Intel Gaudi 3 Accelerator Performance Surpasses NVIDIA H100

Intel claims that the Gaudi 3 accelerator has gained an advantage in competition with NVIDIA H100. Gaudi 3 performs 1.7x faster than H100 when training Llama2-13B at FP8 precision across a 16-accelerator cluster. In addition, Gaudi 3 has also significantly improved the H200/H100 inference performance, increasing by 1.3 times to 1.5 times, and the energy efficiency has increased to 2.3 times.