Nvidia GPUs with nearly 8,000 CUDA cores spotted in benchmark database (updated)

[Darth Shiv]

under 0% - AMD

Spicy!

 

[Morris Minor]

Lol

 

[mongeese]

Since the last few (read: most) comments aren't particularly topic, let's try to bring it back into focus.


It's also possible that the mystery GPU is a GV100 (Volta) replacement, in which case it would be a Tesla model, rather than a GeForce or Quadro. The V100 add-in card has a base clock of 1.23 GHz and the HBM2 is clocked to 0.876 GHz, so the indicated clocks in the unknown GPU aren't wildly different from these.

Just reference, this is the same test done using a Titan X (Pascal):

System manufacturer System Product Name - Geekbench

Benchmark results for a System manufacturer System Product Name with an Intel Core i7-9700K processor.

browser.geekbench.com

Finding a random Tesla V100 result, that's done using a 6 core/12 thread CPU, gives:

Google Google Compute Engine - Geekbench

Benchmark results for a Google Google Compute Engine with an Intel Xeon processor.

browser.geekbench.com

The results are a little down on the unnamed device in some area, ahead in others:

System manufacturer System Product Name - Geekbench

Benchmark results for a System manufacturer System Product Name with an Intel Core i7-8700K processor.

browser.geekbench.com

Sobel 46 vs 69.3 Gpixels/sec
Canny 7.39 vs 12.0 Gpixels/sec
Stereo Matching 890.4 vs 873.2 Gpixels/sec
Histogram Eq 26.0 vs 30.5 Gpixels/sec
Gaussian 10.6 vs 16.2 Gpixels/sec
DoF 5.85 vs 7.38 Gpixels/sec
Face Detection 302.9 vs 307.0 images/sec
Horizon Detection 3.68 vs 5.44 Gpixels/sec
Feature Matching 0.899 vs 5.44 Gpixels/sec
Particle Physics 23082 vs 19714 fps
SFFT 1.86 vs 2.60 Tflops

The Feature Matching results are vastly different, though. Here's another V100 result, but this time with a far larger CPU (48 cores):

Dell Inc. PowerEdge R740 - Geekbench

Benchmark results for a Dell Inc. PowerEdge R740 with an Intel Xeon Platinum 8268 processor.

browser.geekbench.com

It would seem that the CPU, unfortunately, has a significant impact on the CUDA Compute results. making it difficult to properly compare these new findings.

Yeah you could be right about the Tesla thing. I'm starting to think it has too many cores for GeForce. So maybe a big Tesla card for all this exascale and AI nonsense.

So if we're going to really get into it. The GV100 is an 84 SM card with four SMs always disabled for manufacturing tolerances. We'd expect this card to have 128 SMs, maybe 8 GPCs with 8 TPCs, but subtract four, so 124 SMs. Which is what there is. Also 1.2 GHz memory implies HBM2 on a 4096-bit bus which is Tesla only right now. And third, assuming there are 32 FP64 cores per SM, then this card would have about 13 TFLOPS at ~1.6 GHz which is the calculation from the Big Red 200 GPU.

On the other hand Volta cards are exclusively 80 SM, whereas this thing is appearing in all sorts of configs.

It's curious.

On the topic of clock speeds, the GV100 is available with base clocks ranging from 936 MHz to 1312 MHz, and boost clocks from 1290 MHz to 1637 MHz. So I'd expect the GA100 to have clock speeds covering a similar or greater range.

 

[neeyik]

On the topic of clock speeds, the GV100 is available with base clocks ranging from 936 MHz to 1312 MHz, and boost clocks from 1290 MHz to 1637 MHz. So I'd expect the GA100 to have clock speeds covering a similar or greater range.

Geekbench sadly only reports the default base/boost clock and not the real-time one - for example, in the Titan X (Pascal) link I gave, it states 1.53 GHz but it actually ran at 1.89 GHz during the tests, almost 24% faster. Mind you, if all these numbers are genuine, this new chip will be absolutely monstrous in transistor count, so I wouldn't expect it to be much, if at all, past 1.4 GHz.

 

[ChickyBoom]

Yeh, but can it run... nah screw that. Will it have drivers that won't crash GTA5??

 

[Gus Fring]

Maybe they've mutated specialized RTX cores into more general purpose cores, so that if not needed as RTX they can be put to use in more general Gfx/data/ML work. All that data we spew out has to be wrangled into *something* ??

 

[treetops]

I have a 2080 Ti which I got used for much less than retail 4 months ago and Netflix app always plays 4K, you can check the resolution by pressing Ctrl + Shift + Alt + D. You need the HEVC codec from app store.

I will skip the 3080Ti as my OCed 2080Ti can run anything maxed at 4K60Hz which is the max of my monitor. Also it will be much more expensive vs 2080Ti because of no competition and lack of stock due to corona virus. I will upgrade in 5 years when I get a 4K120Hz OLED(?) monitor.

Hey do you know if Hulu has a similar command, can't find any info.