AWS Enters the High-performance Compute Market With New AI Chip

In a major step toward vertical integration, AWS announced three new Elastic Compute Cloud (EC2) instances built with custom silicon for high-performance computing (HPC). These new chips also mark Amazon’s entry into the high-performance computing space, easing the company's reliance on third-party designs and competing with manufacturers such as NVIDIA, Intel, and AMD.

AWS has long supported cloud-based computation and distributed computing. The recent announcement signals not only an incoming performance increase but also a shift toward custom silicon, providing companies with improved, purpose-built systems that can maximize performance.

 

The performance increase provided by next-gen AWS chips gives users in various fields increased access to computational resources at a lower overall cost. Image used courtesy of TechCrunch

 

In this article, we’ll take a closer look at Amazon’s new chips, their intended role in the HPC/AI space, and their effectiveness compared to offerings from other manufacturers. We'll also review the new AWS instances to help engineers determine which tool is right for the job.

 

Graviton3E: High-performance, Low-cost Computing

Extending its custom processor line, AWS' Graviton3E (G3E) is the latest addition to the Graviton family. Following the Graviton3, the G3E is built with next-gen applications in mind. Preliminary results show that the new chip delivers up to 35% improved vector-instruction performance than Graviton3 processors, making it a strong candidate for computationally intensive systems.

 

The C7gn instance (shown above) can be used for network-intensive applications, where throughput and bandwidth are critical. Image used courtesy of Business Wire

 

In early 2023, users can begin experimenting with the new processor on two new AWS instances: the Hpc7g (HPC) and C7gn (network-intensive) instances.

In both instances, the G3E provides a major uptick in performance. The Hpc7g instance produces up to 2x improved performance than the C6gn instances and 20% improved performance over Hpc6a instances. The C7gn instance, on the other hand, focuses on network bandwidth and throughput, doubling the bandwidth and increasing packet processing by 50% compared to the current generation.

 

Inferentia2: Making Machine Learning Faster

While the G3E processor is designed to improve computational efficiency, the Inferentia2 processor is built for artificial intelligence. The Inferentia2 allows designers to quickly test, evaluate, and deploy their inference models.

 

The Inf2 instances provide a major improvement in model deployment for cutting-edge machine- and deep-learning applications. Image used courtesy of TechCrunch

 

The Inferentia2 will be featured in the new Inf2 instances, where AWS has reported a 3x increase in computing power, a 4x increase in throughput, and 10x lower latency than current-gen Inf1 instances. In addition, Inf2 instances can support up to 175 billion parameters, making them strong contenders for large-scale models. The new Inf2 instances are currently available in preview.

 

Putting AWS Hardware in AWS Data Centers

As AWS continues to produce hardware built for high-performance computing, it puts itself in a competitive position in the HPC market. The G3E processor tackles HPC problems similar to both AMD and Intel processors, while the Inferentia2 chip can excel at parallel and data-heavy computations, much like NVIDIA GPUs. In addition, AWS' decision to replace outsourced hardware with in-house designs may impact its customer base for legacy manufacturers.

This is not to say that AWS is abandoning all third-party hardware. The new R7iz instances, for example, use 4th-gen Intel Xeon processors for more efficient access to EDA tools. However, it demonstrates that the trend toward in-house chip production is not limited to any one sector. Thanks to the newly minted instances, engineers and designers in all fields can benefit from the improvements in computing technology.