Nvidia, the world’s largest graphic chips supplier, reportedly will mass produce three reduced artificial intelligence (AI) chips for China’s customers in the second quarter of this year.
The California-based company originally planned to unveil three AI chips – H20, L20 and L2 – for the Chinese markets last November but the launch was delayed to 2024 as the company wanted to check whether the chips comply with the US export controls.
The three chips are in full compliance with US export policies, and will be manufactured by Taiwan’s Wistron Corp in the second quarter, according to an article published by Wccftech, a Canadian technology news website.
The first batch of the H20 chips will probably be delivered to customers around mid- or late-second quarter of 2024, said the report.
“In terms of parameters, H20’s performance density and computational power comply with the US exports policies,” a Chinese writer with the Shenzhen Xinbang Information Technology Co Ltd says in an article published on Tuesday.
He says H20 has a speed of 296 trillion floating point operations per second (teraflops or tflops) in FP8 Tensor Core operations, compared with H100’s 1,979 tflops and H200’s 3,958 tflops. He says, H200, the world’s most powerful AI chip, is 13 times faster than the H20.
Meanwhile, published accounts indicate the H20 is a comparatively lean and mean machine. Dylan Patel, an analyst at Semianalysis.com, says in an article published last November that H20 actually is over 20% faster than the H100 in large language model (LLM) reasoning, which can be used to generate content using very large datasets.
He says that, although the H100 is 6.68 times faster than the H20, people should also consider the MFU (model FLOPs utilization) rate, or the actual utilization rate, when measuring their performance.
As the H100’s MFU is only 38.1% while the H20’s can reach 90%, the performance of the H20 in actual multi-GPU interconnection environments is close to 50% of that of the H100.
Other technology writers say H20 has an advantage in power consumption as its thermal design power is 400 watts, lower than H100’s 700 watts.
Chinese markets
This saga dates back to In August 2022, when the Biden administration banned the exports of Nvidia’s A100 and H100 chips and AMD’s MI250 chip to China and Russia as these chips have high interconnect bandwidths – 600 gigabytes per second or above.
Nvidia later unveiled the A800 and H800 processors, which work at 400 and 300 gigabytes per second, respectively, targeting the Chinese markets. IT experts said A800 and H800 have performance rates of about 70% of the A100’s and H100’s, respectively.
On October 17, the US Commerce Department’s Bureau of Industry and Security (BIS) said it would use “performance” and “performance density” as new parameters to categorize restricted chips. Under the new rules, Nvidia’s A800, H800, L40, L40S and RTX 4090 chips are banned from being shipped to China. Nvidia wants to fill the resulting gap by shipping H20 to the country.
If this new chip can achieve 50% of H100’s speed while its power consumption is 43% less than that of H100, it may be attractive to Chinese customers, said some analysts.
“Although the H20 has a lower computational power than the H100, it will be set at a lower price, and it supports Nvidia’s special features such as NVLink and CUDA platforms,” said Kuo Ming-Chi, a Taiwanese technology analyst at the Hong Kong-based TF International Securities Group Ltd. “Chinese customers still have a strong interest in the H20 chip.”
Advantage over Chinese competitors shrinking
A Liaoning-based columnist says in an article on December 30 that H20 still has an advantage in performance and efficiency over Chinese AI chips but such an advantage is shrinking.
He says that, with Beijing’s policy and financial support, many Chinese chip makers are growing quickly and will one day break Nvidia’s monopoly in the AI chip markets.
In fact, some Chinese technology firms have already shifted to use local chips, including Huawei’s 910B and Biren’s BR100 chips, as they don’t want to wait for Nvidia’s delivery.
The pressure on Nvidia could inrease. On December 2, United States Commerce Secretary Gina Raimondo said in a forum that if any US firm redesigns its chip around a particular cut line that enables Chinese companies to do AI, the US government will control it the very next day.
The possibility can’t be ruled out that, if the performance of H20 can indeed reach 50% of H100’s, Raimondo may further tighten the export rules.
Read: Raimondo calls out Nvidia for China shipments
Follow Jeff Pao on Twitter at @jeffpao3
Put your money on expecting the US to ban exports of the H20 to China and any other chip NVIDIA creates for the China market.