Intel Plans Special Edition Gaudi 3 for China

In the wake of Intel's recent unveiling of its latest AI acceleration chip, Gaudi 3, the company is gearing up to introduce a customized variant designed specifically for the Chinese market. This specialized edition encompasses two hardware iterations: the HL-328 OAM-compatible Mezzanine Card and the HL-388 PCIe acceleration card.

Reports indicate that Intel has divulged this information within its Gaudi 3 whitepaper. The HL-328 is slated for release on June 24th, followed by the HL-388 on September 24th.

In terms of hardware specifics, the Chinese special edition of Gaudi 3 mirrors the original with its 96MB SRAM on-chip storage, 128GB HBM2e high-bandwidth memory boasting a 3.7TB/s bandwidth, PCIe 5.0 x16 interface, and comparable decoding standards. However, due to U.S. export control regulations governing AI chips, stringent compliance dictates that the comprehensive computing performance (TPP) must not exceed 4800 to facilitate export to China. Consequently, the 16-bit performance of the Chinese special edition Gaudi 3 cannot surpass 150 TFLOPS.

Intel's disclosed data reveals Gaudi 3's capability to achieve 1835 TFLOPS in FP16/BF16, outshining NVIDIA's H100 by 40% in large-scale model training and exhibiting 50% higher inference efficiency.

Evidently, the Chinese special edition of Gaudi 3 necessitates a substantial reduction in AI performance to adhere to export regulations, mandating significant cuts in core numbers (the original version featuring 8 matrix math engines and 64 tensor cores) and operating frequencies.