Nvidia Revs North American Channel To Sell Workstation Version Of Banned-In-China A800 GPU
The AI chip giant says it planned to introduce the new A800 40GB Active for workstations ‘regardless’ of recent U.S. export restrictions that banned the sale of the A800 server GPU in China.
[Editor’s note: This story was updated on Nov. 8 to reflect clarifications and commentary on the difference between the A800 40GB Active for workstation PCs and A800 for servers. The information was provided by Nvidia to CRN after publication of the original version.]
Nvidia is getting partners in North America and other regions set up to sell a tweaked version of the A800 GPU, an AI chip originally designed to sidestep U.S. export restrictions against China before new rules halted sales to the country last month.
Nvidia partners, including U.S.-based electronics manufacturer PNY Technologies and system integrator Colfax International, have begun promoting this week the introduction of the Nvidia A800 40GB Active PCIe card, which the chip designer is calling the “ultimate workstation development platform for AI, data science and high-performance computing” on its American website.
PNY began selling the Nvidia A800 40GB Active GPU on Monday through partners in North America, Latin America, Europe,Africa and India, a spokesperson told CRN. Excluded countries include China and dozens of others, such as Russia and most Middle Eastern countries (full list below).
Other partners openly promoting the A800 40GB Active include Japan-based ASK Corp. and Elsa.
A U.S. distribution executive told CRN he expects the A800 to start shipping in the next few weeks.
“I expect they will sell out fairly quickly given the demand for Nvidia’s high-end AI GPUs,” said Kent Tibbils, vice president of marketing at Fremont, Calif.-based ASI. “Overall, the AI/ML market is still driving server growth across multiple markets, and we expect this to continue through 2024.”
Nvidia product marketer Allen Bourgoyne told CRN that the A800 40GB Active is a follow-up to 2018’s Quadro GV100 workstation GPU, which the company no longer produces.
While the A800 40GB Active uses the same GPU that serves as the basis for the Nvidia A800 server GPU introduced to Chinese customers a year ago, the company intended to introduce the A800 40GB Active to a wider customer base “regardless of what happened to China,” according to Bourgoyne.
“It’s just using a chip family that originally was targeted for data center, and we were able to apply that to a workstation design,” said Bourgoyne.
The A800’s History: Built For China, Then Banned A Year Later
Nvidia originally designed the A800 to satisfy U.S. export restriction rules set against China for AI chips in October 2022 after the American government ordered the company to stop selling its most powerful data center GPUs, the A100 and H100, to customers in the Asian country.
The main goal of the export restrictions, set by the U.S. Department of Commerce, is to prevent China from gaining access to state-of-the-art technologies to boost its military.
Using the same Ampere architecture that powers the A100, Nvidia introduced the A800 to Chinese customers a year ago, sidestepping U.S. export restrictions at the time by designing the GPU to offer a chip-to-chip data transfer rate lower than the threshold for AI chips targeted by the sanctions.
Whereas the A100 has a chip-to-chip bandwidth of 600 GB/s, which is the minimum threshold for chips banned by U.S. export restrictions against China, the A800’s chip-to-chip bandwidth falls below that threshold with only 400 GB/s.
Nvidia subsequently introduced the H800 as an alternative to the company’s most recent flagship data center GPU, H100, and like the A800, it has a chip-to-chip bandwidth under 600 GB/s.
Rivals Intel and AMD reportedly pursued similar strategies after the U.S. export restrictions prevented the companies from selling powerful AI chips into China.
Last month, the U.S. government effectively banned Nvidia from selling the A800 and H800 into China by expanding export restrictions to include AI chips that exceed a certain performance level when multiple chips are connected within a system, which is critical to training increasingly large AI models.
The rules also impacted Nvidia’s recently launched L40S GPU as well as chips from Intel and AMD.
The Commerce Department said the new performance density parameter is meant to prevent companies from introducing workarounds that would allow Chinese firms to purchase a “larger number of smaller datacenter AI chips which, if combined, would be equally as powerful as restricted chips.”
The new U.S. rules enacted in October also restricted the export of advanced chips to a broader set of countries, including Iran and Russia, according to Reuters.
The Wall Street Journal reported on Monday that the expanded export restrictions resulted in Nvidia cancelling more than $5 billion in AI chip orders set for Chinese customers next year.
An Nvidia spokesperson told the newspaper that the chip designer was in the process of reallocating supply of impacted AI chips, such as the A800, to the U.S. and other regions.
Nvidia Pitches A800 40GB PCIe Card As ‘Ultimate’ Workstation Platform
Nvidia’s A800 40GB Active product is an actively cooled dual-slot PCIe card that leverages “the same GPU that was used in the original A800 server product,” according to Bourgoyne.
Called the “ultimate workstation development platform for AI, data science and high-performance computing,” the A800 40GB Active GPU is designed to “bring the power of a supercomputer to your workstation and accelerate end-to-end data science workflows,” according to Nvidia’s website.
Bourgoyne said the Nvidia product team for the A800 40GB Active decided to use the same GPU as the A800 server product based on which chips were available that could meet the product’s requirements, which included a need to provide fast double-precision performance for HPC applications.
“They have to look at, here’s all the parts you have available, which of those will meet all your constraints from a physical, electrical, cooling [and] price [standpoint]—all of those things—and then they select the right parts,” he said.
Nvidia’s A800 40GB Active PCIe card for workstations shares several of the same specifications as Nvidia’s A100 40GB PCIe card for servers, such as 6,912 CUDA cores, 432 Tensor cores, 40GB of high-bandwidth HBM2 memory and a 240-watt maximum power consumption.
The A800 is also capable of achieving the same 9.7 teraflops in double-precision performance and 19.5 teraflops in single-precision performance as the A100’s PCIe and SXM form factors.
The main difference is the rate at which the A800 can communicate with other A800s, with the GPU featuring an NVLink chip-to-chip bandwidth of 400 GB/s versus the A100’s 600 GB/s.
In marketing materials, Nvidia compared the A800 40GB Active GPU to the Quadro GV100 PCIe card and said the former is 4.2 times faster for AI inference with the BERT Large model, 90 percent faster for AI training with the BERT Large model, 90 percent faster for the GTC benchmark and 70 percent faster for the LAMMPS benchmark.
Nvidia said the A800 40GB Active GPU, like other AI chips in its portfolio, comes with a three-year subscription to Nvidia AI Enterprise, the company’s software suite that includes AI frameworks, libraries, pre-trained models and tools for developing and running AI applications.
PNY’s Full List Of Excluded Countries For A800 Active GPU
PNY’s full list of excluded countries for the Nvidia A800 Active GPU are as follows:
Afghanistan, Armenia, Azerbaijan, Bahrain, Belarus, Burma, Cambodia, Central African Republic, China, Democratic Republic of Congo, Cuba, Cyprus, Egypt, Eritrea, Georgia, Haiti, Iran, Iraq, Jordan, Kazakhstan, North Korea, Kuwait, Kyrgyzstan, Laos, Lebanon, Libya, Macau, Moldova, Mongolia, Oman, Pakistan, Qatar, Russia, Saudi Arabia, Somalia, South Sudan, Republic of Sudan, Syria, Tajikistan, Turkmenistan, United Arab Emirates, Uzbekistan, Venezuela, Vietnam, Yemen and Zimbabwe.