Post Syndicated from Stephanie Doyle original https://www.backblaze.com/blog/ai-101-training-vs-inference/

What do Sherlock Holmes and ChatGPT have in common? Inference, my dear Watson!

“We approached the case, you remember, with an absolutely blank mind, which is always an advantage. We had formed no theories. We were simply there to observe and to draw inferences from our observations.”
—Sir Arthur Conan Doyle, The Adventures of the Cardboard Box

As we all continue to refine our thinking around artificial intelligence (AI), it’s useful to define terminology that describes the various stages of building and using AI algorithms—namely, the AI training stage and the AI inference stage. As we see in the quote above, these are not new concepts: they’re based on ideas and methodologies that have been around since before Sherlock Holmes’ time. 

If you’re using AI, building AI, or just curious about AI, it’s important to understand the difference between these two stages so you understand how data moves through an AI workflow. That’s what I’ll explain today.

The TL:DR

The difference between these two terms can be summed up fairly simply: first you train an AI algorithm, then your algorithm uses that training to make inferences from data. To create a whimsical analogy, when an algorithm is training, you can think of it like Watson—still learning how to observe and draw conclusions through inference. Once it’s trained, it’s an inferring machine, a.k.a. Sherlock Holmes. 

Whimsy aside, let’s dig a little deeper into the tech behind AI training and AI inference, the differences between them, and why the distinction is important. 

Obligatory Neural Network Recap

Neural networks have emerged as the brainpower behind AI, and a basic understanding of how they work is foundational when it comes to understanding AI.  

Complex decisions, in theory, can be broken down into a series of yeses and nos, which means that they can be encoded in binary. Neural networks have the ability to combine enough of those smaller decisions, weigh how they affect each other, and then use that information to solve complex problems. And, because more complex decisions require more points of information to come to a final decision, they require more processing power. Neural networks are one of the most widely used approaches to AI and machine learning (ML). 

What Is AI Training?: Understanding Hyperparameters and Parameters

In simple terms, training an AI algorithm is the process through which you take a base algorithm and then teach it how to make the correct decision. This process requires large amounts of data, and can include various degrees of human oversight. How much data you need has a relationship to the number of parameters you set for your algorithm as well as the complexity of a problem. 

We made this handy dandy diagram to show you how data moves through the training process:

And hey—we’re leaving out a lot of nuance in that conversation because dataset size, parameter choice, etc. is a graduate-level topic on its own, and usually is considered proprietary information by the companies who are training an AI algorithm. It suffices to say that dataset size and number of parameters are both significant and have a relationship to each other, though it’s not a direct cause/effect relationship. And, both the number of parameters and the size of the dataset affect things like processing resources—but that conversation is outside of scope for this article (not to mention a hot topic in research). 

As with everything, your use case determines your execution. Some types of tasks actually see excellent results with smaller datasets and more parameters, whereas others require more data and fewer parameters. Bringing it back to the real world, here’s a very cool graph showing how many parameters different AI systems have. Note that they very helpfully identified what type of task each system is designed to solve:

So, let’s talk about what parameters are with an example. Back in our very first AI 101 post, we talked about ways to frame an algorithm in simple terms: 

Machine learning does not specify how much knowledge the bot you’re training starts with—any task can have more or fewer instructions. You could ask your friend to order dinner, or you could ask your friend to order you pasta from your favorite Italian place to be delivered at 7:30 p.m. 

Both of those tasks you just asked your friend to complete are algorithms. The first algorithm requires your friend to make more decisions to execute the task at hand to your satisfaction, and they’ll do that by relying on their past experience of ordering dinner with you—remembering your preferences about restaurants, dishes, cost, and so on. 

The factors that help your friend make a decision about dinner are called hyperparameters and parameters. Hyperparameters are those that frame the algorithm—they are set  outside the training process, but can influence the training of the algorithms. In the example above, a hyperparameter would be how you structure your dinner feedback. Do you thumbs up or down each dish? Do you write a short review? You get the idea. 

Parameters are factors that the algorithm derives through training. In the example above, that’s what time you prefer to eat dinner, which restaurants you enjoy after eating, and so on. 

When you’ve trained a neural network, there will be heavier weights between various nodes. That’s a shorthand of saying that an algorithm will prefer a path it knows is significant, and if you want to really get nerdy with it, this article is well-researched, has a ton of math explainers for various training methods, and includes some fantastic visuals. For our purposes, here’s one way people visualize a “trained” algorithm: 

The “dropout method” is essentially adding weight to the relationships an AI algorithm has found to be significant for the dataset it’s working on. It can then de-prioritize (or sometimes even eliminate) the other relationships. 

Once you have a trained algorithm, then you can use it with a reasonable degree of certainty that it will give you good results, and that leads us to inference. 

What Is AI Inference?

Once you’ve trained your algorithm, you can send it out in the world to do its job (and make yours easier). When you present a trained AI algorithm with a problem and it gives you an answer, that’s called inference. It’s using the way it was trained to draw conclusions or make predictions, depending on how it was built, and once an algorithm is in the “inference stage”, it’s no longer learning (usually). 

Here’s our diagram for how data might move through an inference process: 

As we’ve seen in the real world, this does not mean that the AI model will give you a perfect answer every time—depending on, for example, the AI’s access to new or current information, the complexity of training, etc., the algorithm might give you a better or worse (more or less accurate) answer. But, it will be able to look at a new piece of information or dataset, and return a response. 

