Intel announced its third-generation AI accelerator, Gaudi 3, at Vision 2024. This chip, built on the architecture of its predecessor, Gaudi 2, has 2 silicon dies connected by high-bandwidth communication. It provides double the AI compute power of the previous chip using 8-bit floating-point infrastructure and also increases efficiency using the BFloat 16 number format. Gaudi 3 shows a 40% faster training time and superior inferencing performance in comparison to Nvidia’s H100. Intel’s Gaudi 3 also uses less energy for LLM tasks due to its large-matrix math engines. It has more HBM and better memory bandwidth than H100 and uses TSMC’s N5 process technology. In related news, Nvidia has announced a competitor, the Blackwell B200, based on TSMC’s N4P process, which performs slightly better in terms of efficiency and density. The next generation of Intel’s Gaudi is code-named Falcon Shores and it is expected to use Intel’s 18A technology.
The ainewsarticles.com article you just read is a brief synopsis; the original article can be found here: Read the Full Article…
