Intel Vision On paper, Intel's Habana Gaudi3 AI accelerators don't look ready to take on Nvidia's H100 thanks to older process technology and slower HBM memory that delivers fewer FLOPS. But Gelsinger's gang insists their latest parts can not only go toe-to-toe with the H100 in inference, but best it in training.
Report: Intel's Gaudi-3 50% faster than Nvidia's H100 processor | World Business Watch
Announced at Intel Vision on Tuesday, the accelerator comes four years after the x86 giant abandoned its Nervana architecture in favor of Habana's Gaudi accelerators, and nearly two years after its Gaudi2 chip arrived.
For its third generation, the Habana team opted for a multi-die architecture, joining AMD and Nvidia, which have already embraced advanced packaging to scale their chips beyond the benchmark limit.
Compared to AMD's MI300X, or even Intel's Ponte Vecchio GPUs, the Gaudi3 is relatively conservative in this regard, using a high-speed interconnect to make a pair of compute arrays behave as one.
