NVIDIA A800

What the A800 is

The NVIDIA A800 is a data-center GPU — a processor built for servers that train and run large neural networks rather than for graphics on a desktop. NVIDIA put it into production in the third quarter of 2022 as an alternative to its A100 for customers in China ^[1]. It was not a worldwide product: NVIDIA sold the A800 only into China, Hong Kong and Macao ^[2].

Why it exists

Chips of this class became subject to export controls, and a vendor that wants to keep selling into a restricted market has to offer a part that stays below the control thresholds. NVIDIA said the A800 meets the US government's clear test for reduced export control and cannot be programmed to exceed it ^[1].

The design change that separated it from the A100 was the link between chips. On the A800 the chip-to-chip data transfer rate is 400 GB/s, against 600 GB/s on the A100, and NVIDIA described no other significant differences between the two GPUs ^[1]. That link matters in multi-GPU systems, where many accelerators work on one model and must constantly exchange data; the traffic between cards, not the arithmetic inside a single card, can become the bottleneck. Lowering the interconnect rate therefore changes a measurable number without changing how any one chip computes.

Specifications

The A800 shipped in three versions: 40 GB of HBM2 memory with a PCIe interface, 80 GB of HBM2e with PCIe, and 80 GB of HBM2e with an SXM interface ^[2]. HBM is high-bandwidth memory stacked close to the processor; PCIe is the standard server expansion slot, while SXM is NVIDIA's own module format for densely packed systems.

Compute was carried over from the A100: the card keeps 6,912 CUDA cores, 9.7 TFLOPS of FP64 and 19.5 TFLOPS of FP32 performance ^[2]. CUDA cores are the small arithmetic units that do the parallel work; FP64 refers to 64-bit floating-point arithmetic, used where results need more precision, and FP32 to 32-bit arithmetic, the usual mode for training networks.

Its memory bandwidth ranges from 1.6 to 2.0 TB/s, it uses a PCIe 4.0 system bus, and its thermal design power ranges from 50 to 400 W ^[2]. Those ranges come from the three versions: a passively cooled PCIe card that draws tens of watts and a densely packed SXM module that draws hundreds of watts are the same silicon configured for very different machines.

Availability in China

By November 2022 A800 cards were being sold on the Chinese e-commerce platform Taobao, with one seller in Shenzhen telling the Global Times that he held 1,000 units in stock for immediate delivery ^[1]. A data-center accelerator appearing on a consumer marketplace, in volume, says something about how quickly demand had formed around the part.

Where it sits

NVIDIA's China revenue was US$7.111 billion in the year before the A800 launch, 26.4% of its total ^[2]. A GPU designed for one market, with a deliberately narrowed interconnect and a stated inability to be programmed past the threshold ^[1], shows how large that market was for the company and how much engineering effort a vendor was willing to spend to keep serving it inside the rules.