Intel Unveils Modified Gaudi 3 AI Chips for China, with Significant Performance Reduction
Intel's launch of modified Gaudi 3 chips for the Chinese market is a response to the US sanctions on chip exports to China. These modified chips will experience a substantial drop in performance, with a reduction of approximately 92% in AI capabilities. Despite the performance limitations, the Chinese editions of the Gaudi 3 chips will retain the same specifications as the original versions.
Intel, the American chip giant, is set to launch modified versions of its Gaudi 3 AI chips exclusively for the Chinese market. The announcement comes as Intel aims to navigate the US sanctions on chip exports to China. According to a white paper released by Intel, the two China-exclusive models, HL-328 and HL-388, are scheduled to be launched on June 24 and September 24, respectively. However, these modified chips will experience a significant drop in performance compared to the original versions.
In order to comply with US export controls, the Chinese editions of the Gaudi 3 chips will need to reduce their AI performance by approximately 92%. This means that the computing performance of these chips will need to be below 4,800 TPP (total processing power) to be eligible for export to China. As a result, the 16-bit performance of the Chinese editions of the Gaudi 3 chips will not exceed 150 teraflops.
Similar to Nvidia's Chinese-edition H20 chip, the modified Intel chips will offer a performance of 148 teraflops in FP16/BF16, just below the permitted limit. In comparison, the original version of the Gaudi 3 chip can achieve an impressive 1,835 teraflops in FP16/BF16. These formats are commonly used in AI computing to strike a balance between accuracy and efficiency when processing large amounts of data.
Despite the performance reduction, the HL-328 and HL-388 versions of the Gaudi 3 chips will retain the same specifications as the original versions. They will feature a 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 will have a thermal design power (TDP) of just 450 watts, compared to the higher TDP of 600 watts and 900 watts for the original versions.
This move by Intel follows Nvidia's strategic re-entry into the Chinese market with the H20 AI chip in the first quarter of 2024. The modified-for-China H20 chip offers significantly lower AI computing power compared to the original H100 chip available in the rest of the world.
Intel's launch of modified Gaudi 3 chips for the Chinese market is a response to the US sanctions on chip exports to China.
These modified chips will experience a substantial drop in performance, with a reduction of approximately 92% in AI capabilities.
Despite the performance limitations, the Chinese editions of the Gaudi 3 chips will retain the same specifications as the original versions.
Source: TECHNODE