Nvidia’s China-market H20 chips hit another speed bump
Integration woes delay Nvidia's hopes of maintaining grip on Middle Kingdom
Nvidia has reportedly delayed the launch of its latest Chinese-market AI accelerators until early next year over issues integrating the chip into server platforms.
Citing two sources familiar with the matter, Reuters reports that Nvidia has informed customers its rumored H20 GPU, which is designed to limbo under the Biden administration's latest round on export controls to China, won't be ready until February or March.
The latest round of sanctions, revealed last month, placed stiffer controls on the sale of GPUs and AI accelerators sold to China and other entities of concern. Specifically, the rules lowered the performance limits for chips to the point where most of Nvidia's current generation Ada Lovelace, Hopper and even some older Ampere generation parts could no longer be sold in the Middle Kingdom without a license.
The rules were so strict that even Nvidia's flagship gaming GPU, the RTX 4090 was banned from being sold to Chinese buyers. However, less than a month after the export controllers were announced, Nvidia — which had been preemptively ordered to cease shipments to China early — appeared to have uncovered a loophole.
Multiple reports indicated the California-based chipmaker had three new CPUs in the works including the H20, L20, and L2. According to SemiAnalysis, which managed to get an early look at the specs of these cards, revealed they're likely variants of Nvidia's H200, L40, and L4 which have been nerfed to comply with performance and density limits.
H20 is particularly notable as it will allegedly offer 96GB of memory, 4Tb/s of memory bandwidth, and 296 teraFLOPS of FP8 performance. While far less than the full-fat H200, the chip is said to deliver 20 percent faster performance than the older H100 in AI inferencing workloads. This could be down to a faster complement of high-bandwidth memory.
This wouldn't be the first time Nvidia and others have reworked their chips to comply with export restrictions. Last fall, after the first round of sanctions went into effect, Nvidia revealed a down-tuned version of its popular A100 accelerator of the Chinese market dubbed the A800. The chipmaker quickly followed it up with the Hopper-based H800 earlier this year.
Intel also announced a slower version of its Habana Gaudi2 accelerator for the Chinese market as well. The x86-giant also developed a variant of its GPU Max cards for "other markets," which we assumed at the time to mean China.
- Alibaba shuts down quantum lab, donates it to university
- AI chip outfit Graphcore's sales to China hit by US export rules
- Nvidia revenue explodes, led by datacenter products and … InfiniBand?
- China relents: Broadcom-VMware merger approved at last
Nvidia's ongoing efforts to jump through Uncle Sam's hoops has sparked the ire of competing chipmakers. Earlier this month Cerebras CEO put Nvidia on blast calling these efforts 'un-American" and likening the chipmaker to an AI arms dealer. There may be more on that story later.
While the H20 had been expected to ship earlier this month once the export restrictions went into effect on November 16, Chinese vendors hoping to get their hands on the accelerator will have to wait a little longer. Reportedly server manufacturers were facing issues integrating the chip into their platforms.
Having said that, Nvidia is reportedly pushing ahead with the launch of the slower L20 and L2 GPUs. ®