On April 14th, reports emerged revealing that Intel is preparing to launch a "special edition" of its Gaudi 3 AI chip specifically tailored for the Chinese market.
The China-specific Gaudi 3 includes two variants: the HL-328 OAM-compatible mezzanine card and the HL-388 PCIe acceleration card. The HL-328 is set to debut on June 24th, while the HL-388 will follow on September 24th.
Compared to the original version, the China-specific Gaudi 3 boasts the same 96MB SRAM on-chip memory, 128GB of HBM2e high-bandwidth memory with a bandwidth of 3.7TB/s, PCIe 5.0 x16 interface, and standard decoding capabilities.
However, due to US export controls on AI chips, the chip's overall computational performance (TPP) must be below 4800 to be exported to China. This implies that the China-specific Gaudi 3's 16-bit performance cannot exceed 150 TFLOPS.
While the original Gaudi 3 achieves performance of up to 1835 TFLOPS on FP16/BF16, the China-specific variant may need to reduce its AI performance by approximately 92% to comply with US export regulations.
Nevertheless, the decrease in performance also results in a significant reduction in power consumption. According to disclosed information, the TDP of the China-specific Gaudi 3 PCIe card and OAM card are both 450 watts, compared to 600 watts and 900 watts for the original versions, respectively.
