against it. The stakes? Often a dinner, winner’s choice of restaurant, sometimes $1000 cash. Reddy usually lost.

A few of Reddy’s losing predictions, according to a record kept by Bell, who always bet against him:

  • By 1996 video-on-demand would be available in 5 cities of over 0.5 million people and 250,000 people will have access to the service, with a substantial number of users.
  • In 2003 AI would be thought more important than the transistor
  • In 2003 a production car that drives itself will be available for less than 20 percent more than a non-self-driving car
  • By 2002, 10,000 workstations would communicate at a speed of gigabits per second

Which brings us back to Reddy’s new Babel fish wager.  

In fairness, Reddy critiqued his own prediction. “Why I might lose?” he said. Because “technology isn’t enough, you need accessibility and ease of use. It has to be completely unintrusive, like a Babel fish, to fit in our ear, recognize the language and translate it.”

Bell, as usual, took the opposite side, betting against Reddy.

Hennessy supported Reddy’s view. “For once I think Raj could be right,” he said.

But there are “two issues I worry about,” he continued. One is “the last 10 percent problem, though the choice of 100 of the most frequent languages makes it more doable. The other is that if we don’t figure out how to extend Moore’s Law, you’ll put that thing in your ear, and it will burn your head up. One of the things we were wrong about [in the past] was that we didn’t understand how much computing power it would take to do AI.”

Lazowska asked for a clarification, that is, did the envisioned device have to do the translation on board, or could it rely on the user carrying a smart phone. Reddy indicated that it could rely on whatever it wanted, as long as the system doesn’t demand the user’s attention.

“It has to be nonintrusive. I don’t want to have to think about it,” Reddy explained. “If I am talking in Hindi, you can hear me in English in real time.”

With an external gadget permitted in the mix, Lazowska was also willing to bet that Reddy’s prediction will come true.

Van Dam, however, wasn’t convinced.  “There is a spectrum of cockeyed techno-optimism,” he said, “which Raj [Reddy] occupies. I try to be a pragmatist.”

Van Dam’s concerned are the financial incentives, which would have to be significant, he said, to get a device like this to market.

It would have to “nail all 100 languages, manage background noise, [and] be robust,” he said. And, he mused, the user interface will present a big challenge for designers. “I think we will get close but not quite there,” he said, and so, though he loves the quest, he reluctantly bet against it.

Whose bets will pay off? We’ll have to wait ten years to find out.

Cryptocurrency Blockchains Don’t Need To Be Energy Intensive

Post Syndicated from Edd Gent original https://spectrum.ieee.org/computing/networks/cryptocurrency-blockchains-dont-need-to-be-energy-intensive

Blockchain is a generic term for the way most cryptocurrencies record and share their transactions. It’s a type of distributed ledger that parcels up those transactions into chunks called “blocks” and then chains them together cryptographically in a way that makes it incredibly difficult to go back and edit older blocks. How often a new block is made and how much data it contains depends on the implementation. For Bitcoin, that time frame is 10 minutes; for some cryptocurrencies it’s less than a minute.

Unlike most ledgers, which rely on a central authority to update records, blockchains are maintained by a decentralized network of volunteers. The ledger is shared publicly, and the responsibility for validating transactions and updating records is shared by the users. That means blockchains need a simple way for users to reach agreement on changes to the ledger, to ensure everyone’s copy of the ledger looks the same and to prevent fraudulent activity. These are known as consensus mechanisms, and they vary between blockchains.

Blockchain consensus mechanisms decide which user gets to create the next block in the chain, prescribe how other users can verify the block is valid, and ensure users add only genuine transactions through incentives, deterrents, or both. Here we’ll discuss four primary consensus mechanisms.

The granddaddy of all consensus mechanisms—behind Bitcoin, Litecoin, Monero, and (for the time being at least) Ethereum—is called proof of work. Essentially, PoW makes adding transactions to the blockchain computationally—and therefore financially—very expensive, so as to discourage fraudulent activity. At the same time, users who go to the trouble of creating valid blocks, known as mining, are rewarded with cryptocurrency.

The only way miners can game a PoW system is if they control over 51 percent of the blockchain’s mining power, which is almost impossible for a large network like Bitcoin. The downside to PoW is that it requires huge amounts of electricity to power all these computations, which is both inefficient compared with other financial systems and bad for the environment.

Three alternatives to proof of work are being used in other cryptocurrencies and could offer real competition for ­Bitcoin’s PoW (the industry gold standard) in the years ahead.

Each alternative, of course, has its own upsides and downsides. The three consensus mechanisms outlined here—proof of stake, proof of burn, and proof of capacity—each consume far less energy than PoW. But proof of stake (PoS) and proof of burn (PoB), for instance, could lead to a “rich getting richer” scenario because they both reward users who hold lots of their coins. PoS could also encourage hoarding among its holders. As an upside, proof of capacity has a lower cost and less of an environmental impact compared with PoW because memory uses much less energy than processing. On the other side of the coin, PoC invokes the fear that if it becomes popular, it could also lead to massive price inflation of memory chips and nonvolatile storage. That may already be playing out, after the launch of the PoC currency Chia, in March, led memory prices to spike with shortages in some markets. Most important, none of these alternatives have had their security tested at scales comparable with those of Bitcoin.

About the Author

Edd Gent is a freelance science and technology journalist based in Bangalore, India.

This article appears in the July 2021 print issue as “Four Ways to Secure Blockchains.”

For precision, the Sapphire Clock outshines even the best atomic clocks

Post Syndicated from Ewen Levick original https://spectrum.ieee.org/computing/hardware/for-precision-the-sapphire-clock-outshines-even-the-best-atomic-clocks

Tick, tick, tick. The rhythm of clocks is universal. But how exactly alike are those ticks? For some vital applications, even vanishingly small deviations can be a problem.

