US government requirements for the acceptable performance of accelerators allowed to be shipped to China are mostly formulated in the form of measurable characteristics, but are rarely made public. Intel, as it became known this week, will also face restrictions on the delivery of Gaudi accelerators to China.
Image Source: Intel
The processor giant, which is going through a difficult period in its history, can hardly be called a noticeable player in the segment of computing accelerators for artificial intelligence systems. Of the 10% of the market that Nvidia leaves to other market participants, the bulk is accounted for by AMD, and Intel’s role in this type of commercial activity is generally minimal. By the way, this omission did not go unnoticed by the new Intel CEO Lip-Bu Tan, and he promised to strengthen the company’s position in the segment of accelerators for AI systems.
According to the Financial Times, Intel has already notified its Chinese customers of the need to obtain an export license to supply its computing accelerators to China, which have a performance level higher than the limits declared by the US authorities. Thus, the threshold for memory bandwidth is 1400 GB/s, for interface bandwidth – 1100 GB/s, while in total these values cannot exceed 1700 GB/s. Like Nvidia H20 accelerators, Intel’s Gaudi family solutions are obviously subject to restrictions.
Nvidia has already estimated its possible losses from new US sanctions in this area at $5.5 billion based on the results of the first quarter alone. According to unofficial data, these losses will easily exceed $10 billion based on the results of the entire current year. AMD is currently content with more modest estimates of damage at the level of $800 million, and Intel has not yet published its calculations on this topic. The mention of the need to obtain export licenses for the supply of accelerators to China is just a formality. In practice, all three companies have very little chance of acquiring such licenses.