When I asked our friend ChatGPT about this, it broke down inference and defined it in relation to the logical underpinnings of different types of algorithms. (As a mere human, how could I resist using inference while writing about inference?) Sure, ChatGPT may be merely mimicking human-like dialogue based on its ingestion of the whole internet, then giving me the most likely answer based on its specialized neural network that predicts the likelihood of the next word in response to my conversational query and without access to data after January 2022, but here’s what our AI friend had to say for itself: 

• Machine Learning Inference: In machine learning, after a model has been trained on a dataset, it is deployed to make predictions or classifications on new, unseen data. During inference, the model takes the input data, processes it, and produces an output or a prediction based on the patterns and relationships it learned during training. For example, a trained machine learning model might be used to classify images, predict stock prices, or identify spam emails.
• Deep Learning Inference: Deep learning, a subfield of machine learning, involves neural networks with multiple layers. In deep learning, inference means feeding an input through a trained deep neural network and obtaining an output. This is commonly used in applications like image recognition, natural language processing, and speech recognition. Inference in deep learning can be computationally intensive, and specialized hardware, such as GPUs and TPUs, are often used to accelerate the process.
• Rule-Based Systems Inference: In rule-based AI systems, inference is the process of applying logical rules to make decisions or draw conclusions. These systems use a set of predefined rules and facts to reason and infer new information. For example, in an expert system for medical diagnosis, inference involves applying a set of rules to patient symptoms and medical knowledge to arrive at a diagnosis.
• Probabilistic Inference: In probabilistic AI, inference is used to estimate probabilities and uncertainty. Bayesian networks and Markov decision processes are examples of probabilistic models that perform inference to compute probabilities of events or make decisions based on probabilities.

You’ll notice that each of these are saying basically the same thing: the AI algorithm applies its decision-making paradigm to a problem. 

Why Stop Learning During the Inference Stage?

In general, it’s important to keep these two stages—training and inference—of an AI algorithm separate for a few reasons: 

• Efficiency: Training is typically a computationally intensive process, whereas inference is usually faster and less resource-intensive. Separating them allows for efficient use of computational resources.
• Generalization: The model’s ability to generalize from training data to unseen data is a key feature. It should not learn from every new piece of data it encounters during inference to maintain this generalization ability.
• Reproducibility: When using trained models in production or applications, it’s important to have consistency and reproducibility in the results. If models were allowed to learn during inference, it would introduce variability and unpredictability in their behavior.

There are some specialized AI algorithms that want to continue learning during the inference stage—your Netflix algorithm is a good example, as are self-driving cars, or dynamic pricing models used to set airfare pricing. On the other hand, the majority of problems we’re trying to solve with AI algorithms deliver better decisions by separating these two phases—think of things like image recognition, language translation, or medical diagnosis, for example.

Training vs. Inference (But, Really: Training Then Inference)

To recap: the AI training stage is when you feed data into your learning algorithm to produce a model, and the AI inference stage is when your  algorithm uses that training to make inferences from data. Here’s a chart for quick reference: 

The difference may seem inconsequential at first glance, but defining these two stages helps to show implications for AI adoption particularly with businesses. That is, given that it’s much less resource intensive (and therefore, less expensive), it’s likely to be much easier for businesses to integrate already-trained AI algorithms with their existing systems. 

And, as always, we’re big believers in demystifying terminology for discussion purposes. Let us know what you think in the comments, and feel free to let us know what you’re interested in learning about next.

The post AI 101: Training vs. Inference appeared first on Backblaze Blog | Cloud Storage & Cloud Backup.

Post Syndicated from Molly Clancy original https://www.backblaze.com/blog/anything-as-a-service-all-the-as-a-service-acronyms-you-didnt-know-you-needed/

Have you ever felt like you need a dictionary just to understand what tech-savvy folks are talking about? Well, you’re in luck, because we’re about to decode some of the most common jargon of the digital age, one acronym at a time. Welcome to the world of “as a Service” acronyms, where we take the humble alphabet and turn it into a digital buffet. 

So, whether you’re SaaS-savvy or PaaS-puzzled, or just someone desperately searching for a little HaaS (Humor as a Service …yeah, we made that one up), you’ve come to the right place. Let’s take a big slurp from this alphabet soup of tech terms.

The One That Started It All: SaaS

SaaS stands for software as a service, and it’s the founding member of the “as a service” nomenclature. (Though, very confusingly, there’s also Sales as a Service—it’s just not shortened to SaaS. Usually.)

Imagine your software as a pizza delivery service. You don’t need to buy all the ingredients, knead the dough, and bake it yourself. Instead, you simply order a slice, and it magically appears on your table (a.k.a. screen). SaaS products are like that, but instead of pizza they serve up everything from messaging to video conferencing to email marketing to …well, really you name it. Which brings us to…

The Kind of Ironic One: XaaS

XaaS stands for, variously, “everything” or “anything” as a service. No one is really sure about the term’s provenance, but it’s a fair guess to say it came into existence when, well, everything started to become a service, probably sometime around the mid-2010s. The thinking is: if it exists in the digital realm, you can probably get it “as a service.” 

The Hardware Related Ones: HaaS, IaaS, and PaaS

HaaS (Hardware as a Service): Instead of purchasing hardware yourself, like computers, servers, networking equipment, and other physical infrastructure components, with HaaS, you can lease or rent the equipment for a given period. It would be like renting a pizza kitchen to make your specialty pies specifically for your sister’s wedding or your grandma’s birthday.

IaaS (Infrastructure as a Service): Infrastructure as a service is kind of like hardware as a service, but it comes with some additional goodies thrown in. Instead of renting just the kitchen, you rent the whole restaurant, chair, tables, and servers (no pun intended) included. IaaS delivers virtualized computing resources, like virtual machines, storage (that’s us!), and networking, over the internet.

