Generative AI Infrastructure at AWS | Amazon Web Services

Building and training generative artificial intelligence (AI) models, as well as predicting and providing accurate and insightful outputs requires a significant amount of infrastructure.

There’s a lot of data that goes into generating the high-quality synthetic text, images, and other media outputs that large-language models (LLMs), as well as foundational models (FMs), create. To start, the data set generally has somewhere around one billion variables present in the model that it was trained on (also known as parameters). To process that massive amount of data (think: petabytes), it can take hundreds of hardware accelerators (which are incorporated into purpose-built ML silicon or GPUs).

Given how much data is required for an effective LLM, it becomes costly and inefficient if an organization can’t access the data for these models as quickly as their GPUs/ML silicon are processing it. Selecting infrastructure for generative AI workloads impacts everything from cost to performance to sustainability goals to the ease of use. To successfully run training and inference for FMs organizations need:

1. Price-performant accelerated computing (including the latest GPUs and dedicated ML Silicon) to power large generative AI workloads.
2. High-performance and low-latency cloud storage that’s built to keep accelerators highly utilized.
3. The most performant and cutting-edge technologies, networking, and systems to support the infrastructure for a generative AI workload.
4. The ability to build with cloud services that can provide seamless integration across generative AI applications, tools, and infrastructure.

Overview of compute, storage, & networking for generative AI

Amazon Elastic Compute Cloud (Amazon EC2) accelerated computing portfolio (including instances powered by GPUs and purpose-built ML silicon) offers the broadest choice of accelerators to power generative AI workloads.

To keep the accelerators highly utilized, they need constant access to data for processing. AWS provides this fast data transfer from storage (up to hundreds of GBs/TBs of data throughput) with Amazon FSx for Lustre and Amazon S3.

Accelerated computing instances combined with differentiated AWS technologies such as the AWS Nitro System, up to 3,200 Gbps of Elastic Fabric Adapter (EFA) networking, as well as exascale computing with Amazon EC2 UltraClusters helps to deliver the most performant infrastructure for generative AI workloads.

Coupled with other managed services such as Amazon SageMaker HyperPod and Amazon Elastic Kubernetes Service (Amazon EKS), these instances provide developers with the industry’s best platform for building and deploying generative AI applications.

This blog post will focus on highlighting announcements across Amazon EC2 instances, storage, and networking that are centered around generative AI.

AWS compute enhancements for generative AI workloads

Training large FMs requires extensive compute resources and because every project is different, a broad set of options are needed so that organization of all sizes can iterate faster, train more models, and increase accuracy. In 2023, there were a lot of launches across the AWS compute category that supported both training and inference workloads for generative AI.

One of those launches, Amazon EC2 Trn1n instances, doubled the network bandwidth (compared to Trn1 instances) to 1600 Gbps of Elastic Fabric Adapter (EFA). That increased bandwidth delivers up to 20% faster time-to-train relative to Trn1 for training network-intensive generative AI models, such as LLMs and mixture of experts (MoE).

Watashiha offers an innovative and interactive AI chatbot service, “OGIRI AI,” which uses LLMs to incorporate humor and offer a more relevant and conversational experience to their customers. “This requires us to pre-train and fine-tune these models frequently. We pre-trained a GPT-based Japanese model on the EC2 Trn1.32xlarge instance, leveraging tensor and data parallelism,” said Yohei Kobashi, CTO, Watashiha, K.K. “The training was completed within 28 days at a 33% cost reduction over our previous GPU based infrastructure. As our models rapidly continue to grow in complexity, we are looking forward to Trn1n instances which has double the network bandwidth of Trn1 to speed up training of larger models.”

AWS continues to advance its infrastructure for generative AI workloads, and recently announced that Trainium2 accelerators are also coming soon. These accelerators are designed to deliver up to 4x faster training than first generation Trainium chips and will be able to be deployed in EC2 UltraClusters of up to 100,000 chips, making it possible to train FMs and LLMs in a fraction of the time, while improving energy efficiency up to 2x.

AWS has continued to invest in GPU infrastructure over the years, too. To date, NVIDIA has deployed 2 million GPUs on AWS, across the Ampere and Grace Hopper GPU generations. That’s 3 zetaflops, or 3,000 exascale super computers. Most recently, AWS announced the Amazon EC2 P5 Instances that are designed for time-sensitive, large-scale training workloads that use NVIDIA CUDA or CuDNN and are powered by NVIDIA H100 Tensor Core GPUs. They help you accelerate your time to solution by up to 4x compared to previous-generation GPU-based EC2 instances, and reduce cost to train ML models by up to 40%. P5 instances help you iterate on your solutions at a faster pace and get to market more quickly.

And to offer easy and predictable access to highly sought-after GPU compute capacity, AWS launched Amazon EC2 Capacity Blocks for ML. This is the first consumption model from a major cloud provider that lets you reserve GPUs for future use (up to 500 deployed in EC2 UltraClusters) to run short duration ML workloads.

AWS is also simplifying training with Amazon SageMaker HyperPod, which automates more of the processes required for high-scale fault-tolerant distributed training (e.g., configuring distributed training libraries, scaling training workloads across thousands of accelerators, detecting and repairing faulty instances), speeding up training by as much as 40%. Customers like Perplexity AI elastically scale beyond hundreds of GPUs and minimize their downtime with SageMaker HyperPod.

