Amazon Web Services (AWS) CEO Adam Selipsky blew the lid off the cloud provider’s latest custom silicon today, announcing the first instances powered by the cloud provider’s next-generation Graviton3 data center CPUs and Trainium artificial intelligence (AI) accelerators.

“The Graviton3 chips are a big leap forward, 25% faster on average for general compute workloads [compared] to Graviton2 and they perform even better for certain specialized workloads,” Selipsky boasted during his opening keynote at AWS re:Invent.

Introduced in 2018, Graviton was Amazon’s first crack at building a cloud-centric CPU. Based on an Arm architecture, the chip saw broad success, especially among price-conscious customers willing to work around technical challenges. AWS followed it up two years later with Graviton2, which thanks to a more advanced Arm Neoverse microarchitecture and smaller, 7-nanometer process node, offered even greater performance.

“Today thousands of customers are using Graviton2-based instances and reaping the benefits and price performance over a very wide range of workloads, including big data analytics, game servers, and high-performance computing,” Selipsky said.

The first instances using AWS’s Graviton3 are now available in preview starting today on Elastic Compute Cloud (EC2) C7g instances.

The new chips offer 25% higher performance in general compute, twice the floating-point performance in scientific and cryptographic workloads, and up to three times higher performance in machine learning applications, compared to the previous generation chip, Selipsky boasted.

Graviton3 achieves these performance metrics while consuming half the power of “comparable instances,” he added.

While Amazon has yet to release specifics regarding the chips themselves, the chips will beat Intel and AMD to offer DDR5 memory support, a feature that isn’t expected to reach x86 servers until Intel’s Sapphire Rapids hits the market early next year.

Alongside Graviton3, AWS announced the availability of its Trainium AI accelerator.

“Training machine-learning models and running inferences are highly compute intensive, and many of you have asked us to find a way to lower the cost of those machine-learning workloads,” Selipsky said.

Trainium, announced at last year’s re:Invent conference, is the companion to Amazon’s earlier Inferentia AI inferencing accelerator, which launched in 2019. As its name suggests, Trainium accelerates AI training models.

“We expect [Trainium] to deliver the best price performance for training deep learning models in the cloud and the fastest on EC2,” Selipsky said.

Amazon’s TRN1 instances are the first to use the chips and are tailored for image recognition, natural-language processing, fraud detection, and forecast training models.

The instances are also AWS’ first to offer 800 Gb/s networking to enable training that scales across multiple compute nodes.

“Sometimes with machine learning workloads you need more processing than any single instance can handle, and we can network these together,” Selipsky said.

Using the technology, Amazon claims it can build customers tens of thousands of AI accelerators interconnected by more than a petabyte of networking throughput.

“Now with both Trainium and Inferentia powered instances, customers can have the best priceperformance for machine learning from scaling training workloads, to accelerating deep learning workloads in production with high-performance inference,” Selipsky said.