Nvidia Ampere A100 announced

Uttamattamakin · May 14, 2020

3 hours ago, leadeater said:

For those that care about the GPU architecture details itself.

(Chart showing the technical specs)

https://www.anandtech.com/show/15801/nvidia-announces-ampere-architecture-and-a100-products

https://devblogs.nvidia.com/nvidia-ampere-architecture-in-depth/

Wow... I was so much wanting a 2080 to get those tensor cores to use for my work that I would quite frankly have given a good time to whoever would be willing to donate one for my science. For the ... 3080 or whatever it will be called. For marketing I could see them saying this thing has over TWICE the transistor count of the previous generation so we're going to call it the 4090. For that I'd have someone's baby.
Before I say that though

I have to wonder how much it would cost? Double the price of a 2080 TI?

Sweet God one could do a high resolution fully realistic similation of the first stages of the big bang accounting for full General Relativity with the tensor cores, and particle physics interactions with the CUDA cores. While also watching the whole thing unfold in VR.

Results45 · May 14, 2020

.

Nvidia: has built and run a private multi-exaflop supercomputer for the past 1-2 years

Also Nvidia: uses Epyc Rome CPUs in DGX A100 servers

Intel: double roasted

leadeater · May 14, 2020

1 hour ago, Results45 said:

Also Nvidia: uses Epyc Rome CPUs in DGX A100 servers

Yea I noticed that, the DGX V100 used Intel but not this time

leadeater · May 14, 2020

4 hours ago, Inkz said:

Just curious do you guys think there is anything we can infer about their gaming series from this? Like how previous generations have compared for the professional cards?

Only some of the architecture efficiency benefits and the area to transistor density increase gains. GV100/Volta only told us a small amount about Turing and this will be the same. The fine details about the SMs in the GV100 (Gx100) dies is completely different to the lesser dies in the stack, a lot different.

What this does tell us is that the next gaming GPU dies won't have more than the number of CUDA cores seen here. Just as an FYI the GA100 die actually has 8192 CUDA cores in 128 SMs but actual products only have 108 SMs active so 6912 CUDA cores. The remaining 20 SMs are being used for binning/yield so we may even see a refresh in a year or so to something like GA200 with more SMs active as the manufacturing process gets more refined and yields increase.

TU102 is a bit under 100mm2 smaller than GV100, had 4608 CUDA cores rather than 5120 and 576 Tensor cores compared to 640. Also keep in mind Ampere has no RT cores and actually has less Tensor cores than Volta. A smaller die will need less SM headroom for yield so a higher ratio will be active.

Without doing a better evaluation I'd spit ball about 6000 CUDA cores and 300-350 Tensor cores for the next gaming architecture with an unknown amount of RT cores as I couldn't even begin to guess on that one.