Amazon AWS just launched its Trainium3 AI chip

BREAKING: Amazon AWS just launched its Trainium3 AI chip

Trainium3 UltraServers:

– Up to 4.4× higher performance – ~4× more memory bandwidth

– 4× better performance-per-watt vs Trainium2 UltraServers

– Scale to 144 chips (362 FP8 PFLOPs) per server and 1M+ chips in UltraClusters

AWS is positioning Trainium3 as a lower-cost alternative to Nvidia GPUs.

It says Trainium-based instances can cut training – and often inference

– costs by as much as ~50% vs comparable GPU clusters on large models,

for customers willing to adopt its Neuron software stack and Trainium instances.

Source: https://x.com/AskPerplexity/status/1995901639794860178?s=20