PaaS (Platform as a Service): Think of PaaS as a step even further than IaaS—you’re not just renting a pizza restaurant, you’re renting a test kitchen where you can develop your award-winning pie. PaaS provides developers the ability to build, manage, and deploy applications with services like development frameworks, databases, and infrastructure management. It’s the ultimate DIY platform for tech enthusiasts.

The Bad One: RaaS

RaaS stands for Ransomware as a Service, and this is one “as a service” variant you don’t want to mess with. Basically, cybercriminals can purchase ransomware just as easily as you would purchase any app on the app store (it’s probably more complicated than that, but you get the general gist). This makes it easy for even the least savvy cybercriminal to get into the ransomware game. Not great. 

The Ones That Help With the Last One: BaaS and DRaaS

BaaS (Backup as a Service): Backup as a Service is a cloud-based data protection solution that allows individuals and organizations to back up their data to a remote cloud. (Hey! That’s us too!) Instead of managing on-premises backup infrastructure, users can securely store their data off-site, often on highly redundant and geographically distributed servers.

DRaaS (Disaster Recovery as a Service): DRaaS stands for disaster recovery as a service, and it’s the antidote to RaaS. Of course, you need good backups to begin with, but adding DRaaS allows businesses to ensure specific recovery time objectives (RTOs, FYI) so they can get back up and running in the event they’re attacked by ransomware or there’s a natural disaster at your primary storage location. DRaaS solutions used to be made almost exclusively with the large enterprise in mind, but today, it’s possible to architect a DRaaS solution for your business affordably and easily.

The Analytical One: DaaS

DaaS stands for data as a service, and it’s your data’s personal chauffeur. It fetches the information you need and serves it up on a silver platter. DaaS offers data on-demand, making structured data accessible to users over the internet. It simplifies data sharing and access, often in real-time, without the need for complex data management.

The Development-Focused Ones: CaaS, BaaS (again), and FaaS

CaaS (Containers as a Service): CaaS simplifies the deployment, scaling, and orchestration of containerized applications. It’s the tech version of a literal container ship. The individual containers “ship” individual pieces of software, and a CaaS tool helps carry all of those individual containers. Check out container management software Docker’s logo for a visualization:

BaaS (Backend as a Service): It wouldn’t be the first time an acronym has two meanings. BaaS, in this context, provides a backend infrastructure for mobile and web app developers, offering services like databases, user authentication, and APIs. Imagine your own team of digital butlers tending to the back end of your apps. They handle all the behind-the-scenes stuff, so you can focus on making your app shine. 

FaaS (Function as a Service): FaaS is a serverless computing model where developers focus on writing and deploying individual functions or code snippets. These functions run in response to specific events, promoting scalability and efficiency in application development. It’s like having a team of tiny, code-savvy robots doing your bidding.

Go Forth and Abbreviate

Now that you’ve sampled all of the flavors the vast “as a service” world has to offer, we hope you’ve gained a clearer understanding of these sometimes confounding terms. So whether you’re a business professional navigating the cloud or just curious about the tech world, you can wield these acronyms with confidence. 

Did we miss any? I’m sure. Let us know in the comments.

The post Anything as a Service: All the “as a Service” Acronyms You Didn’t Know You Needed appeared first on Backblaze Blog | Cloud Storage & Cloud Backup.

Post Syndicated from Pat Patterson original https://www.backblaze.com/blog/2023-performance-improvements/

You don’t always need the absolute fastest cloud storage—your performance requirements depend on your use case, business objectives, and security needs. But still, faster is usually better. And Backblaze just announced innovation on B2 Cloud Storage that delivers a lot more speed: most file uploads will now be up to 30% faster than AWS S3. 

Today, I’m diving into all of the details of this performance improvement, how we did it, and what it means for you.

The TL:DR

The Results: Customers who rely on small file uploads (1MB or less) can expect to see 10–30% faster uploads on average based on our tests, all without any change to durability, availability, or pricing. 

What Does This Mean for You? 

All B2 Cloud Storage customers will benefit from these performance enhancements, especially those who use Backblaze B2 as a storage destination for data protection software. Small uploads of 1MB or less make up about 70% of all uploads to B2 Cloud Storage and are common for backup and archive workflows. Specific benefits of the performance upgrades include:

• Secures data in offsite backup faster.
• Frees up time for IT administrators to work on other projects.
• Decreases congestion on network bandwidth.
• Deduplicates data more efficiently.

Veeam® is dedicated to working alongside our partners to innovate and create a united front against cyber threats and attacks. The new performance improvements released by Backblaze for B2 Cloud Storage furthers our mission to provide radical resilience to our joint customers.

—Andreas Neufert, Vice President, Product Management, Alliances, Veeam

When Can I Expect Faster Uploads?

Today. The performance upgrades have been fully rolled out across Backblaze’s global data regions.

How We Did It

Prior to this work, when a customer uploaded a file to Backblaze B2, the data was written to multiple hard disk drives (HDDs). Those operations had to be completed before returning a response to the client. Now, we write the incoming data to the same HDDs and also, simultaneously, to a pool of solid state drives (SSDs) we call a “shard stash,” waiting only for the HDD writes to make it to the filesystems’ in-memory caches and the SSD writes to complete before returning a response. Once the writes to HDD are complete, we free up the space from the SSDs so it can be reused.

Since writing data to an SSD is much faster than writing to HDDs, the net result is faster uploads. 

That’s just a brief summary; if you’re interested in the technical details (as well as the results of some rigorous testing), read on!