Deep-learning inference is another example of how AWS is continuing its cloud infrastructure innovations, including the low-cost, high-performance Amazon EC2 Inf2 instances powered by AWS Inferentia2. These instances are designed to run high-performance deep-learning inference applications at scale globally. They are the most cost-effective and energy-efficient option on Amazon EC2 for deploying the latest innovations in generative AI.

Another example is with Amazon SageMaker, which helps you deploy multiple models to the same instance so you can share compute resources—reducing inference cost by 50%. SageMaker also actively monitors instances that are processing inference requests and intelligently routes requests based on which instances are available—achieving 20% lower inference latency (on average).

AWS invests heavily in the tools for generative AI workloads. For AWS ML silicon, AWS has focused on AWS Neuron, the software development kit (SDK) that helps customers get the maximum performance from Trainium and Inferentia. Neuron supports the most popular publicly available models, including Llama 2 from Meta, MPT from Databricks, Mistral from mistral.ai, and Stable Diffusion from Stability AI, as well as 93 of the top 100 models on the popular model repository Hugging Face. It plugs into ML frameworks like PyTorch and TensorFlow, and support for JAX is coming early this year. It’s designed to make it easy for AWS customers to switch from their existing model training and inference pipelines to Trainium and Inferentia with just a few lines of code.

Cloud storage on AWS enhancements for generative AI

Another way AWS is accelerating the training and inference pipelines is with improvements to storage performance—which is not only critical when thinking about the most common ML tasks (like loading training data into a large cluster of GPUs/accelerators), but also for checkpointing and serving inference requests. AWS announced several improvements to accelerate the speed of storage requests and reduce the idle time of your compute resources—which allows you to run generative AI workloads faster and more efficiently.

To gather more accurate predictions, generative AI workloads are using larger and larger datasets that require high-performant storage at scale to handle the sheer volume in of data.

With Amazon S3 Express One Zone a new storage class purpose-built to high-performance and low-latency object storage for an organizations most frequently accessed data, making it ideal for request-intensive operations like ML training and inference. Amazon S3 Express One Zone is the lowest-latency cloud object storage available, with data access speed up to 10x faster and request costs up to 50% lower than Amazon S3 Standard, from any AWS Availability Zone within an AWS Region.

AWS continues to optimize data access speeds for ML frameworks too. Recently, Amazon S3 Connector for PyTorch launched, which loads training data up to 40% faster than with the existing PyTorch connectors to Amazon S3. While most customers can meet their training and inference requirements using Mountpoint for Amazon S3 or Amazon S3 Connector for PyTorch, some are also building and managing their own custom data loaders. To deliver the fastest data transfer speeds between Amazon S3, and Amazon EC2 Trn1, P4d, and P5 instances, AWS recently announced the ability to automatically accelerate Amazon S3 data transfer in the AWS Command Line Interface (AWS CLI) and Python SDK. Now, training jobs download training data from Amazon S3 up to 3x faster and customers like Scenario are already seeing great results, with a 5x throughput improvement to model download times without writing a single line of code.

To meet the changing performance requirements that training generative AI workloads can  require, Amazon FSx for Lustre announced throughput scaling on-demand. This is particularly useful for model training because it enables you to adjust the throughput tier of your file systems to meet these requirements with greater agility and lower cost.

EC2 networking enhancements for generative AI

Last year, AWS introduced EC2 UltraCluster 2.0, a flatter and wider network fabric that’s optimized specifically for the P5 instance and future ML accelerators. It allows us to reduce latency by 16% and supports up to 20,000 GPUs, with up to 10x the overall bandwidth. In a traditional cluster architecture, as clusters get physically bigger, latency will also generally increase. But, with UltraCluster 2.0, AWS is increasing the size while reducing latency, and that’s exciting.

AWS is also continuing to help you make your network more efficient. Take for example a recent launch with Amazon EC2 Instance Topology API. It gives you an inside look at the proximity between your instances, so you can place jobs strategically. Optimized job scheduling means faster processing for distributed workloads. Moving jobs that exchange data the most frequently to the same physical location in a cluster can eliminate multiple hops in the data path. As models push boundaries, this type of software innovation is key to getting the most out of your hardware.

In addition to Amazon Q (a generative AI powered assistant from AWS), AWS also launched Amazon Q networking troubleshooting (preview).

You can ask Amazon Q to assist you in troubleshooting network connectivity issues caused by network misconfiguration in your current AWS account. For this capability, Amazon Q works with Amazon VPC Reachability Analyzer to check your connections and inspect your network configuration to identify potential issues. With Amazon Q network troubleshooting, you can ask questions about your network in conversational English—for example, you can ask, “why can’t I SSH to my server,” or “why is my website not accessible”.

Conclusion

AWS is bringing customers even more choice for their infrastructure, including price-performant, sustainability focused, and ease-of-use options. Last year, AWS capabilities across this stack solidified our commitment to meeting the customer focus and goal of: Making generative AI accessible to customers of all sizes and technical abilities so they can get to reinventing and transforming what is possible.

Additional resources