For those applications, help is on the way, in the form of the most precise clock yet created. Developed by Andre Luiten when he was completing his studies at the University of Western Australia, it’s built around a small, extremely cold crystal of sapphire. Luiten calls it the Cryogenic Sapphire Oscillator, and it could bolster technologies as varied as military radar and quantum computing. He and his colleagues are working on these applications at the University of Adelaide, also in Australia, where he now serves as director of the Institute for Photonics and Advanced Sensings.

The new clock—also known as the Sapphire Clock—isn’t better than an atomic clock; it’s different. That’s because accuracy and precision are different things: Accuracy is how well a clock can measure a true second, now defined as the time it takes cesium atoms under controlled conditions to oscillate between two energy states exactly 9,192,631,770 times. Since 2013, even more accurate types of atomic clocks have been built, but over 400 atomic clocks based on cesium-133 atoms are still used to create civil time across the globe. If you’re reading this article on a smartphone or a laptop, the time displayed on the edge of your screen is derived from one of those atomic clocks.

For many applications, such as satellite-based global positioning systems, accuracy is paramount. And make no mistake, even cesium atomic clocks are stunningly accurate. The NIST-F2 cesium clock operated by the U.S. National Institutes of Standards and Technology in Boulder, Colo., is so accurate that it would have to run for 300 million years to gain or lose a second.

But for some applications, accuracy is less important than precision. Precision has to do not with delineating the perfect second but rather with creating extremely regular ticks, or oscillations. Imagine a game of darts. Atomic clocks are able to land all their darts, or oscillations, broadly around the bull’s-eye so that the average position is right on target, even though any given dart might be a centimeter or two away from dead center. Luiten’s device doesn’t aim for the bull’s-eye: instead, it is able to land all its darts at exactly the same point on the dartboard. In other words, each tick is really, really, really just like another.

To achieve very high precision, Luiten needed to find a material that could sustain electromagnetic oscillations for longer than a beam of cesium atoms can. Another way of putting this is that he needed a crystal with a greater spectral purity, one that would respond only to an exceedingly narrow range of frequencies, almost like a low-loss guitar string that can vibrate for an extremely long time and thus at a very pure frequency.

That turned out to be sapphire, a crystal of aluminum oxide that can be synthesized in the laboratory. When cooled to –267 °C (6 kelvins) and made to oscillate, the symmetry of this type of crystal causes it to lose less energy than almost any other known material. This characteristic makes sapphire an ideal surface on which to propagate electromagnetic radiation. Diamond would work, but it’s costly in large, ultrapure samples. Silicon is cheap, but because it’s a semiconductor it would produce large electrical losses.

“We use a cylindrical chunk of sapphire that’s roughly the same size as the largest natural sapphire that’s ever been found,” Luiten says. “We inject microwaves, and they naturally travel around the circumference of the sapphire.”

The microwaves are injected at the same frequency as the sapphire’s natural resonance, causing them to ripple across the outer surface of the crystal like sound waves traveling along a curved wall. “When you whisper in St Paul’s in London, the sound travels around the entire circumference of the cathedral,” Luiten says. “We’re using the same idea, except only a particular set of frequencies work.”

To match the frequency to the natural resonance of sapphire—the point at which the waves of the “whisper” reinforce after each oscillation—Luiten and his colleagues in Adelaide adjust the temperature to take advantage of impurities in the crystal. “Sapphire is structurally robust, so when subjected to outside forces it still rings at the same frequency,” Luiten notes.

Unfortunately, sapphire’s remarkable properties manifest themselves only near absolute zero. So some method had to be found to keep the crystal supercold. In the early 1990s, when Luiten was doing his Ph.D., he would put the sapphire at the bottom of a giant Thermos flask and fill it up with liquid helium. But the liquid would boil away every six or seven days, and he and his colleagues would have to fill it again.

Luiten decided to install the sapphire in a cryogenic refrigerator, which uses helium gas to keep the crystal cold and stable. However, the gas arrived in high-pressure pulses that caused the temperature to fluctuate and the sapphire to shake, which degraded its ability to keep time. Luiten’s colleague John Hartnett pioneered methods to reduce the vibrations created by the cooling system, using metal-isolation techniques and a small bath of liquid helium instead of gaseous helium.

“The liquid helium allows us to have a good thermal connection between the sapphire and the fridge but prevents vibrations from getting through,” Luiten says.

The Cryogenic Sapphire Oscillator had finally taken shape, and Hartnett’s work was honored in 2010 with IEEE’s W. G. Cady Award. The next challenge was to bring the Sapphire Clock into the outside world. “The oscillator was this crazy scientific tool that could do these amazing tests, but its use was limited to that,” Luiten says.

Luiten and Hartnett spun off the technology into a company called QuantX Labs, which they both now direct. Turns out they were far from done, because the clock had two problems: One, at roughly the same size as a small refrigerator, the clock was too big for many applications. Two, it was expensive, although just how expensive the company won’t say. Despite these problems, there was one organization in Australia with both the need for unrivaled precision and the money to pay for it: the Royal Australian Air Force (RAAF).

To monitor for illegal fishing or other activity off Australia’s vast and barely populated north coast, the RAAF operates an over-the-horizon radar system known as the Jindalee Operational Radar Network (JORN) with support from BAE Systems Australia. JORN uses three transmit-and-receive sites, with each transmitter separated from its receiver by roughly 100 kilometers (62 miles) to prevent interference.

The transmitter stations refract high-frequency signals off the ionosphere, and the receivers listen for echoes created by ships and aircraft. “JORN can see out to 3,000 kilometers,” BAE’s former project lead, Steve Wynd, explains. “But because we’re going up and refracting those transmissions back down, it has a minimum range of about 1,000 kilometers.”

The receiver stations consist of 480 antenna pairs arranged in two parallel lines along the red desert sand, each 3 km long. They rely on the Doppler effect, in which objects moving toward the radar return higher frequency echoes than objects moving away—that is, the signal undergoes a phase shift.