The Path to Performance Upgrades

As you might recall from many Drive Stats blog posts and webinars, Backblaze stores all customer data on HDDs, affectionately termed ‘spinning rust’ by some. We’ve historically reserved SSDs for Storage Pod (storage server) boot drives. 

Until now. 

That’s right—SSDs have entered the data storage chat. To achieve these performance improvements, we combined the performance of SSDs with the cost efficiency of HDDs. First, I’ll dig into a bit of history to add some context to how we went about the upgrades.

HDD vs. SSD

IBM shipped the first hard drive way back in 1957, so it’s fair to say that the HDD is a mature technology. Drive capacity and data rates have steadily increased over the decades while cost per byte has fallen dramatically. That first hard drive, the IBM RAMAC 350, had a total capacity of 3.75MB, and cost $34,500. Adjusting for inflation, that’s about $375,000, equating to $100,000 per MB, or $100 billion per TB, in 2023 dollars.

Today, the 16TB version of the Seagate Exos X16—an HDD widely deployed in the Backblaze B2 Storage Cloud—retails for around $260, $16.25 per TB. If it had the same cost per byte as the IBM RAMAC 250, it would sell for $1.6 trillion—around the current GDP of China!

SSDs, by contrast, have only been around since 1991, when SanDisk’s 20MB drive shipped in IBM ThinkPad laptops for an OEM price of about $1,000. Let’s consider a modern SSD: the 3.2TB Micron 7450 MAX. Retailing at around $360, the Micron SSD is priced at $112.50 per TB, nearly seven times as much as the Seagate HDD.

So, HDDs easily beat SSDs in terms of storage cost, but what about performance? Here are the numbers from the manufacturers’ data sheets:

The SSD is over 20 times as fast at sustained data transfer than the HDD, but look at the difference in random transfer rates! Even when the HDD is at its fastest, transferring blocks from the outer edge of the disk, the SSD is over 2,200 times faster reading data and nearly 900 times faster for writes.

This massive difference is due to the fact that, when reading data from random locations on the disk, the platters have to complete an average of 0.5 revolutions between blocks. At 7,200 rotations per minute (RPM), that means that the HDD spends about 4.2ms just spinning to the next block before it can even transfer data. In contrast, the SSD’s data sheet quotes its latency as just 80µs (that’s 0.08ms) for reads and 15µs (0.015ms) for writes, between 84 and 280 times faster than the spinning disk.

Let’s consider a real-world operation, say, writing 64kB of data. Assuming the HDD can write that data to sequential disk sectors, it will spin for an average of 4.2ms, then spend 0.25ms writing the data to the disk, for a total of 4.5ms. The SSD, in contrast, can write the data to any location instantaneously, taking just 27µs (0.027ms) to do so. This (somewhat theoretical) 167x speed advantage is the basis for the performance improvement.

Why did I choose a 64kB block? As we mentioned in a recent blog post focusing on cloud storage performance, in general, bigger files are better when it comes to the aggregate time required to upload a dataset. However, there may be other requirements that push for smaller files. Many backup applications split data into fixed size blocks for upload as files to cloud object storage. There is a trade-off in choosing the block size: larger blocks improve backup speed, but smaller blocks reduce the amount of storage required. In practice, backup blocks may be as small as 1MB or even 256kB. The 64kB blocks we used in the calculation above represent the shards that comprise a 1MB file.

The challenge facing our engineers was to take advantage of the speed of solid state storage to accelerate small file uploads without breaking the bank.

Improving Write Performance for Small Files

When a client application uploads a file to the Backblaze B2 Storage Cloud, a coordinator pod splits the file into 16 data shards, creates four additional parity shards, and writes the resulting 20 shards to 20 different HDDs, each in a different Pod.

Each Pod writes the incoming shard to its local filesystem; in practice, this means that the data is written to an in-memory cache and will be written to the physical disk at some point in the near future. Any requests for the file can be satisfied from the cache, but the data hasn’t actually been persistently stored yet.

We need to be absolutely certain that the shards have been written to disk before we return a “success” response to the client, so each Pod executes an fsync system call to transfer (“flush”) the shard data from system memory through the HDD’s write cache to the disk itself before returning its status to the coordinator. When the coordinator has received at least 19 successful responses, it returns a success response to the client. This ensures that, even if the entire data center was to lose power immediately after the upload, the data would be preserved.

As we explained above, for small blocks of data, the vast majority of the time spent writing the data to disk is spent waiting for the drive platter to spin to the correct location. Writing shards to SSD could result in a significant performance gain for small files, but what about that 7x cost difference?

Our engineers came up with a way to have our cake and eat it too by harnessing the speed of SSDs without a massive increase in cost. Now, upon receiving a file of 1MB or less, the coordinator splits it into shards as before, then simultaneously sends the shards to a set of 20 Pods and a separate pool of servers, each populated with 10 of the Micron SSDs described above—a “shard stash.” The shard stash servers easily win the “flush the data to disk” race and return their status to the coordinator in just a few milliseconds. Meanwhile, each HDD Pod writes its shard to the filesystem, queues up a task to flush the shard data to the disk, and returns an acknowledgement to the coordinator.

Once the coordinator has received replies establishing that at least 19 of the 20 Pods have written their shards to the filesystem, and at least 19 of the 20 shards have been flushed to the SSDs, it returns its response to the client. Again, if power was to fail at this point, the data has already been safely written to solid state storage.

We don’t want to leave the data on the SSDs any longer than we have to, so, each Pod, once it’s finished flushing its shard to disk, signals to the shard stash that it can purge its copy of the shard.

