The company has not disclosed specific performance numbers for Trainium2, but it claims that its Trn2 instances can scale out with up to 100,000 Trainium2 chips to achieve up to 65 ExaFLOPS of low-precision compute performance for AI workloads. This scaling could significantly reduce the training time for a 300-billion parameter large language model from months to weeks. AWS partners, such as Anthropic, are ready to deploy the new accelerator.
Key takeaways:
- Amazon Web Services has introduced Trainium2, a new accelerator for artificial intelligence workloads that significantly increases performance compared to its predecessor.
- Trainium2 is designed specifically for training foundation models and large language models, featuring four times higher training performance, two times higher performance per watt, and three times as much memory.
- Amazon aims to enable its clients to access up to 65 'AI' ExaFLOPS performance for their workloads, with its Trn2 instances scalable with up to 100,000 Trainium2 chips.
- Amazon has partners, such as Anthropic, ready to deploy the AWS Trainium2 accelerators, with the expectation that it will be at least 4x faster than the first-generation Trainium chips for some key workloads.