Hardware

Broadcom’s router for distributed AI, built to link over a million accelerators across separate data centres.

A new class of Rubin GPU on cheaper GDDR7 memory, aimed at the context phase of long-context inference.

NVIDIA’s first SRAM-based low-latency inference accelerator, built on licensed Groq technology, sitting alongside Rubin.

Keeps Blackwell’s two dies and node but raises NVFP4 compute by 1.5x and memory to 288 GB.

Moves to HBM4 and grows from 208bn to 336bn transistors compared with Blackwell Ultra.

Replaces Grace with the new Vera CPU and the Blackwell GPUs with two Rubin GPUs on HBM4.

Pairs Grace with two dual-die Blackwell GPUs and adds FP4 inference.

Moves from one die to two on TSMC 4NP and adds FP4 inference.

Same Hopper compute as H100, with 141 GB of HBM3e at 4.8 TB/s, which NVIDIA says nearly doubles inference speed on Llama…

Introduced FP8 and the Transformer Engine, which NVIDIA said can speed up transformer networks as much as 6x versus the previous generation.