Raphael ThysIntel kicked off its Vision 2024 conference with the unveiling of the Gaudi 3 AI processing chip. Gaudi 3 is designed to accelerate AI workloads at an enterprise level, and Intel reckons it's faster than the current industry benchmark - Nvidia's H100 GPU.
Gaudi 3 uses the same architecture and underlying principles as Gaudi 2, but it uses TSMC's 5nm process (compared to TSMC's 7nm in Gaudi 2), making it more efficient.
Gaudi 3 has up to 128GB of HBM2e with 3.7 TB/s of bandwidth and 900W TDP.
Intel showed benchmarks comparing Gaudi 3 to market leader Nvidia and its dominating H100 GPU, running different large language models like LLAMA2-7B, LLAMA2-13B, and GPT 3-175B. Gaudi 3 is up to 1.7x faster.
Intel also says that Gaudi 3 is up to 2.3 times more power efficient than Nvidia's H100.
Intel has already sent air-cooled Gaudi 3 models for sampling to partners with general availability to follow in Q3. Liquid-cooled units will come in Q4.