NVIDIA’s massive A100 GPU, In this mini-episode of our explainer show, Upscaled, we break down NVIDIA’s latest GPU, the A100, and its new graphics architecture Ampere. Instead, it’s the direct replacement to the Volta-based V100 — a 2017 GPU purpose-built for data centers.
About The News:
NVIDIA’s massive A100 GPU was a little hazy on the finer details of Ampere, but what we do know is that the A100 GPU is huge. Its die size is 826 square millimeters, which is larger than both the V100 (815mm2) and the NVIDIA’s flagship gaming card, the RTX 2080 Ti (754mm2).
This 20x performance jump is partially due to the massive increase in cores, but architectural improvements and new ways of doing math (which we dive into in our video) are likely contributing much more. The A100 is also helped by its memory: It has 40GB of HBM2 memory, compared to the 16GB the V100 launched with (the company bumped the memory on Volta cards to 32GB later), which means each DGX A100 system has 320GB of VRAM to play with.
So what can this tell us about NVIDIA’s much-anticipated new gaming cards? Well, concretely, some of these AI improvements will find their way into GeForce cards, improving performance in upscaling tasks like DLSS, or denoising, which is a key aspect of ray-tracing.
In a briefing with reporters, NVIDIA CEO Jensen Huang all-but-confirmed that, while there’s “great overlap in the architecture” between Ampere and the upcoming consumer cards, these gaming cards won’t feature HBM2 memory, and the sizing of the different elements in the chips will be very different, as they’ll be more focused on graphics performance than high-precision math.
The current rumor du jour comes from the YouTuber Moore’s Law is Dead, who suggests that the anticipated “3080 Ti” will have around 30-percent more cores than the 2080 Ti, which could make for a flagship GPU on a much-more reasonable 450mm2 die. Will a smaller die mean GPU prices go down? Given those specs will likely put NVIDIA’s offerings ahead of AMD’s rumored new cards, we doubt it.