Nvidia has downgraded three graphic processing models (GPUs) for the Chinese language markets after it was banned by the US authorities from transport A800 and H800 chips to China final month.
The California-based chipmaker is anticipated on Thursday to launch a minimum of three new synthetic intelligence (AI), the H20 L20, and L2, and maybe extra, to exchange the banned processors, media reported.
The efficiency density, or velocity per die dimension, of the three chips is lowered to between 14.9% and 26.8% of that of the H100. Nvidia slowed their speeds with some {hardware} and software program changes, based on know-how consultants.
The H100 is 6.68 instances quicker than the H20, know-how analyst Dylan Petal says in an article revealed by SemiAnalysis on November 9. Nonetheless, the H20 is 20% quicker than the H100 in massive language mannequin (LLM) reasoning, he added.
LLMs are deep studying algorithms that may acknowledge, summarize, translate, predict and generate content material utilizing very massive datasets, based on Nvidia’s web site.
Some Chinese language corporations had given up ordering Nvidia’s AI chips as they didn’t know when and whether or not their orders can be canceled amid the USA’ tightening chip export controls.
Baidu, China’s search engine, had already ordered 1,600 Ascend 910B chips from Huawei for about 450 million yuan (US$61.83 million) in August and obtained about 1,000 of them, Reuters reported on November 7, citing two unnamed sources.
One of many sources mentioned the Ascend processors at the moment are essentially the most refined AI chips out there in China, though they aren’t as quick as Nvidia’s.
“The H20’s total computing energy is simply equal to twenty% of that of the H100, which means that there’s room for worth reduce,” a Shanghai-based columnist writes in an article revealed on Monday. “Nonetheless, utilizing the H20 will nonetheless be extra pricey than utilizing China’s AI chips, comparable to Huawei’s 910B.”
The author says Nvidia will lose its competitiveness in China over the long term if it can not promote its most cutting-edge merchandise within the nation.
New parameters
In August final 12 months, the Biden administration ordered US chipmakers to cease exporting graphic processors that function at interconnect bandwidths of 600 gigabytes per second or above to China and Russia. Nvidia’s A100 and H100 chips and AMD’s MI250 chip are within the class affected by this rule.
Nvidia later unveiled the A800 and H800 processors, which work at 400 and 300 gigabytes per second respectively, concentrating on the Chinese language markets. Some analysts discovered that the A800 and H800 had been really lowered variations of the A100 and H100, respectively.
On October 17, the US Commerce Division’s Bureau of Trade and Safety (BIS) mentioned it is not going to categorize restricted chips through the use of “interconnect bandwidth” as a parameter. As a substitute, it would use “efficiency” and “efficiency density” as new parameters.
Below the brand new guidelines, a chip with a complete processing efficiency of 4,800 or extra or a efficiency density of 5.92 or extra will likely be banned from being shipped to China. A800, H800, L40, L40S and RTX 4090 chips are within the class of this rule.
China’s orders involving US$5 billion price of Nvidia chips have reportedly been canceled.
As of now, the H20, L20 and L2 can nonetheless be exported to China as they fulfill the efficiency and performance-density necessities. However they’re changing into unattractive to Chinese language corporations.
A Beijing-based author surnamed Huang in an article describes the H20, L20 and L2 because the “castrated variations” of the extra superior H100, AD102 and AD104 chips, respectively.
He says it’s price mentioning that the H20 is even slower than the entry-level A30 chip, which was launched in April 2021.
The H100 is designed for graphics-intensive workloads whereas the A100 is designed for high-performance computing (HPC) and AI workloads. The H100 is 2 instances quicker than the A100, which can be two instances quicker than the A30.
Jiang Tao, a senior vp of iFLYTEK, a Hefei-based AI answer supplier, mentioned on October 20 that the corporate makes use of Huawei’s Ascend 910B chips for computing. With out offering knowledge, he claimed that the chip has reached the benchmark of Nvidia A100.
iFLYTEK has been unable to buy American gadgets because it was added to the entity listing of the US in 2019. It was accused of supplying its surveillance gear to Xinjiang camps that detain Uyghurs and different ethnic minority folks.
Learn: Finish to decoupling tops China’s pre-summit calls for
Comply with Jeff Pao on Twitter at @jeffpao3
