MishaEngel wrote: For now the specs .of the new NVidia cards look promising, let's hope they turn out to be great.
Hi.
I have problems to find the fp32 performance of the new nVidea Ampere Graphics card, to find out how fast they will be in Resolve. I can find a couple of links stating the fp32 performance. But as I remember, did I see the same numbers as rumors from the same sites. And think they just kept the numbers they already had as rumors. So I don't think that I can trust these numbers. I am looking for a fp32 performance number we can trust on.
But here is a link, I will like the share with you:
QUOTE: 'SwAY256 - Tuesday, September 1, 2020
I strongly suspect that the "Shader FLOPS" are not "2x FP32" but in fact BFLOAT16 FLOPS as seen in the A100 Ampere GPU. If there were really FP32 FLOPS, it would be labeled as FP32. As Microsoft is always talking about TFLOPS for the next xbox, I guess that NVIDIA wanted to respond.
That means that the 10k CUDA cores are a marketing trick to say that there are 5k CUDA Cores that are capable of working on floats that have the same dynamic as FP32 but far less precision (BFLOAT16).
If that's correct, I don't know the impact of using BF16 instead of FP32 in game rendering.
Anonomouse - Tuesday, September 1, 2020
Nvidia's website literally says "2x FP32 throughput" (
https://www.nvidia.com/da-dk/geforce/gr ... wmfg_en-us)
SwAY256 - Wednesday, September 2, 2020
Yes but that's what they said for A100. Turned out that it was a "mixed precision FP32" (BF16 in fact). Given that A100 has 19.5 FP32 TFLOPS and 39 BF16 TFLOPS, I wouldn't be surprised that GF3090 has 18/36.'
From:
https://www.anandtech.com/show/16057/nv ... 0-rtx-3090User Comments page 26.
If SwAY256 are right, will the RTX 3090 have a Resolve performance of 18 fp32 TFLOPS.
What we need is Independent Reviews with the correct fp32 performance.
Regards Carsten.