“We propagate signals out, and if the target is moving toward or away from us, then we see a Doppler shift. Over time, we are able to develop target direction and velocity to develop target tracks,” Wynd says.

The signals’ refraction off the ionosphere allows the radar to see over the horizon, but the movement of the ionosphere introduces variations in the signal, as do reflections from the Earth’s surface. The radar cross section of the Earth’s surface can be large, on the order of a million times as great as the cross section of targets. That immensity can make targets hard to identify.

“One of the challenges we have is resolving targets from the background clutter,” Wynd says. “If the clutter is too high, then the signal disappears.”

This is where precise timing really matters. The frequency of the outgoing signal is controlled using the ticks of a reference clock, currently a quartz-based oscillator. If those ticks aren’t very precise, then the outgoing signal becomes irregular, and it’s harder to measure changes in the returning echoes. In addition, if the ticks of the clocks at the transmission and receiver stations get out of sync, then the whole system inaccurately measures the distance to the target.

In both scenarios the radar generates a noisier picture, meaning that smaller or slower targets shift or even become indistinguishable. On the other hand, stable transmission frequencies and better synchronization allow more precise measurements of the phase shift, which means that JORN becomes better at separating targets of interest from the clutter.

According to the Australian military, the Sapphire Clock is a “huge leap,” providing a picture of slow-moving or erratic targets that is three orders of magnitude clearer than what the quartz oscillator can achieve. This is due to the different crystal structure of quartz, which gives rise to a less-well-defined resonance frequency and therefore a lower spectral purity in the output signals of the quartz oscillator. Sapphire is also less sensitive to vibrations and is easier to obtain in ultrapure form than other mineral crystals, such as diamond. Although the low-temperature requirement is a disadvantage in comparison with quartz, the results speak for themselves. “It’s the difference between a plasma TV from 15 years ago versus what you see in an ultra-HD television now,” Wynd insists. “This clock produces a clearer picture.”

The sheer size of JORN’s receiver stations, however, creates another problem. Returning waves come from different angles, causing them to hit the antenna pairs at slightly different times.

“We’ve got a 3-kilometer array that’s physically locked,” Wynd says. “If the target is 30 degrees to the left, the wavefront will hit the left antenna pair slightly earlier than the next one, and so forth.”

To compensate, returning signals are reconstructed using snapshots taken from each antenna pair at the exact moment the wave hits. In effect, the operators electronically steer the radar to face the direction of the echo. The ticks of the Sapphire Clock allow JORN to time each snapshot with greater precision than they ever could with the quartz-based oscillator. “We take readings off each antenna at a slightly different time,” says Wynd. “The greater the precision of that timing source and its distribution, the better the radar can resolve targets.”

The Sapphire Clock’s potential was obvious. The Australian military funded the production of two clock prototypes and flew them to Queensland for a trial. The team then discovered an issue with using the world’s most precise clock: How do you know if it’s working?

Because the clock is three orders of magnitude more precise than any other timepiece, it’s difficult to measure whether it’s working correctly. Fortunately, with two of them, “we could compare one against the other,” Wynd explains.

The clocks remained at the radar site for eight weeks, conducting tests against each other and as part of JORN. Although the technology is not yet a permanent part of the radar, plans are well underway to integrate it into the system. “The science works, but JORN has performance requirements in terms of availability and supportability,” Wynd explains. “That engineering approach is different from engineering a prototype.”

“Subject to satisfactory progress, it is intended that the Cryogenic Sapphire Oscillator be transitioned into JORN,” an Australian military spokesperson said.

Quantum computing is another application for the Sapphire Clock because it, too, requires very precise timing. First, a quick recap of the theory: Traditional computer chips flip electrical currents off and on to create binary bits of information, represented by either a 0 or a 1. Quantum computers, on the other hand, rely on qubits—atomic particles that exist in a complex superposition state, one often described (perhaps simplistically) as being a 0 and a 1 at the same time. The effect is to greatly increase the amount of information that a system of qubits can encode, and thus process. The potential performance of a quantum computer scales exponentially with the number of qubits.

The trouble with qubits, however, is that they are unstable and thus prone to error. If external conditions change—say, because of an imposed electromagnetic field—performance can suffer dramatically. “That degradation is a significant limiter,” explains Michael Biercuk, director of the University of Sydney’s Quantum Control Laboratory and the founder of the startup Q-CTRL. “It’s the Achilles’ heel of the field.”

A lot of effort has gone into creating better hardware to hold the outside world at bay and protect qubits, but it’s not enough. “It’s not just the outside world that can mess you up,” Biercuk explains. “As the quantum hardware has gotten much better, we’ve had to begin worrying about how the master clock used to synchronize all of the devices is performing.”

The master clock’s ticks help synchronize the microwaves that match the natural frequency of the qubits so that the microwaves are tuned to manipulate the qubits. An unstable clock can change the frequency of the microwaves, which can cause errors that are indistinguishable from instability in the qubit itself.

“In order to have a good composite system–the master clock plus the qubit–we need a stable source of microwaves,” Biercuk says. “This is what the Sapphire Clock produces for us.”

The Quantum Control Laboratory purchased a Sapphire Clock in 2018 and is using it to create more-robust and stable quantum computers. Preliminary results show that with the use of the Sapphire Clock, the useful lifetime of qubits has been extended by a factor of nine over that of off-the-shelf alternatives.

“The Sapphire Clock gives a pure starting frequency, which we can modulate to implement quantum logic operations that are robust against other sources of error,” Biercuk says. “Combining this system with an atomic [clock] reference can provide not only an absolute frequency measure but also excellent long-term stability over months and years.”

Should the Sapphire Clock help make quantum computers practical, it would indirectly advance pharmaceutical research and cryptography. Much early-stage pharmaceutical research and development uses computers to simulate or analyze molecules in the context of disease mechanisms. Quantum computers could simulate and compare much larger molecules than traditional computers can. In cryptography, quantum methods could break encryption algorithms that would now take centuries to break, making almost every part of our digital lives vulnerable.

