Intel is preparing to launch modified Gaudi 3 chips for the Chinese market when it releases the new generation AI accelerator chip this year, according to British technology media outlet The Register. Intel has disclosed the existence of the two China-exclusive models in its Gaudi 3 white paper, with HL-328 set to launch on June 24 and HL-388 on September 24.
Why it matters: Intel’s launch of customized Gaudi 3 chips for China is part of the American chip giant’s efforts to work within US sanctions on chip exports to China.
Details: Based on the information provided in Intel’s Gaudi 3 white paper, the Chinese editions of the Gaudi 3 models may ultimately require a reduction of around 92% in AI performance to comply with US export controls.
- Due to US export restrictions, high-performance AI chips need a computing performance below 4,800 TPP (total processing power) to be eligible for export to China. This implies that the 16-bit performance of the Chinese editions of the Gaudi 3 chips will not exceed 150 teraflops.
- Similar to the Nvidia’s Chinese-edition H20 chip, the two modified Intel chips will offer a 148 teraflops of FP16/BF16 performance, just below the permitted limit.
- According to Intel, the original version of the Gaudi 3 chip can achieve 1,835 teraflops in FP16/BF16, formats used in computing to represent a balance between accuracy and efficiency for chips commonly used for AI to process large amounts of data quickly and efficiently.
- Compared to the original version, the HL-328 and the HL-388 versions have the same 96 MB cache, 128GB of HBM2e memory with a bandwidth of 3.7TB/s, PCIe 5.0 x16 interfaces, and decoding standards. However, both models are capable of just 450 watts of thermal design power (TDP), while the original versions of respective models have a higher TDP of 600 watts and 900 watts.
Context: In the first quarter of 2024, Nvidia strategically re-entered the Chinese market with the H20 AI chip. In terms of performance, the modified-for-China H20 chip’s AI computing power is slightly less than 15% of the H100, the original version of the chip available in the rest of the world.