Real-World Performance Gains

As I mentioned above, that calculated 167x performance advantage of SSDs over HDDs is somewhat theoretical. In the real world, the time required to upload a file also depends on a number of other factors—proximity to the data center, network speed, and all of the software and hardware between the client application and the storage device, to name a few.

The first Backblaze region to receive the performance upgrade was U.S. East, located in Reston, Virginia. Over a 12-day period following the shard stash deployment there, the average time to upload a 256kB file was 118ms, while a 1MB file clocked in at 137ms. To replicate a typical customer environment, we ran the test application at our partner Vultr’s New Jersey data center, uploading data to Backblaze B2 across the public internet.

For comparison, we ran the same test against Amazon S3’s U.S. East (Northern Virginia) region, a.k.a. us-east-1, from the same machine in New Jersey. On average, uploading a 256kB file to S3 took 157ms, with a 1MB file taking 153ms.

So, comparing the Backblaze B2 U.S. East region to the Amazon S3 equivalent, we benchmarked the new, improved Backblaze B2 as 30% faster than S3 for 256kB files and 10% faster than S3 for 1MB files.

These low-level tests were confirmed when we timed Veeam Backup & Replication software backing up 1TB of virtual machines with 256k block sizes. Backing the server up to Amazon S3 took three hours and 12 minutes; we measured the same backup to Backblaze B2 at just two hours and 15 minutes, 40% faster than S3.

Test Methodology

We wrote a simple Python test app using the AWS SDK for Python (Boto3). Each test run involved timing 100 file uploads using the S3 PutObject API, with a 10ms delay between each upload. (FYI, the delay is not included in the measured time.) The test app used a single HTTPS connection across the test run, following best practice for API usage. We’ve been running the test on a VM in Vultr’s New Jersey region every six hours for the past few weeks against both our U.S. East region and its AWS neighbor. Latency to the Backblaze B2 API endpoint averaged 5.7ms, to the Amazon S3 API endpoint 7.8ms, as measured across 100 ping requests.

What’s Next?

At the time of writing, shard stash servers have been deployed to all of our data centers, across all of our regions. In fact, you might even have noticed small files uploading faster already. It’s important to note that this particular optimization is just one of a series of performance improvements that we’ve implemented, with more to come. It’s safe to say that all of our Backblaze B2 customers will enjoy faster uploads and downloads, no matter their storage workload.

The post How We Achieved Upload Speeds Faster Than AWS S3 appeared first on Backblaze Blog | Cloud Storage & Cloud Backup.

Post Syndicated from Stephanie Doyle original https://www.backblaze.com/blog/ai-101-do-the-dollars-make-sense/

Welcome back to AI 101, a series dedicated to breaking down the realities of artificial intelligence (AI). Previously we’ve defined artificial intelligence, deep learning (DL), and machine learning (ML) and dove into the types of processors that make AI possible. Today we’ll talk about one of the biggest limitations of AI adoption—how much it costs. Experts have already flagged that the significant investment necessary for AI can cause antitrust concerns and that AI is driving up costs in data centers. 

To that end, we’ll talk about: 

• Factors that impact the cost of AI.
• Some real numbers about the cost of AI components. 
• The AI tech stack and some of the industry solutions that have been built to serve it.
• And, uncertainty.

Defining AI: Complexity and Cost Implications

While ChatGPT, DALL-E, and the like may be the most buzz-worthy of recent advancements, AI has already been a part of our daily lives for several years now. In addition to generative AI models, examples include virtual assistants like Siri and Google Home, fraud detection algorithms in banks, facial recognition software, URL threat analysis services, and so on. 

That brings us to the first challenge when it comes to understanding the cost of AI: The type of AI you’re training—and how complex a problem you want it to solve—has a huge impact on the computing resources needed and the cost, both in the training and in the implementation phases. AI tasks are hungry in all ways: they need a lot of processing power, storage capacity, and specialized hardware. As you scale up or down in the complexity of the task you’re doing, there’s a huge range in the types of tools you need and their costs.   

To understand the cost of AI, several other factors come into play as well, including: 

• Latency requirements: How fast does the AI need to make decisions? (e.g. that split second before a self-driving car slams on the brakes.)
• Scope: Is the AI solving broad-based or limited questions? (e.g. the best way to organize this library vs. how many times is the word “cat” in this article.)
• Actual human labor: How much oversight does it need? (e.g. does a human identify the cat in cat photos, or does the AI algorithm identify them?)
• Adding data: When, how, and what quantity new data will need to be ingested to update information over time? 

This is by no means an exhaustive list, but it gives you an idea of the considerations that can affect the kind of AI you’re building and, thus, what it might cost.

The Big Three AI Cost Drivers: Hardware, Storage, and Processing Power

In simple terms, you can break down the cost of running an AI to a few main components: hardware, storage, and processing power. That’s a little bit simplistic, and you’ll see some of these lines blur and expand as we get into the details of each category. But, for our purposes today, this is a good place to start to understand how much it costs to ask a bot to create a squirrel holding a cool guitar.

First Things First: Hardware Costs

Running an AI takes specialized processors that can handle complex processing queries. We’re early in the game when it comes to picking a “winner” for specialized processors, but these days, the most common processor is a graphical processing unit (GPU), with Nvidia’s hardware and platform as an industry favorite and front-runner. 

The most common “workhorse chip” of AI processing tasks, the Nvidia A100, starts at about $10,000 per chip, and a set of eight of the most advanced processing chips can cost about $300,000. When Elon Musk wanted to invest in his generative AI project, he reportedly bought 10,000 GPUs, which equates to an estimated value in the tens of millions of dollars. He’s gone on record as saying that AI chips can be harder to get than drugs. 