Substantial challenges remain, of course. It would be great, for instance, if researchers could find a way to shrink both the size and the cost of the cryocooler that encases the sapphire. The team is reengineering the device to work at 50 K by increasing the concentration of magnetic impurities in the crystal without introducing additional losses. That’s a temperature that liquid nitrogen can’t quite get to, but it’s way easier than 6 K. It would make the cooler less expensive, less power hungry, and a good deal smaller, too.

The team has submitted a provisional patent for this breakthrough and is already attracting interest from the aviation and telecommunications industries. A major contract is reportedly in the works.

“There’s interest in putting the clock on aeroplanes, and we’re hoping for an opportunity in 5G telecommunications systems,” Luiten says.

If successful, Luiten and his team will be one step closer to climbing and measuring a scientific Everest. The result of their long climb could soon become a common sight, a quiet and unobtrusive machine that tells a remarkable story in a pure, precise language: tick, tick, tick.

About the Author

Ewen Levick is a journalist based in Australia who writes on military technology and travel. He is the editor of Mongolia Weekly, an associate editor for Australian Defence Magazine, and the author of Overland.

This article appears in the July 2021 print issue as “The Most Precise Timekeeper in the World.”

In the Race to Hundreds of Qubits, Photons May Have “Quantum Advantage”

Post Syndicated from Charles Q. Choi original https://spectrum.ieee.org/tech-talk/computing/hardware/race-to-hundreds-of-photonic-qubits-xanadu-scalable-photon

Quantum computers based on photons may have some advantages over  electron-based machines, including operating at room temperature and not temperatures colder than that of deep space. Now, say scientists at quantum computing startup Xanadu, add one more advantage to the photon side of the ledger. Their photonic quantum computer, they say, could scale up to rival or even beat the fastest classical supercomputers—at least at some tasks.

Whereas conventional computers switch transistors either on or off to symbolize data as ones and zeroes, quantum computers use quantum bits or “qubits” that, because of the bizarre nature of quantum physics, can exist in a state known as superposition where they can act as both 1 and 0. This essentially lets each qubit perform multiple calculations at once.

The more qubits are quantum-mechanically connected entangled together, the more calculations they can simultaneously perform. A quantum computer with enough qubits could in theory achieve a “quantum advantage enabling it to grapple with problems no classical computer could ever solve. For instance, a quantum computer with 300 mutually-entangled qubits could theoretically perform more calculations in an instant than there are atoms in the visible universe.

Why Electronic Health Records Haven’t Helped U.S. With Vaccinations

Post Syndicated from Robert N. Charette original https://spectrum.ieee.org/riskfactor/computing/it/ehr-pandemic-promises-not-kept

IEEE COVID-19 coverage logo, link to landing page

Centers for Disease Control and Prevention (CDC) website states that if you have or suspect you have Covid-19, to inform your doctor. However, who do you tell that you actually received your shot?

If you get the shot at a local hospital that is affiliated with your doctor, the vaccination information might show up in your electronic health record (EHR). However, if it doesn’t, or you get the shot through a pharmacy like CVS, your vaccination information may end up stranded on the little paper CDC vaccination card you receive with your shot. The onus will be on you to give your doctor your vaccination information, which will have to be manually (hopefully without error) entered into your EHR.

It was not supposed to be like this.

When President George W. Bush announced in his 2004 State of the Union address that he wanted every U.S. citizen to have an EHR by 2014, one of the motivations was to greatly improve medical care in case of a public health crisis such as a pandemic. The 2003 SARS-Cov-1 pandemic had highlighted the need for the access, interoperability and fusion of health information from a wide variety of governmental and private sources so that the best public health decisions could be made by policy makers in a rapid fashion. As one Chinese doctor remarked in a candid 2004 lessons learned article, in the spring of 2003 SARS had become “out of control due to incorrect information.”

Today, President Bush’s goal that nearly all Americans have access to electronic health records has been met, thanks in large part to the $35 billion in Congressionally-mandated incentives for EHR adoption set out in the 2009 American Recovery and Reinvestment Act. Unfortunately, EHRs have created challenges in combatting the current SARS-Cov-2 pandemic, and at least one doctor is saying that they are proving to be an impediment.

In a recent National Public Radio interview, Dr. Bob Kocher, an adjunct professor at Stanford University School of Medicine and former government healthcare policy official in the Obama Administration, blamed EHRs in part for problems with Americans not being able to easily schedule their vaccinations. A core issue, he states, is that most EHRs are not universally interoperable, meaning health and other relevant patient information they contain is not easily shared either with other EHRs or government public health IT systems. In addition, Kocher notes, key patient demographic information like race and ethnicity may not be captured by an EHR to allow public health officials to understand whether all communities are being equitably vaccinated.

As a result, the information public officials need to determine who has been vaccinated, and where scarce vaccine supplies need to be allocated next, is less accurate and timely than it needs to be. With Johnson & Johnson’s Covid-19 vaccine rolling out this week, the need for precise information increases to ensure the vaccine is allocated to where it is needed most. 

It is a bit ironic that EHRs have not been central in the fight against COVID-19, given how much President Bush was obsessed with preparing the U.S. against a pandemic. Even back in 2007, researchers were outlining how EHRs should be designed to support public health crises and the criticality of seamless data interoperability. However, the requirement for universal EHR interoperability was lost in the 2009 Congressional rush to get hospitals and individual healthcare providers to move from paper medical records to adopt EHRs. Consequently, EHR designs were primarily focused on supporting patient-centric healthcare and especially on aiding healthcare provider billing.

A myriad of other factors has exacerbated the interoperability situation. One is the sheer number of different EHR systems used across the thousands of healthcare providers, each with its own unique way of capturing and storing health information. Even hospitals themselves operate several different EHRs, with the average reportedly having some 16 disparate EHR vendors in use at its affiliated practices. 

