Amazon announced the launch of updated versions of its specialized chips at the AWS re:Invent developer conference in Las Vegas. The Graviton 4 chip, a general-purpose microprocessor chip utilized by SAP and others for extensive workloads, and the Trainium 2 chip, a special-purpose accelerator chip tailored for large neural network programs like generative AI.
Trainium 2 is specifically optimized for training tasks, particularly for large language models (LLMs) and foundation models, such as OpenAI’s GPT-4. It is engineered to manage neural networks with trillions of parameters, focusing on scalability in the AI industry.
The new Trainium 2 chip promises enhanced performance, offering up to four times faster training speed and three times more memory capacity compared to its predecessor. Amazon also introduced the Graviton 4 chip, which boasts a 30% improvement in compute performance and competes with Intel and AMD processors based on the x86 chip standard.
Amazon revealed its collaboration extension with Nvidia to incorporate Nvidia’s cutting-edge chips into its cloud computing service. The Trainium 2 chips will be available in Amazon’s EC2 cloud computing service as “Trn2” instances, with the potential to scale up to 100,000 instances interconnected by the Elastic Fabric Adapter, delivering a total of 65 exaFLOPs of computing power.
Customers leveraging these advanced chips can train a 300-billion parameter LLM in significantly reduced time frames. Furthermore, Amazon’s strategic investments in generative AI, including a substantial stake in Anthropic, highlight its commitment to innovation in AI silicon.
Apart from AI-focused Trainium chips, the Graviton processors cater to conventional workloads and have been adopted by various companies for diverse tasks such as databases, analytics, web servers, and more. SAP reported a 35% enhancement in price performance for analytical workloads using the Graviton chips.
The introduction of these new chips follows the previous launches of Graviton 3 and the original Trainium. Amazon’s advancements in AI silicon align with similar efforts by industry peers like Microsoft and Google, with Amazon extending its partnership with Nvidia to incorporate the GH200 Grace Hopper multi-chip product into its cloud services next year. This collaboration aims to accelerate neural network training processes, further solidifying AWS’s position as a key player in AI research and development.
