NVIDIA A100 “Ampere” GPUs With 7936 Cores Being Sold In China, 15% More Cores Than Original A100s

NVIDIA seems to have been shipping A100 "Ampere" GPUs with higher core counts than the original spec in China as per recent retail listings.

It's no surprise that the Chinese marketplaces are filled with curious products for us to look at and explore, such as the instance when we saw mobile variants of NVIDIA's flagship Ada Lovelace SKUs being "carved" into desktop models using an adapter board or prototype coolers of NVIDIA's RTX 4090 GPU.

Now, Chinese marketplaces have seen the emergence of new variants of NVIDIA's A100 AI GPUs, with specifications that are better than the original design.

All kinds of 7936SP A100s are surfacing one after another. Kinda surprised that at least some of tjese 7936sp A100s were made back in 2020 and are only surfacing now?

Credit: https://t.co/dpfpLUm1Cf https://t.co/ls8LSJcSiS pic.twitter.com/CUswegFKzP

— Jiacheng Liu (@unnatural__log) April 14, 2024

Before we go into the mystery, let's take a quick recap on NVIDIA's A100s. The GPU comes in two different configurations based on the Ampere architecture, with the 40 GB and the 80 GB variants, both coming in with their respective differences in the memory and speed departments. Both GPU offerings feature 6912 CUDA cores, but in the case of the newly-emerged models, the core count comes at 7936 which are 15% more cores than the previous variants and 2 fewer SMs versus the full GA100 GPU (8192 cores).

The new NVIDIA Ampere A100 GPUs also feature 96 GB of HBM2E memory along a 6144-bit bus interface which means we are looking at six 8-hi stacks with 16 GB capacities per module. Based on the pictures provided in the listing, the graphics card comes in PCIe form factor and is a prototype unit that was originally designed for the GRID system.

The card features three 8-pin connectors and a massive GPU that sits underneath a huge IHS. The memory chips are not on the PCB but rather on the GPU package itself since the card is based on HBM and not the traditional GDDR DRAM which is seen on consumer-grade graphics cards.

Clock speeds are rated at 1260 MHz base and the boost clock is not mentioned but the GPU does produce around 20 TFLOPs of Single-Precision and 10 TFLOPs of Double Precision compute. For reference, the A100 PCIe 80 GB produces 19.5 TFLOPs of Single-Precision and 9.74 TFLOPs of double precision compute.

Now the reason behind the difference here isn't quite exactly confirmed, but this could either be a move by NVIDIA for specific clients, probably providing them a customized solution that has surfaced into the markets now although the GPUs were produced back in 2020. A few listings of the same A100 96 GB model can also be found on eBay here and here. The GPU goes for around $17 to $20K.

News Source: @unnatural__log