Nvidia announced its new GPU, the A100 a few days ago. Although calling it a GPU might be a bit of a stretch as this ‘GPU’ is mainly focused towards A.I. Computing and Scientific Computing, still it is based on the Ampere architecture so it might give us an hint about GeForce gaming card that may come out next.
Now graphics cards by their nature are designed to do a limited set of operations really quick and in parallel like the Million plus polygons your typical games uses in each frame, because of this people also realised that you can process a huge amount of data much more faster than a CPU and thus began the use of GPUs in servers and super-computers. And that is exactly what this A100 card is intended for.
Nvidia plans to sell this in their DGX-1 systems as sort of a pod that links 8 of these systems together and can be slotted into a server and sells for $199,999. But Nvidia says it would be worth it considering the performance the A100 can deliver.
The A100 is based on the Ampere architecture and it’s also manufactured on the 7nm process, now moving to to a lower nanometer process from its earlier 12nm process should give them better energy efficiency and enable them to cram in more transistors on each chip and that’s exactly what they did, Nvidia claims that the A100 packs 54 Billion transistors into a 826 sq. millimetre dye, for context Nvidia’s current gaming card the RTX 2080ti has 18.6 Billion transistors in a 754 sq. millimetre dye. This can potentially mean that the A100 packs as much as 3 times the performance as this top of the line gaming card.
Comparing the A100 to a RTX 2080ti might be the wrong call here, Nvidia has really set the A100 to be the successor of the V100 its other Data center GPU released in 2018, being a very impressive GPU in its own right the V100 retailed for around $9,000. The V100, although plenty fast, Nvidia claims that the newer A100 will be as much as 20 times faster than its predecessor. Of course all this processing wouldn’t mean much if it couldn’t move data fast enough, but with 14GB of HBM2 (High Bandwidth Memory) capable of moving 1.5TB per second i.e., 40 times faster than a High End CPU server, it should be fine.
In addition to the 8 A100 GPUs the DGX system also sports a 15TB Gen4 NVME flash storage, a Terabyte of RAM and 2 64-core AMD Rome CPUs. So what does this card tell us about the next graphics cards? Well the Ampere architecture has been positioned as the successor of the Volta architecture, but Volta never spawned any consumer cards and the current RTX series graphic cards are based on the Turing architecture, but Volta did introduce Tensor cores that are specialised for A.I. math that did find its way into consumer cards. And it is possible that the A.I. improvements seen in the Ampere card would also make its way into the gaming cards, helping with A.I. based Upscaling like DLSS or in De-Noising that is used in Ray Tracing.
So will it be possible to get your hands on new GPU as fast as the A100? Probably not, but with the technologies used in this card we might start seeing future cards getting significantly faster with more transistors and better A.I. for that sweet Ray tracing.