Google offers folks the ability to rent their TPUs through the cloud starting at $1.20 per chip hour for on-demand service (less if you commit to a contract). Meanwhile, Intel released a sub-$100 USB stick with a full NPU that can plug into your personal laptop, and folks have created their own models at home with the help of open sourced developer toolkits. Here’s a guide to using them if you want to get in the game yourself. 

Clearly, the spectrum for chips is vast—from under $100 to millions—and the landscape for chip producers is changing often, as is the strategy for monetizing those chips—which leads us to our next section. 

Using Third Parties: Specialized Problems = Specialized Service Providers

Building AI is a challenge with so many moving parts that, in a business use case, you eventually confront the question of whether it’s more efficient to outsource it. It’s true of storage, and it’s definitely true of AI processing. You can already see one way Google answered that question above: create a network populated by their TPUs, then sell access.   

Other companies specialize in broader or narrower parts of the AI creation and processing chain. Just to name a few, diverse companies: there’s Hugging Face, Inflection AI, CoreWeave, and Vultr. Those companies have a wide array of product offerings and resources from open source communities like Hugging Face that provide a menu of models, datasets, no-code tools, and (frankly) rad developer experiments to bare metal servers like Vultr that enhance your compute resources. How resources are offered also exist on a spectrum, including proprietary company resources (i.e. Nvidia’s platform), open source communities (looking at you, Hugging Face), or a mix of the two. 

This means that, whichever piece of the AI tech stack you’re considering, you have a high degree of flexibility when you’re deciding where and how much you want to customize and where and how to implement an out-of-the box solution. 

Ballparking an estimate of what any of that costs would be so dependent on the particular model you want to build and the third-party solutions you choose that it doesn’t make sense to do so here. But, it suffices to say that there’s a pretty narrow field of folks who have the infrastructure capacity, the datasets, and the business need to create their own network. Usually it comes back to any combination of the following: whether you have existing infrastructure to leverage or are building from scratch, if you’re going to sell the solution to others, what control over research or dataset you have or want, how important privacy is and how you’re incorporating it into your products, how fast you need the model to make decisions, and so on. 

Welcome to the Spotlight, Storage

And, hey, with all that, let’s not forget storage. At the most basic level of consideration, AI uses a ton of data. How much? Going knowledge says at least an order of magnitude more examples than the problem presented to train an AI model. That means you want 10 times more examples than parameters. 

That 10x number is just the amount of data you store for the initial training model—clearly the thing learns and grows, because we’re talking about AI. 

Preserving both your initial training algorithm and your datasets can be incredibly useful, too. As we talked about before, the more complex an AI, the higher the likelihood that your model will surprise you. And, as many folks have pointed out, deciding whether to leverage an already-trained model or to build your own doesn’t have to be an either/or—oftentimes the best option is to fine-tune an existing model to your narrower purpose. In both cases, having your original training model stored can help you roll back and identify the changes over time. 

The size of the dataset absolutely affects costs and processing times. The best example is that ChatGPT, everyone’s favorite model, has been rocking GPT-3 (or 3.5) instead of GPT-4 on the general public release because GPT-4, which works from a much larger, updated dataset than GPT-3, is too expensive to release to the wider public. It also returns results much more slowly than GPT-3.5, which means that our current love of instantaneous search results and image generation would need an adjustment. 

And all of that is true because GPT-4 was updated with more information (by volume), more up-to-date information, and the model was given more parameters to take into account for responses. So, it has to both access more data per query and use more complex reasoning to make decisions. That said, it also reportedly has much better results.

Storage and Cost

What are the real numbers to store, say, a primary copy of an AI dataset? Well, it’s hard to estimate, but we can ballpark that, if you’re training a large AI model, you’re going to have at a minimum tens of gigabytes of data and, at a maximum, petabytes. OpenAI considers the size of its training database proprietary information, and we’ve found sources that cite that number as  anywhere from 17GB to 570GB to 45TB of text data. 

That’s not actually a ton of data, and, even taking the highest number, it would only cost $225 per month to store that data in Backblaze B2 (45TB * $5/TB/mo), for argument’s sake. But let’s say you’re training an AI on video to, say, make a robot vacuum that can navigate your room or recognize and identify human movement. Your training dataset could easily reach into petabyte scale (for reference, one petabyte would cost $5,000 per month in Backblaze B2). Some research shows that dataset size is trending up over time, though other folks point out that bigger is not always better.

On the other hand, if you’re the guy with the Intel Neural Compute stick we mentioned above and a Raspberry Pi, you’re talking the cost of the ~$100 AI processor, ~$50 for the Raspberry Pi, and any incidentals. You can choose to add external hard drives, network attached storage (NAS) devices, or even servers as you scale up.

Storage and Speed

Keep in mind that, in the above example, we’re only considering the cost of storing the primary dataset, and that’s not very accurate when thinking about how you’d be using your dataset. You’d also have to consider temporary storage for when you’re actually training the AI as your primary dataset is transformed by your AI algorithm, and nearly always you’re splitting your primary dataset into discrete parts and feeding those to your AI algorithm in stages—so each of those subsets would also be stored separately. And, in addition to needing a lot of storage, where you physically locate that storage makes a huge difference to how quickly tasks can be accomplished. In many cases, the difference is a matter of seconds, but there are some tasks that just can’t handle that delay—think of tasks like self-driving cars. 

For huge data ingest periods such as training, you’re often talking about a compute process that’s assisted by powerful, and often specialized, supercomputers, with repeated passes over the same dataset. Having your data physically close to those supercomputers saves you huge amounts of time, which is pretty incredible when you consider that it breaks down to as little as milliseconds per task.

