In the second quarter of this year, Nvidia, the largest supplier of graphic chips in the world, is reportedly going to mass produce three reduced artificial intelligence ( AI ) chips for China’s customers.
The California-based business originally intended to release three AI chips—H20, L20 and L2—for Chinese industry last November, but it postponed that release until 2024 in order to verify that the chips complied with US export restrictions.  ,
According to an essay published by the American tech news website Wccftech, the three cards will be produced by Taiwan’s Wistron Corp in the second quarter and are fully compliant with US export regulations.
According to the report, the first shipment of H20 bits will likely be delivered to customers in the middle or late next quarter of 2024.
According to a Chinese author for Shenzhen Xinbang Information Technology Co Ltd., H20’s efficiency density and computing power adhere to US export policies in terms of parameters.
In FP8 Tensor Core operations, he claims that H20 has a speed of 296 trillion floating point operations per second (teraflops or tflop ) when compared to H100’s 1, 979, and H200’S 3, 958, respectively. The H200, the most potent Artificial device in the world, is 13 times faster than the H20, according to him.
The H20 is a relatively lean and mean system, according to published transactions. In a piece that was published in November, analyst Dylan Patel from Semianalysis.com claims that the H20 performs large language model ( LLM) reasoning, which can be used to produce content using very large datasets, over 20 % faster than the M100.
He claims that when evaluating the performance of the H100, people should take into account the MFU ( model FLOPs utilization ) rate, also known as the actual utilization rate.  ,
The H20 performs close to 50 % of as well in actual multi-GPU connection conditions as the H100 does because its MFU is just 38.1 % while its may reach 90 %.  ,
According to other technologies writers, H20 consumes less power than H100 because its thermal pattern power is 400 watts. In addition,  ,
Foreign markets and nbsp,
The A100 and H100 chips from Nvidia and AMD as well as the MI250 device from AMD were prohibited from being exported to China and Russia in August 2022 due to their high connection bandwidths—600 gigabytes per minute or higher—by the Biden administration.  ,
The A800 and H800 computers, which operate at 400 and 300 terabytes per minute, between, were eventually unveiled by Nvidia with a focus on the Chinese marketplaces. According to IT professionals, the A800 and H800 have efficiency rates that are roughly 70 % of those of an A100 and an H100, respectively.
The Bureau of Industry and Security ( BIS ) of the US Commerce Department announced on October 17 that restricted chips would now be categorized using “performance” and”performance density” as new criteria. The new regulations forbid the shipment of Nvidia’s A800, H800s, L40S, and RTX 4090 bits to China. Nvidia wants to ship H20 to the nation to close the resulting distance.
According to some experts, this new chip may be appealing to Chinese consumers if it can reach 50 % of H100’s speed while using 43 % less power than that.
Foreign buyers continue to be very interested in the H20 chip, according to Kuo Ming-Chi, a Chinese tech researcher at the Hong Kong-based TF International Securities Group Ltd.” Despite having less computing authority than the HP100, it will be priced more affordably, and it supports Nvidia’s unique features like NVLink and CUDA platforms.”
shrinking advantages over Chinese rivals
In an article published on December 30, a Liaoning-based blogger claims that while H20 still outperforms Chinese AI cards in terms of performance and efficiency, this benefit is eroding.  ,
According to him, many Chinese chip manufacturers are expanding rapidly thanks to Beijing’s policy and financial assistance, and they will eventually overthrow Nvidia in the market for AI chips.  ,
In fact, some Chinese technology companies have already switched to using local chips, such as the 910B and BR100 chips from Huawei and Biren, because they do n’t want to wait for Nvidia’s arrival.
The strain on Nvidia might increase. In a conference on December 2, United States Commerce Secretary Gina Raimondo stated that the US government will take control of any US company if it redesigns its chip to fit into the specific split line that allows Chinese firms to use AI the very next day.
It is possible that Raimondo will strengthen the trade restrictions even more if H20 achievement is able to match that of H100s by 50 %.  ,
Read: Raimondo criticizes Nvidia for supplies to China
At , @jeffpao3 is Jeff Pao’s Twitter account.