Conclusion

We haven't yet tested NVIDIA's Ampere graphics cards, but NVIDIA's product reveal on September 1 was one of many surprises. We've been hearing of leaked specifications, including from board partners, of certain CUDA core counts and TFLOPs numbers, which led to the believe that Ampere could be an uneventful, incremental upgrade to Turing, much like Pascal was to Maxwell, with the new sub-10 nm node significantly dialing up energy efficiency. Boy, were we wrong.

Our first revelation perhaps was NVIDIA's decision to bet big on Samsung 8 nm node instead of the company's mainstay, a TSMC node such as the N7 on which the company is already making the A100. The next big revelation came from the gargantuan increase in CUDA cores, which NVIDIA claims to have double the performance generation over generation, on the backs of sheer numbers. We're inclined to give NVIDIA's performance claims the benefit of doubt, although we certainly intent to test everything in our upcoming performance reviews.

Ampere is more than just generational performance growth. Improvements to the RT core—now in its 2nd gen—introduce new raytracing performance optimizations, while the 3rd gen Tensor core leverages sparsity to drive up AI inference performance, which should impact performance as AI is used by NVIDIA for RTX de-noising and DLSS. The new GDDR6X memory narrows the gap between GDDR and HBM standards for the first time, without insane increases in bus width. PCI-Express Gen 4 may come in handy for enthusiasts benchmarking RTX 3090 SLI, as each card would have the IO bandwidth equivalent of Gen 3 x16.

As we mentioned earlier, NVIDIA doesn't see its responsibility toward gamers end with just graphics hardware and wants to solve some fundamental problems with the modern PC. The RTX IO technology seeks to vastly reduce the impact of I/O overhead of fast new SSDs on your CPU, building on Microsoft's DirectStorage API, while Reflex seeks to give you a competitive edge in eSports by solving latencies at the system level, something that's more within your control than your network ping. The company also wants to democratize 3D animation filmmaking without making people learn 3D from scratch—the Omniverse Machinima is basically Ansel for moving images, and its possibilities for storytellers are endless.

All in all, we are thoroughly appetized for GeForce RTX 3000 Ampere, and we can't wait to tell you more about the way it performs, what you can expect for your money, and whether it's time to upgrade.