One way this problem is being solved is via caching, or creating temporary storage on the same chips (or motherboards) as the processor completing the task. Another solution is to keep the whole processing and storage cluster on-premises (at least while training), as you can see in the Microsoft-OpenAI setup or as you’ll often see in universities. And, unsurprisingly, you’ll also see edge computing solutions which endeavor to locate data physically close to the end user. 

While there can be benefits to on-premises or co-located storage, having a way to quickly add more storage (and release it if no longer needed), means cloud storage is a powerful tool for a holistic AI storage architecture—and can help control costs. 

And, as always, effective backup strategies require at least one off-site storage copy, and the easiest way to achieve that is via cloud storage. So, any way you slice it, you’re likely going to have cloud storage touch some part of your AI tech stack. 

What Hardware, Processing, and Storage Have in Common: You Have to Power Them

Here’s the short version: any time you add complex compute + large amounts of data, you’re talking about a ton of money and a ton of power to keep everything running. 

Fortunately for us, other folks have done the work of figuring out how much this all costs. This excellent article from SemiAnalysis goes deep on the total cost of powering searches and running generative AI models. The Washington Post cites Dylan Patel (also of SemiAnalysis) as estimating that a single chat with ChatGPT could cost up to 1,000 times as much as a simple Google search. Those costs include everything we’ve talked about above—the capital expenditures, data storage, and processing. 

Consider this: Google spent several years putting off publicizing a frank accounting of their power usage. When they released numbers in 2011, they said that they use enough electricity to power 200,000 homes. And that was in 2011. There are widely varying claims for how much a single search costs, but even the most conservative say .03 Wh of energy. There are approximately 8.5 billion Google searches per day. (That’s just an incremental cost by the way—as in, how much does a single search cost in extra resources on top of how much the system that powers it costs.) 

Power is a huge cost in operating data centers, even when you’re only talking about pure storage. One of the biggest single expenses that affects power usage is cooling systems. With high-compute workloads, and particularly with GPUs, the amount of work the processor is doing generates a ton more heat—which means more money in cooling costs, and more power consumed. 

So, to Sum Up

When we’re talking about how much an AI costs, it’s not just about any single line item cost. If you decide to build and run your own models on-premises, you’re talking about huge capital expenditure and ongoing costs in data centers with high compute loads. If you want to build and train a model on your own USB stick and personal computer, that’s a different set of cost concerns. 

And, if you’re talking about querying a generative AI from the comfort of your own computer, you’re still using a comparatively high amount of power somewhere down the line. We may spread that power cost across our national and international infrastructures, but it’s important to remember that it’s coming from somewhere—and that the bill comes due, somewhere along the way. 

The post AI 101: Do the Dollars Make Sense? appeared first on Backblaze Blog | Cloud Storage & Cloud Backup.

Post Syndicated from James Flores original https://www.backblaze.com/blog/apis-for-media-and-film-what-you-need-to-know/

Over the years, the film industry has witnessed constant transformation, from the introduction of sound and color to the digital revolution, 4K, and ultra high definition (UHD). However, a groundbreaking change is now underway, as cloud technology merges with media and entertainment (M&E) workflows, reshaping the way content is created, stored, and shared.

What’s helping to drive this transformation? APIs, or application programming interfaces. For any post facility, indie filmmaker/creator, or media team, understanding what APIs are is the first step in using them to embrace the flexibility, efficiency, and speed of the cloud.

From Tape to Digital: A Digital File Revolution

The journey towards the cloud transformation in the M&E industry started with the shift from traditional tape and film to digital formats. This revolutionary transition converted traditional media into digital entities, moving them from workstations to servers, shuttle drives, and shared storage systems. Simultaneously, the proliferation of email and cloud-hosted applications like Gmail, Dropbox, and Office 365 laid the groundwork for a cloud-centric future.

Seamless Collaboration With API-Driven Tools

As time went on, applications began communicating effortlessly with one another, facilitating tasks such as creating calendar invites in Gmail through Zoom and the ability to start Zoom meetings with a command in Slack. These integrations were made possible by APIs that allow applications to interact and share data effectively.

What Are APIs?

APIs are sets of rules and protocols that enable different software applications to communicate and interact with each other, allowing you to access specific functionalities or data from one application to be used in another. APIs facilitate seamless integration between diverse systems, enhancing workflows and promoting interoperability.

Most of us in the film industry are familiar with a GUI, a graphical user interface. It’s how we use applications day in and day out—literally the screens on our programs and computers. But a lot of the tasks we execute via a GUI (like saving files, reading files, and moving files) really are pieces of executable code hidden from us behind a nice button. Think of APIs as another method to execute those same pieces of code, but with code. Code executing code. (Let’s not get into the Skynet complex, and this isn’t AI either.)

Grinding the Gears: A Metaphor for APIs

An easy way to think about APIs is to think of them as gears. Each application has a gear.  If we adjust the two gears to talk we simply align them to each other allowing their APIs to establish communication.

Once communications are established, you can start to do some cool stuff. For example, you can migrate a Frame.io archive to a Backblaze B2 Bucket. Or you could use the iconik API to move a file we want to edit with into our Lucidlink filespace, then remove it as soon as we finish our project. 

