Amazon dropped details on their Trainium3 UltraServers and teased the next Trainium4 chips at their annual conference. The UltraServer racks cram 144 Trainium3 processors into one system, delivering over four times the performance and efficiency compared to the older version. They built custom networking tech called NeuronSwitch to link these things into clusters that could theoretically hit a million chips for training massive AI models.
The upcoming Trainium4 generation will pump out six times better FP4 performance and pack way more memory. Amazon is adding support for NVIDIA's NVLink tech, so companies can mix their existing NVIDIA setups with Trainium hardware without rebuilding everything from scratch.
Big cloud customers are already using these chips to cut their training bills. Amazon is going hard against Google and Meta in the custom silicon race.
The upcoming Trainium4 generation will pump out six times better FP4 performance and pack way more memory. Amazon is adding support for NVIDIA's NVLink tech, so companies can mix their existing NVIDIA setups with Trainium hardware without rebuilding everything from scratch.
Big cloud customers are already using these chips to cut their training bills. Amazon is going hard against Google and Meta in the custom silicon race.
