During Intel’s Vision 2024 conference, the Gaudi 3 AI processing chip was unveiled. This GPU is designed to accelerate AI workloads at the enterprise level, and Intel claims it is faster than Nvidia’s H100 GPU, the current industry benchmark.
This new Gaudi uses TSMC’s 5nm process (versus 7nm in Gaudi 2), making it more efficient than Gaudi 2. Gaudi 3 shares the same underlying architecture and underlying principles as Gaudi 2. The Gaudi 3 boasts up to 128 GB of HBM2e with bandwidth of 3.7 TB/s and a thermal design power of 900W.
As part of its benchmarking, Intel demonstrated Gaudi 3’s performance compared to Nvidia’s H100 GPU, which runs large language models such as LLAMA2-7B, LLAMA2-13B, and GPT 3-175B. The performance of Gaudi 3 has been increased by up to 1.7 times.
In addition, Intel claims Gaudi 3 consumes 2.3 times less power than Nvidia’s H100. In Q3, Intel will start shipping air-cooled Gaudi 3 models to partners. Q4 will see liquid-cooled units arrive.
