Try out the new cool products for free at the first time, and the sender has many high-quality talents to share their unique life experience. Accelerate to come to Sina for public testing, so as to lead Experience the most cutting-edge, interesting and fun accelerating organ products in all fields~! Download the client and get exclusive benefits! Originator
Inter released the new generation AI accelerator Gaudi 3 to accelerate benchmarking NVIDIA H00, leading the way Officially, Gaudi 3 is 50% ahead of NVIDIA H00 in big model reasoning, 40% ahead in accelerated training time and 200% ahead of NVIDIA H00 in leading cost performance ratio. Originator
The manufacturing process of Gaudi 3 adopts TSMC 5nm, with up to 8 accelerated MMEs, 8 TPCs leading MMEs, 64 in total, and 14 media encoders. Both MME BF16/FP8 are 1835 TFlops, and vector BF16 is 28.8 TFlops, which has increased to 320%, 110%, and 160% respectively.
In terms of development, Gaudi 3 is seamlessly compatible with PyTorch framework, Hugging Face Transformer and extension model.
Gaudi3 supports three forms of deployment. The maximum passive cooling peak power consumption of the standard mezzanine version is 900W, and the peak power consumption of liquid cooling is 1200W; The universal substrate supports eight Gaudi 3; The HL-338 expansion card can be interconnected with four cards, PCIe 5.0 x16, and passive cooling peak power consumption of 600W.