Another is the number of different federal, let alone state or local public health IT systems that data needs to flow among. For example, there are six different CDC software systems, including two brand new systems, that are used for vaccine administration and distribution alongside the scores of vaccine registry systems that the states are using. With retail pharmacies now authorized to give COVID-19 vaccinations, the number of IT systems involved keeps growing. To say the process involved in creating timely and accurate information for health policy decision makers is convoluted is being kind.

In addition, the data protocols used to link EHRs to government vaccine registries are a maddeningly tortuous mess that often impede information sharing rather than fostering it. Some EHR vendors like Cerner and Epic have already successfully rolled out improvements to their EHR products to support state mass vaccination efforts, but they are more the exception than the rule.

Furthermore, many EHR vendors (and healthcare providers) have not really been interested in sharing patient information, worrying that it would make it too easy for patients to move to a competitor. While the federal government has recently clamped down on information blocking, it is unlikely to go away anytime soon given the exceptions in the interoperability rules.

Yet another factor has been that state and local public health departments have been cut steadily since the recession of 2008, along with investments in health information systems. Many, if not most, existing state public health IT systems for allocating, scheduling, monitoring and reporting vaccinations have proven not up to the task of supporting the complex health information logistical requirements in a pandemic, and have required significant upgrades or supplemental support. Even then, these supplemental efforts have been fraught with problems. Of course, the federal government’s IT efforts to support state vaccination efforts in the wake of such obstacles has hardly been stellar, either. 

Perhaps the only good news is that the importance of having seamless health IT systems, from health provider EHR systems to state and federal government public health IT systems, is now fully appreciated. Late last year, Congress authorized $500 million under the Coronavirus Aid, Relief, and Economic Security (CARES) Act to the CDC for its Public Health Data Modernization Initiative. The Initiative aims to “bring together state, tribal, local, and territorial (STLT) public health jurisdictions and our private and public sector partners to create modern, interoperable, and real-time public health data and surveillance systems that will protect the American public.” The Office of the National Coordinator for Health Information Technology will also receive $62 million to increase the interoperability of public health systems.

The $560 million will likely be just a small down payment, as many existing EHRs will need to be upgraded not only to better support a future pandemic or other public health crises, but to correct existing issues with EHRs such as poor operational safety and doctor burnout. Furthermore, state and local public health organizations along with their IT infrastructure will also have to be modernized, which will take years. This assumes the states are even willing to invest the funding required given all the other funding priorities caused by the pandemic. 

One last issue that will likely come to the fore is whether it is now time for a national patient identifier that could be used for uniquely identifying patient information. It would make achieving EHR interoperability easier, but creating one has been banned since 1998 over both privacy and security concerns, and more recently, arguments that it isn’t necessary. While the U.S. House of Representatives voted to overturn the ban again last summer, the Senate has not followed suit. If they do decide to overturn the ban and boost EHR interoperability efforts, let us also hope Senators keep in mind just how vulnerable healthcare providers are to cybersecurity attacks.

Quantum Computer Error Correction Is Getting Practical

Post Syndicated from Michael J. Biercuk original https://spectrum.ieee.org/tech-talk/computing/hardware/quantum-computer-error-correction-is-getting-practical

Quantum computers are gaining traction across fields from logistics to finance. But as most users know, they remain research experiments, limited by imperfections in hardware. Today’s machines tend to suffer hardware failures—errors in the underlying quantum information carriers called qubits—in times much shorter than a second. Compare that with the approximately one billion years of continuous operation before a transistor in a conventional computer fails, and it becomes obvious that we have a long way to go.

Companies building quantum computers like IBM and Google have highlighted that their roadmaps include the use of “quantum error correction” to achieve what is known as fault-tolerant quantum computing as they scale to machines with 1,000 or more qubits.

Quantum error correction—or QEC for short—is an algorithm designed to identify and fix errors in quantum computers. It’s able to draw from validated mathematical approaches used to engineer special “radiation hardened” classical microprocessors deployed in space or other extreme environments where errors are much more likely to occur.

QEC is real and has seen many partial demonstrations in laboratories around the world—initial steps that make it clear it’s a viable approach. 2021 may just be the year when it is convincingly demonstrated to give a net benefit in real quantum-computing hardware.

Unfortunately, with corporate roadmaps and complex scientific literature highlighting an increasing number of relevant experiments, an emerging narrative is falsely painting QEC as a future panacea for the life-threatening ills of quantum computing.

QEC, in combination with the theory of fault-tolerant quantum computing, suggests that engineers can in principle build an arbitrarily large quantum computer that if operated correctly would be capable of arbitrarily long computations. This would be a stunningly powerful achievement. The prospect that it can be realized underpins the entire field of quantum computer science: Replace all quantum computing hardware with “logical” qubits running QEC, and even the most complex algorithms come into reach. For instance, Shor’s algorithm could be deployed to render Bitcoin insecure with just a few thousand error-corrected logical qubits. On its face, that doesn’t seem far from the 1,000+ qubit machines promised by 2023. (Spoiler alert: this is the wrong way to interpret these numbers).

The challenge comes when we look at the implementation of QEC in practice. The algorithm by which QEC is performed itself consumes resources—more qubits and many operations.

Returning to the promise of 1,000-qubit machines in industry, so many resources might be required that those 1,000 qubits yield only, say, 5 useful logical qubits.

Even worse, the amount of extra work that must be done to apply QEC currently introduces more error than correction. QEC research has made great strides from the earliest efforts in the late 1990s, introducing mathematical tricks that relax the associated overheads or enable computations on logical qubits to be conducted more easily, without interfering with the computations being performed. And the gains have been enormous, bringing the break-even point, where it’s actually better to perform QEC than not, at least 1,000 times closer than original predictions. Still, the most advanced experimental demonstrations show it’s at least 10 times better to do nothing than to apply QEC in most cases.