Check out a video about the solution here:

The MovieLabs 2030 Vision and Cloud Integration

As the industry embraced cloud technology, the need for standardization became apparent. Organizations like the Institute of Electrical and Electronics Engineers (IEEE) and the Society of Motion Picture and Television Engineers (SMPTE) worked diligently to establish technical unity across vendors and technologies. However, implementation of these standards lacked persistence. To address this void, the Movie Picture Association (MPA) established MovieLabs, an organization dedicated to researching, testing, and developing new guidelines, processes, and tooling to drive improvements on how content is created. One such set of guidelines is the MovieLabs 2030 Vision.

Core Principles of the MovieLabs 2030 Vision

The MovieLabs 2030 Vision outlines 10 core principles that are aspirational for the film industry to accomplish by 2030. These core principles set the stage with a high importance on cloud technology and interoperability. Interoperability boils down to the ability to use various tools but have them share resources—which is where APIs come in. APIs help make tools interoperable and able to share resources. It’s a key functionality, and it’s how many cloud tools work together today.  

The Future Is Here: Cloud Technology at Its Peak

Cloud technology grants us instant access to digital documents and the ability to carry our entire lives in our pockets. With the right tools, our data is securely synced, backed up, and accessible across devices, from smartphones to laptops and even TVs.

Although cloud technology has revolutionized various industries, the media and entertainment sector lagged behind, relying on cumbersome shuttle drives and expensive file systems for our massive files. The COVID pandemic, however, acted as a catalyst for change, pushing the industry to seriously consider the benefits of cloud integration.

Breaking Down Silos With APIs

In a post-pandemic world, many popular media and entertainment applications are built in the cloud, the same as other software as a service (SaaS) applications like Zoom, Slack, or Outlook. Which is great! But many of these tools are designed to best operate in their own ecosystem, meaning once the files are in their systems, it’s not easy to take them out. This may sound familiar if you are an iPhone user faced with migrating to an Android or vice versa. (But who would do that? ) 

With each of these applications working in their own ecosystem, the result is their own dedicated storage and usage costs which can vary greatly across tools. So many productions end up with projects and project files locked in various different environments creating storage silos—the opposite of centralized interoperability. 

APIs not only foster interoperability in cloud-based business applications, but also empower filmmaking cloud tools like Frame.io, iconik, and Backblaze the ability to send, receive, and delete files (the POST, GET, PUT, and DELETE commands) data from other programs, enabling more dynamic and advanced workflows, such as sending files to colorists or reviewing edits for picture lock.

Customized Workflows and Automation

APIs offer the flexibility to tailor workflows to specific needs, whether within a single company or for vendor-specific processes. The automation possibilities are virtually limitless, facilitating seamless integration between cloud tools and storage solutions.

The Road Ahead for Media and Entertainment

The Movie Labs 2030 Vision offers a glimpse into a future defined by cloud tools and automation. Principally, that cloud technology with open and extensible storage exists and is available today. 

So for any post facility, indie filmmaker/creator, or media team still driving around shuttle drives while James Cameron is shooting Avatar in New Zealand and editing it in Santa Monica, the future is here and within reach. You can get started today with all the power and flexibility of the cloud without the Avatar budget.

The post APIs for Media and Film: What You Need to Know appeared first on Backblaze Blog | Cloud Storage & Cloud Backup.

Post Syndicated from Elton Carneiro original https://www.backblaze.com/blog/backblaze-qencode-video-transcoding-made-simple/

If you do any kind of video streaming, encoding and storing your data is one of your main challenges. Encoding videos in various formats and resolutions for different devices and platforms can be a resource-intensive task, and setting up and maintaining on-premises encoding infrastructure can be expensive.

Today, we’re excited to announce an expanded partnership with Qencode, a media services platform that enables users to build powerful video solutions, including solutions to the challenges of transcoding, live streaming, and media storage. The expanded partnership embeds the Backblaze Partner API within the Qencode platform, making it frictionless for users to add cloud storage to their media production workflows. 

What Is Qencode?

Qencode is a media services platform founded in 2017 that assists with digital video transformation. The Qencode API provides developers within the over-the-top (OTT), broadcasting, and media & entertainment sectors with scalable and robust APIs for:

• Video transcoding
• Live streaming
• Content delivery
• Media storage
• Artificial intelligence

Qencode + Backblaze

Recognizing the growing demand for integrated and efficient cloud storage within media production, Qencode and Backblaze built an alliance which creates a new paradigm for cutting-edge video APIs fortified by a reliable and efficient cloud storage solution. This integration empowers flexible workflows consisting of uploading, transcoding, storing, and delivering video content for media and OTT companies of all sizes. By integrating the platforms, this partnership provides top-tier features while simplifying the complexities and reducing the risks often associated with innovation.

We want to set new standards for value in an industry that is fragmented and complex. By merging Qencode’s advanced video processing capabilities with Backblaze’s reliable cloud storage, we’re addressing a critical industry need for seamless integration and efficiency. Integrating Backblaze’s Partner API takes our platform to the next level, providing users with a single, streamlined interface for all their video and media needs.

Murad Mordukhay, CEO of Qencode

Qencode + Backblaze Use Cases

The easy-to-use interface and affordability make Qencode an ideal choice for businesses who need video processing at scale without compromising spend or flexibility. Qencode enables businesses of all sizes to customize and control a complete end-to-end solution, from sign-on to billing, which includes seamless access to Backblaze storage through the Qencode software as a service (SaaS) platform. 

Simplifying the User Experience

Expanding this partnership with Qencode takes our API technology a step further in making cloud storage more accessible to businesses whose mission is to simplify user experience. We are excited to work with a specialist like Qencode to bring a simple and low cost storage solution to businesses who need it the most.

The post Backblaze + Qencode: Video Transcoding Made Simple appeared first on Backblaze Blog | Cloud Storage & Cloud Backup.