Intel rumored to launch cut-down AI chips for China, just like NVIDIA did with its AI GPUs

Intel is under continued pressure from US restrictions -- and could reportedly, like NVIDIA -- that would see Team Blue gimping their AI processors for the Chinese market.

2

This isn't a new move, with NVIDIA modifying its AI GPUs to meet US restrictions into China, and now it seems Intel could do it soon with "special edition" versions of its new Gaudi 3 AI accelerators. These two new gimped AI accelerators from Intel are rumored to be launched at the end of June, and the end of September, reports TrendForce.

Intel announced its new Gaudi 3 AI accelerator recently, and in the Gaudi 3 white paper it was discovered that Intel said it would be launching two special edition versions of Gaudi 3 that will be specifically for the Chinese market. There will be two hardware variants launched later this year: the HL-328 OAM-compatible Mezzanine card, and the HL-388 PCIe Accelerator Card. TrendForce reports that the HL-328 will launch on June 24, and the HL-388 on September 24.

• Read more: Intel Gaudi 3 AI accelerator: 128GB HBM2e at up to 3.7TB/sec, up to 900W power
• Read more: NVIDIA samples two new AI GPUs for China, both comply with US restrictions
• Read more: NVIDIA's tweaked H20 AI GPU mass production in Q2 2024, destined for China

The modified, made-for-China Gaudi 3 AI accelerators will share the same feature set with 96MB of on-chip SRAM, 128GB of HBM2e memory with 3.7TB/sec of memory bandwidth, a full PCIe 5.0 x16 interface, and decoding standards. However, US export restrictions on AI processors have a comprehensive computing performance (TPP) of high-performance AI chips to be below 4800 to export to China.

This means that the special edition made-for-China Gaudi 3 AI accelerator cannot have 16-bit performance over 150 TFLOPS. The Gaudi 3, in its original form, has 1835 TFLOPS in FP16/BF16 and will thus need to be cut down to meet US export restrictions for China.

Intel's upcoming made-for-China edition will need to reduce the core count (the original Gaudi 3 has 8 Matrix Multiplication Engines (MME) and 64 Tensor Processor Cores (TPC engines) and operating frequency. To get to under 150 TFLOPS of AI power, Intel would need to cut a gigantic 92% of its power to meet US export control requirements.

The new Gaudi 3 made-for-China variant will have 148 TFLOPS of FP16/BF16 performance, making it under the 150 TFLOPS threshold.