Find out how NVIDIA created the new A800 GPU to bypass the US ban on selling advanced chips to China!
NVIDIA offers A800 GPU to bypass US ban on China!
Two months after it was forbidden Because the US government has blocked China from selling high-performance AI chips, NVIDIA has introduced a new A800 GPU designed to bypass those restrictions.
The new NVIDIA A800 is based on the same Ampere microarchitecture as the A100, which was used as a performance baseline by the US government.
Despite its numerically larger model number (the lucky number 8 was probably chosen to appeal to the Chinese), this is an out-of-tune piece, with slightly reduced performance to comply with export control limitations.
The NVIDIA A800 GPU, which entered production in the third quarter, is another alternative product to the NVIDIA A100 GPU for customers in China.
The A800 meets the U.S. government's clear test for reduced export control and cannot be programmed to overcome it.
NVIDIA probably expects the A little slower The NVIDIA A800 GPU will enable it to continue supplying China with A100-level chips used to power high-performance supercomputers and data centers for artificial intelligence applications.
As I'll show you in the next section, except for high-end applications, there won't be a really significant performance difference between the A800 and the A100. So NVIDIA customers who want or need the A100 will have no problem opting for the A800.
However, this can only be a stopgap solution, as NVIDIA is forced to sell A100-level chips to China until the US government changes its mind.
Read more : AMD and NVIDIA banned from selling AI chips to China!
How fast is the NVIDIA A800 GPU?
The US government considers the NVIDIA A100 to be the performance baseline for its export control restrictions to China.
Any chip that is equal to or faster than the Ampere-based chip, which was launched on May 14, 2020, is prohibited from being sold or exported to China. But, as they say, the devil is in the details.
The US government hasn't specified how much slower the chips need to be to qualify for export to China, so NVIDIA could technically get away with slightly downgrading the A100 while still offering nearly the same level of performance.
And that’s what NVIDIA did with the A800: it’s basically the A100 with 33% slower NVLink interconnect speed. NVIDIA also limited the maximum number of GPUs supported in a single server to 8.
That only slightly reduces the performance of the A800 servers, compared to the A100 servers, while offering the same Amount of GPU processing performance. Most users won't notice the difference.
The only significant impediment is at the highest end: Chinese companies are now limited to a maximum of eight GPUs per server, up from up to sixteen.
To show you what I mean, I looked at the A800's specs and compared them to the A100 below:
NVIDIA A100 vs A800: 80GB PCIe version
Specifications | A100 80GB PCIe | A800 80GB PCIe |
FP64 | 9.7 TFLOPS | |
FP64 tension core | 19.5 TFLOPS | |
FP32 | 19.5 TFLOPS | |
Tension float 32 | 156 TFLOPS | |
BFLOAT 16 tension core | 312 TFLOPS | |
FP16 tension core | 312 TFLOPS | |
INT8 Tensor Core | 624 STOPS | |
GPU Memory | 80GB HBM2 | |
GPU Memory Bandwidth | 1.935 GB/s | |
TDP | 300 watts | |
Multi-instance GPU | Up to 7 MIG to 10 GB | |
Interconnect | NVLink: 600 GB/s PCIe Gen4: 64 GB/s | NVLink: 400 GB/s PCIe Gen4: 64 GB/s |
Server Options | 1-8 GPU |
NVIDIA A100 vs A800: 80GB SXM version
Specifications | A100 80GB SXM | A800 80GB SXM |
FP64 | 9.7 TFLOPS | |
FP64 tension core | 19.5 TFLOPS | |
FP32 | 19.5 TFLOPS | |
Tension float 32 | 156 TFLOPS | |
BFLOAT 16 tension core | 312 TFLOPS | |
FP16 tension core | 312 TFLOPS | |
INT8 Tensor Core | 624 STOPS | |
GPU Memory | 80GB HBM2 | |
GPU Memory Bandwidth | 2.039 GB/s | |
TDP | 400 watts | |
Multi-instance GPU | Up to 7 MIG to 10 GB | |
Interconnect | NVLink: 600 GB/s PCIe Gen4: 64 GB/s | NVLink: 400 GB/s PCIe Gen4: 64 GB/s |
Server Options | 4/ 8 / 16 GPU | 4/8 GPU |
NVIDIA A100 vs A800: 40GB PCIe version
Specifications | A100 40GB PCIe | A800 40GB PCIe |
FP64 | 9.7 TFLOPS | |
FP64 tension core | 19.5 TFLOPS | |
FP32 | 19.5 TFLOPS | |
Tension float 32 | 156 TFLOPS | |
BFLOAT 16 tension core | 312 TFLOPS | |
FP16 tension core | 312 TFLOPS | |
INT8 Tensor Core | 624 STOPS | |
GPU Memory | 40GB HBM2 | |
GPU Memory Bandwidth | 1.555 GB/s | |
TDP | 250 watts | |
Multi-instance GPU | Up to 7 MIG to 10 GB | |
Interconnect | NVLink: 600 GB/s PCIe Gen4: 64 GB/s | NVLink: 400 GB/s PCIe Gen4: 64 GB/s |
Server Options | 1-8 GPU |
Please support my work!
Support my work via bank transfer/paypal/credit card!
Name: Adrian Wong
Wire transfer : CIMB7064555917 (Swift code: CIBBMYKL)
Credit Card / Paypal: https://paypal.me/techarp
Dr. Adrian Wong has been writing about technology and science since 1997, and even published a book with Prentice Hall called Breaking the BIOS barrier (ISBN 978-0131455368) While I was in medical school.
He continues to devote countless hours each day to writing about technology, medicine and science, in his quest for facts in a post-truth world.
Recommended reading
Support Tech ARP!
Please support us with Visiting our sponsorsparticipating in the ARP Technology Forumseither donating to our fund. Thank you!
Leave feedback about this