This is why a major public-sector research program run by the U.S. intelligence community has spent the last four years seeking to finally cross the break-even point in experimental hardware, for just one logical qubit. We may well unambiguously achieve this goal in 2021—but that’s the beginning of the journey, not the end.

Crossing the break-even point and achieving useful, functioning QEC doesn’t mean we suddenly enter an era with no hardware errors—it just means we’ll have fewer. QEC only totally suppresses errors if we dedicate infinite resources to the process, an obviously untenable proposition. Moreover, even forgetting those theoretical limits, QEC is imperfect and relies on many assumptions about the properties of the errors it’s tasked with correcting. Small deviations from these mathematical models (which happen all the time in real labs) can reduce QEC’s effectiveness further.

Instead of thinking of QEC as a single medicine capable of curing everything that goes wrong in a quantum computer, we should instead consider it an important part of a drug cocktail.

As special as QEC is for abstract quantum computing mathematically, in practice it’s really just a form of what’s known as feedback stabilization. Feedback is the same well-studied technique used to regulate your speed while driving with cruise control or to keep walking robots from tipping over. This realization opens new opportunities to attack the problem of error in quantum computing holistically and may ultimately help us move closer to what we actually want: real quantum computers with far fewer errors.

Fortunately, there are signs that within the research community a view to practicality is emerging. For instance, there is greater emphasis on approximate approaches to QEC that help deal with the most nefarious errors in a particular system, at the expense of being a bit less effective for others.

The combination of hardware-level, open-loop quantum control with feedback-based QEC may also be particularly effective. Quantum control permits a form of “error virtualization” in which the overall properties of the hardware with respect to errors are transformed before implementation of QEC encoding. These include reduced overall error rates, better error uniformity between devices, better hardware stability against slow variations, and a greater compatibility of the error statistics with the assumptions of QEC. Each of these benefits can reduce the resource overheads needed to implement QEC efficiently. Such a holistic view of the problem of error in quantum computing—from quantum control at the hardware level through to algorithmic QEC encoding—can improve net quantum computational performance with fixed hardware resources.

None of this discussion means that QEC is somehow unimportant for quantum computing. And there will always remain a central role for exploratory research into the mathematics of QEC, because you never know what a clever colleague might discover. Still, a drive to practical outcomes might even lead us to totally abandon the abstract notion of fault-tolerant quantum computing and replace it with something more like fault-tolerant-enough quantum computing. That might be just what the doctor ordered.

Michael J. Biercuk, a professor of quantum physics and quantum technology at the University of Sydney, is the founder and CEO of Q-CTRL.

AI Recodes Legacy Software to Operate on Modern Platforms

Post Syndicated from Dexter Johnson original https://spectrum.ieee.org/tech-talk/computing/software/ai-legacy-software-analysis-tool

Last year, IBM demonstrated how AI can perform the tedious job of software maintenance through the updating of legacy code. Now Big Blue has introduced AI-based methods for re-coding old applications so that they can operate on today’s computing platforms.

The latest IBM initiatives, dubbed Mono2Micro and Application Modernization Accelerator (AMA), give app architects new tools for updating legacy applications and extracting new value from them. These initiatives represent a step towards a day when AI could automatically translate a program written in COBOL into Java, according to Nick Fuller, director of hybrid cloud services at IBM Research.

Fuller cautions that these latest AI approaches are currently only capable of breaking the legacy machine code of non-modular monolithic programs into standalone microservices. There still remains another step in translating the programming language because, while the AMA toolkit is in fact designed to modernize COBOL, at this point it only provides an incremental step in the modernization process, according to Fuller. “Language translation is a fundamental challenge for AI that we’re working on to enable some of that legacy code to run in a modern software language,” he added.

In the meantime, IBM’s latest AI tools offer some new capabilities. In the case of Mono2Micro, it first analyzes the old code to reveal all the hidden connections within it that application architects would find extremely difficult and time consuming to uncover on their own, such as the multiple components in the underlying business logic that contain numerous calls and connections to each other. 

Mono2Micro leverages AI clustering techniques to group similar code together, revealing more clearly how groups of code interact. Once Mono2Micro ingests the code, it analyzes the source and object code both statically (analyzing the program before it runs) and dynamically (analyzing the program while it’s running).

The tool then refactors monolithic Java-based programs and their associated business logic and user interfaces into microservices. This refactoring of the monolith into standalone microservices with specific functions minimizes the connections that existed in the software when it was a monolithic program, changing the application’s structure without altering its external behavior.

The objective of the AMA toolkit is to both analyze and refactor legacy applications written in even older languages (COBOL, PL/I). For the AMA toolkit, static analysis of the source code coupled with an understanding of the application structure is used to create a graph that represents the legacy application. When used in conjunction with deep-learning methods, this graph-based approach facilitates data retention as AMA goes through deep-learning processes.

IBM’s AI strategy addresses the key challenges for machine learning when the data input is code and the function is analysis: volume and multiple meanings. Legacy, mission-critical applications are typically hundreds of thousands to millions of lines of code. In this context, applying machine learning (ML) techniques to such large volumes of data can be made more efficient through the concept of embeddings.

These embedding layers represent a way to translate the data into numerical values. The power of embeddings comes from them mapping a large volume of code with multiple possible meanings to numerical values. This is what is done, for example, in translating natural human language to numerical values using “word” embeddings. It is also done in a graph context as it relates to code analysis.

“Embedding layers are tremendous because without them you would struggle to get anything approaching an efficiently performing machine-learning system,” said Fuller.

He added that in the case of code analysis, the ML system gets better in recommending microservices for the refactored legacy application by replicating the application functionality.

Fuller noted: “Once you get to that point, you’re not quite home free, but you’re essentially 70 percent done in terms of what you’re looking to gain, namely a mission critical application that is refactored into a microservices architecture.”

