AWS's new Trainium2 chips, now available, enable faster training for large language models, offering significant performance boosts over previous generations and rival offerings.
"These UltraServers will feature 64 interconnected Trainium2 chips, scaling up to 83.2 peak petaflops of compute, an impressive leap for machine learning applications.
AWS claims Trainium2 will deliver '3x higher token-generation throughput' for Meta's Llama 405B model compared to rivals, enhancing performance in cloud computing.
The upcoming AI compute cluster, powered by hundreds of thousands of Trainium2 chips, is projected to be the largest in the world, five times more powerful than its predecessor.
Collection
[
|
...
]