Intel's Habana Gaudi2 accelerators, produced by subsidiary Habana Labs, were just 10% slower than Nvidia's system. The Gaudi2 system is based on a seven-nanometer manufacturing node, compared to the five-nanometer Hopper GPU. Intel plans to update the system with FP8 precision quantization later this month, which it claims will double its performance. Other competitors included Google, whose latest Tensor Processing Units fell short of Nvidia's numbers, and Qualcomm, whose Cloud AI100 chipset showed strong performance due to recent software updates.
Key takeaways:
- Nvidia Corp.’s most advanced chips were the top performers in MLCommons' benchmarks for testing AI inference, with Intel Corp.’s hardware coming a surprisingly close second.
- The benchmarks are based on a large language model (LLM) with 6 billion parameters, known as GPT-J 6B, designed to summarize texts from news articles.
- Nvidia's GH200 Grace Hopper Superchip and HGX 100 system displayed the most impressive results across all of MLPerf’s data center tests, including tasks such as computer vision, speech recognition, and medical imaging.
- Intel's Habana Gaudi2 accelerators came a close second to Nvidia’s chips, with the company promising a two-times performance boost with an upcoming update and a new 5nm Gaudi3 chipset in the works.