Amazon just dropped its first 3nm chips for AI work in AWS data centers. Trainium2 handles the heavy training stuff with better memory design and more power efficiency, while Graviton5 takes care of regular cloud tasks without needing crazy cooling upgrades. Both chips help AWS pack more compute into the same racks, and they work together to support the massive model training setups everyone's running these days.
The wild part is Amazon's planning to bake in NVLink support for its next batch of processors. That means you can mix Amazon silicon with Nvidia GPUs in the same cluster without everything grinding to a halt at the interconnect layer. It's basically Amazon hedging its bets while still building out its own hardware, which makes sense when you're trying to keep costs down and avoid getting bottlenecked by GPU shortages during another AI hype cycle.
The wild part is Amazon's planning to bake in NVLink support for its next batch of processors. That means you can mix Amazon silicon with Nvidia GPUs in the same cluster without everything grinding to a halt at the interconnect layer. It's basically Amazon hedging its bets while still building out its own hardware, which makes sense when you're trying to keep costs down and avoid getting bottlenecked by GPU shortages during another AI hype cycle.