Coding for Qubits: How to Program in Quantum Computer Assembly Language

Post Syndicated from W. Wayt Gibbs original https://spectrum.ieee.org/tech-talk/computing/software/qscout-sandia-open-source-quantum-computer-and-jaqal-quantum-assembly-language

Quantum computing arguably isn’t quite full-fledged computing till there’s quantum software as well as hardware. One open-source quantum computer project at Sandia National Laboratory in Albuquerque, New Mexico aims to address this disparity with a custom-made assembly language for quantum computation. 

Over the next several years, physicist Susan Clark and her team at Sandia plan to use a $25 million, 5-year grant they won from the U.S. Department of Energy to run code provided by academic, commercial, and independent researchers around the world on their “QSCOUT” platform as they steadily upgrade it from 3 qubits today to as many as 32 qubits by 2023. 

QSCOUT stands for the Quantum Scientific Computing Open User Testbed and consists of ionized ytterbium atoms levitating inside a vacuum chamber. Flashes of ultraviolet laser light spin these atoms about, executing algorithms written in the team’s fledgling quantum assembly code—which they’ve named Just Another Quantum Assembly Language or JAQAL. (They’ve in fact trademarked the name as Jaqal with lowercase letters “aqal,” so all subsequent references will use that handle instead.) 

Although Google, IBM, and some other companies have built bigger quantum machines and produced their own programming languages, Clark says that QSCOUT offers some advantages to those keen to explore this frontier of computer science. Superconducting gates, like those in the Google and IBM machines, are certainly fast. But they’re also unstable, losing coherence (and data) in less than a second.

Thanks to ion-trapping technology similar to that developed by the company IonQ (which has published a nice explainer here), Clark says QSCOUT can maintain its computation’s coherence—think of it like a computational equivalent of retaining a train of thought—over as much as 10 seconds. “That’s the best out there,” Clark says. “But our gates are a little slower.” 

The real advantage of QSCOUT is not performance, however, but the ability it gives users to control as much or as little of the computer’s operation as they want to—even adding new or altered operations to the basic instruction set architecture of the machine. “QSCOUT is like a breadboard, while what companies are offering are like printed circuits,” says Andrew Landahl, who leads the QSCOUT software team.

“Our users are scientists who want to do controlled experiments. When they ask for two quantum gates to happen at the same time, they mean it,” he says. Commercial systems tend to optimize users’ programs to improve their performance. “But they don’t give you a lot of details of what’s going on under the hood,” Clark says. In these early days, when it is still so unclear how best to deal with major problems of noise, data persistence, and scalability, there’s a role for a quantum machine that just does what you tell it to do.

To deliver that combination of precision and flexibility, Landahl says, they created Jaqal, which includes commands to initialize the ions as qubits, rotate them individually or together into various states, entangle them into superpositions, and read out their end states as output data. (See “A ‘Hello World’ Program in Jaqal,” below.) 

The first line of any Jaqal program, e.g.

from qscout.v1.std usepulses *

loads a gate pulse file that defines the standard operations (“gates,” in the lingo of quantum computing). This scheme allows for easy extensibility. Landahl says that the next version will add new instructions to support more than 10 qubits and add new functions. Plus, he says, users can even write their own functions, too. 

One addition high on the wish list, Clark says, is a feature taken for granted in classical computing: the ability to do a partial measurement of a computation in progress and to then make adjustments based on the intermediate state. The interconnectedness of qubits makes such partial measurements tricky in the quantum realm, but experimentalists have shown it can be done.

Practical programs will intermix quantum and classical operations, so the QSCOUT team has also released on Github a Python package called JaqalPaq that provides a Jaqal emulator as well as commands to include Jaqal code as an object inside a larger Python program.

Most of the first five project proposals that Sandia accepted from an initial batch of 15 applicants will perform benchmarking of various kinds against other quantum computers. But, Clark says, “One of the teams [led by Phil Richerme at Indiana University, Bloomington] is solving a small quantum chemistry problem by finding the ground states of a particular molecule.”

She says she plans to invite a second round of proposals in March, after the team has upgraded the machine from 3 to 10 qubits.

A “Hello World” Program in Jaqal

One of the simplest non-trivial programs typically run on a new quantum computer, Landahl says, is code that entangles two qubits into one of the so-called Bell states, which are superpositions of the classical 0 and 1 binary states. The Jaqal documentation gives an example of a 15-line program that defines two textbook operations, executes those instructions to prepare a Bell state, and then reads out measurements of the two qubits’ resulting states. 

But as a trapped-ion computer, QSCOUT supports a nifty operation called a Mølmer–Sørensen gate that offers a shortcut. Exploiting that allows the 6-line program below to accomplish the same task—and to repeat it 1024 times:

register q[2]        // Define a 2-qubit register

loop 1024 {          // Sequential statements, repeated 1024x
    prepare_all     // Prepare each qubit in the |0⟩ state
    Sxx q[0] q[1]    // Perform the Mølmer–Sørensen gate
    measure_all     // Measure each qubit and output results

Darpa Hacks Its Secure Hardware, Fends Off Most Attacks

Post Syndicated from Samuel K. Moore original https://spectrum.ieee.org/tech-talk/computing/embedded-systems/darpa-hacks-its-secure-hardware-fends-off-most-attacks

Last summer, Darpa asked hackers to take their best shots at a set of newly designed hardware architectures. After 13,000 hours of hacking by 580 cybersecurity researchers, the results are finally in: just 10 vulnerabilities. Darpa is calling it a win, not because the new hardware fought off every attack, but because it “proved the value of the secure hardware architectures developed under its System Security Integration Through Hardware and Firmware (SSITH) program while pinpointing critical areas to further harden defenses,” says the agency.

Researchers in SSITH, which is part of Darpa’s multibillion dollar Electronics Resurgence Initiative, are now in the third and final phase of developing security architectures and tools that guard systems against common classes of hardware vulnerabilities that can be exploited by malware. [See “How the Spectre and Meltdown Hacks Really Worked.”] The idea is to find a way past the long-standing security model of “patch and pray”, where vulnerabilities are found and software is updated.

