AWS unveils next-gen AI accelerator to challenge Nvidia’s dominance
Amazon Web Services (AWS) has revealed its upcoming AI accelerator, Trainium3, promising a fourfold performance increase over its predecessor, Trainium2. Announced during AWS’s re:Invent conference, the chip is set to launch in late 2024 and could mark a significant leap forward in the competition for AI hardware supremacy.
Breaking down Trainium3
Trainium3 is expected to be the first machine learning accelerator built on a 3nm process, delivering a 40% efficiency improvement over Trainium2. While specific technical details remain scarce, AWS claims a single Trainium3 UltraServer could achieve up to 1.3 exaFLOPS of AI compute performance when leveraging sparsity techniques.
AWS has yet to disclose details about the chip’s memory architecture or the precise floating-point precision used for its peak performance calculations. However, the teaser suggests Trainium3 aims to rival or exceed the capabilities of Nvidia's upcoming Blackwell architecture, which also introduces innovations in ultra-low precision AI computations.
Trainium2 goes mainstream
As the industry awaits Trainium3, AWS is rolling out Trainium2 instances for general availability. These chips offer up to 1.3 petaFLOPS of dense FP8 compute and feature 96GB of high-bandwidth memory delivering 2.9TBps of bandwidth per accelerator.
The Trainium2 UltraServer configuration—comprising 64 interconnected accelerators—delivers 83.2 petaFLOPS of FP8 performance, with potential scaling to 332.8 petaFLOPS using sparsity techniques. AWS also claims Trainium2-based EC2 instances provide 30–40% better price-performance than Nvidia GPU-based instances.
For large-scale AI workloads, Trainium2 is set to power AWS’s collaboration with Anthropic under Project Rainier. This supercomputing initiative, expected to debut in 2025, will feature “hundreds of thousands” of Trainium2 chips, offering 5x the exaFLOPS required to train Anthropic’s latest AI models.
AWS not ditching Nvidia yet
Despite its aggressive investment in custom silicon through Annapurna Labs, AWS is not sidelining Nvidia. The company continues to deploy Nvidia’s cutting-edge GPUs, including the forthcoming Grace-Blackwell Superchips, as part of Project Ceiba, a supercomputer expected to deliver 51 exaFLOPS of dense BF16 compute.
AWS’s dual approach of advancing proprietary AI chips while maintaining strong ties with Nvidia reflects the growing demand for diversified, scalable AI solutions.
As the computing giant continues to unveil details about Trainium3, the industry will watch closely to see if these next-gen chips can live up to their promise and deliver a viable challenge to Nvidia’s supremacy in the AI era.
