NVIDIA has announced a trio of brand new data center GPU accelerators which are designed for the China HPC & AI segment and bypass the chip ban.
NVIDIA To Resume Advanced Chip Shipments To China With The A800 Data Center GPU Accelerator
Back in August, it was reported that the US Government had halted all of its top chip makers, which includes NVIDIA and AMD, from selling their GPUs to China and Russia. The chip ban on China was not only to halt the supplement of high-end GPUs but those chips which can be used for a military advantage. NVIDIA's H100 and A100 being the world's top AI and HPC chips are already being used by the US Military for advanced military & scientific purposes so anyone who possesses these chips possesses an advantage on the field.
"The Nvidia A800 GPU, which went into production in Q3, 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 exceed it,"
a Nvidia spokesperson said in a statement to Reuters.
The main chip supply that was affected by this included AMD's Instinct MI200 lineup and NVIDIA's A100 "Ampere" and H100 "Hopper" accelerators. Now customers did have a short window where they could still procure a small supply of these chips from Hong Kong but that wasn't going to last forever.
With the latest report, it looks like NVIDIA has come up with an interesting solution. Since the H100 and the A100 are specifically banned, NVIDIA has designed a brand new GPU accelerator known as the A800. The new GPU features a slightly toned down design compared to the A100 but is made to carter the China AI and HPC market.
Based on the specifications, the NVIDIA A800 will be utilizing the same chip architecture as the Ampere A100 GPU. It will come in three variants, two PCIe variants at 40 GB & 80 GB and an SXM variant at 80 GB. The GPUs will offer up to 9.7 TFLOPs of FP64, 19.5 TFLOPs of FP64 Tensor Core, 19.5 TFLOPs of FP32, 156 TFLOPs (312 TFLOPs with sparsity) TF32, 312 TFLOPs (624 TFLOPs with Sparsity) BFLOAT16 and 624 TOPS (1248 TOPs with Sparsity) INT8 performance. The 40 GB GPU features HBM2 memory with up to 1.555 TB/s bandwidth while the 80 GB variants feature HBM2e with up to 2 TB/s bandwidth.
The models range from 250W (40 GB PCIe), 300W (80 GB PCIe), and 400W (80 GB SXM) TDPs. As expected, the PCIe models come with a dual-slot air or single-slot air-cooled design while the SXM variant features a passively cooled heatsink. NVLINK is limited to 2 GPUs for the PCIe variants and the SXM variant is limited to 400 GB/s. The limitation that comes with NVLINK speeds will mean that there won't be enough bandwidth to be utilized between multiple GPUs so you won't get as near as the real thing (A100). NVIDIA will be providing the PCIe A800 GPUs in 1-8 card partners or NVIDIA-certified systems while the SXM variant will be available through NVIDIA HGX-certified or partner systems.
"China is a significant market for Nvidia and it makes ample business sense to reconfigure your product to avoid trade restrictions,”
The China GPU market is said to amount to $400 Million worth of chip sales as NVIDIA puts it and can be seen as a significant chunk of their market share.
News Source: Reuters