In an essay introducing the bug bounty, Keith Rebello, the project’s leader, wrote that patching and praying is a particularly ineffective strategy for IoT hardware, because of the cost and inconsistency of updating and qualifying a hugely diverse set of systems. [See “DARPA: Hack Our Hardware”]

Rebello described the common classes of vulnerabilities as buffer errorsprivilege escalationsresource management attacksinformation leakage attacksnumeric errorscode injection attacks, and cryptographic attacks. SSITH teams came up with RISC-V-based architectures meant to render them impossible. These were then emulated using FPGAs. A full stack of software including a bunch of apps known to be vulnerable ran on the FPGA. They also allowed outsiders to add their own vulnerable applications. The Defense Department then loosed hackers upon the emulated systems using a crowdsourced security platform provided by Synack in a bug bounty effort called Finding Exploits to Thwart Tampering (FETT).

“Knowing that virtually no system is unhackable, we expected to discover bugs within the processors. But FETT really showed us that the SSITH technologies are quite effective at protecting against classes of common software-based hardware exploits,” said Rebello, in a press release. “The majority of the bug reports did not come from exploitation of the vulnerable software applications that we provided to the researchers, but rather from our challenge to the researchers to develop any application with a vulnerability that could be exploited in contradiction with the SSITH processors’ security claims. We’re clearly developing hardware defenses that are raising the bar for attackers.”

Of the 10 vulnerabilities discovered, four were fixed during the bug bounty, which ran from July to October 2020. Seven of those 10 were deemed critical, according to the Common Vulnerability Scoring System 3.0 standards. Most of those resulted from weaknesses introduced by interactions between the hardware, firmware, and the operating system software. For example, one hacker managed to steal the Linux password authentication manager from a protected enclave by hacking the firmware that monitors security, Rebello explains.

In the program’s third and final phase, research teams will work on boosting the performance of their technologies and then fabricating a silicon system-on-chip that implements the security enhancements. They will also take the security tech, which was developed for the open-source RISC-V instruction set architecture, and adapt it to processors with the much more common Arm and x86 instruction set architectures. How long that last part will take depends on the approach the research team took, says Rebelllo. However, he notes that three teams have already ported their architectures to Arm processors in a fraction of the time it took to develop the initial RISC-V version.

Practical Guide to Security in the AWS Cloud, by the SANS Institute

Post Syndicated from IEEE Spectrum Recent Content full text original https://spectrum.ieee.org/whitepaper/practical-guide-to-security-in-the-aws-cloud-by-the-sans-institute

Security in the AWS Cloud

AWS Marketplace would like to present you with a digital copy of the new book, Practical Guide to Security in the AWS Cloud, by the SANS Institute. This complimentary book is a collection of knowledge from 18 contributing authors, who share their tactics, techniques, and procedures for securely operating in the cloud. 

Smart City Video Platform Finds Crimes and Suspects

Post Syndicated from Michelle Hampson original https://spectrum.ieee.org/tech-talk/computing/networks/smart-tracking-platform-anveshak-proves-effective-on-the-streets-of-bangalore

Journal Watch report logo, link to report landing page

In many cities, whenever a car is stolen or someone is abducted, there are video cameras on street corners and in public spaces that could help solve those crimes. However, investigators and police departments typically don’t have the time or resources required to sift through hundreds of video feeds and tease out the evidence they need. 

Aakash Khochare, a doctoral student at the Indian Institute of Science, in Bangalore, has been working for several years on a platform that could be useful. Khochare’s Anveshak, which means “Investigator” in Hindi, won an IEEE TCSC SCALE Challenge 2019 award in 2019.  That annual competition is sponsored by the IEEE Technical Committee on Scalable Computing. Last month, Anveshak was described in detail in a study published in IEEE Transactions on Parallel and Distributed Systems.

Anveshak features a novel tracking algorithm that narrows down the range of cameras likely to have captured video of the object or person of interest. Positive identifications further reduce the number of camera feeds under scrutiny. “This reduces the computational cost of processing the video feeds—from thousands of cameras to often just a few cameras,” explains Khochare.

The algorithm selects which cameras to analyze based on factors such as the local road network, camera location, and the last-seen position of the entity being tracked.

Anveshak also decides how long to buffer a video feed (ie., download a certain amount of data) before analyzing it, which helps reduce delays in computer processing. 

Last, if the computer becomes overburdened with processing data, Anveshak begins to intelligently cut out some of the video frames it deems least likely to be of interest. 

Khochare and his colleagues tested the platform using open dataset images generated by 1,000 virtual cameras covering a 7 square kilometer region across the Indian Institute of Science’s Bangalore campus. They simulated an object of interest moving through the area and used Anveshak to track it within 15 seconds. 

Khochare says Anveshak brings cities closer to a practical and scalable tracking application. However, he says future versions of the program will need to take privacy issues into account.

“Privacy is an important consideration,” he says. “We are working on incorporating privacy restrictions within the platform, for example by allowing analytics to track vehicles, but not people. Or, analytics that track adults but not children. Anonymization and masking of entities who are not of interest or should not be tracked can also be examined.”

Next, he says, the team plans to extend Anveshak so that it is capable of tracking multiple objects at a time. “The use of camera feeds from drones, whose location can be dynamically controlled, is also an emerging area of interest,” says Khochare.

He says the goal is to eventually make Anveshak an open source platform for researchers to use.

Application Note – Advanced Time-Domain Measurements

Post Syndicated from IEEE Spectrum Recent Content full text original https://spectrum.ieee.org/whitepaper/application-note-advanced-timedomain-measurements

Advanced time-domain measurements

Detection of weak signals in the presence of strong disturbers is challenging and requires a high dynamic range. In this application note, we show how high-performance digitizers with built-in FPGA help overcome these challenges using real-time noise suppression techniques such as baseline stabilization, linear filtering, non-linear threshold, and waveform averaging.