stock Quadro RTX 6000 being twice as fast as an overclocked 3080 in an dual gpu F@H PC?

[Sanedish]

The system specs are:

Ryzen 9 5950X

Zotac RTX 3080 Trinity OC LHR

Stock Nvidia Quadro RTX 6000 (see picture)

BeQuiet! Dark Power 13 1000 Watts

 

 

The Quadro:

 

There where no modifications done to it.

[WereCat]

If F@H requires a lot of VRAM then yea, 24GB is more than 10GB/12GB on 3080 and will significantly impact performance if you need more of it.

[Sanedish]

26 minutes ago, WereCat said:

If F@H requires a lot of VRAM then yea, 24GB is more than 10GB/12GB on 3080 and will significantly impact performance if you need more of it.

ik, but vram was usually quite chill

[Kazi]

What about PCI E slots / lines.

For example my B550 board have three PCI Ex16 slots (physically) but only one of them is PCI E 4.0 x16 from CPU other two are PCI E 3.0x4 from chipset.

[Sanedish]

5 minutes ago, Kazi said:

What about PCI E slots / lines.

For example my B550 board have three PCI Ex16 slots (physically) but only one of them is PCI E 4.0 x16 from CPU other two are PCI E 3.0x4 from chipset.

Both PCI E 4 slots, the 3080 was in the upper slot

[Gorgon]

4 hours ago, Sanedish said:

The system specs are:

Ryzen 9 5950X

Zotac RTX 3080 Trinity OC LHR

Stock Nvidia Quadro RTX 6000 (see picture)

BeQuiet! Dark Power 13 1000 Watts

 

The Quadro:

There where no modifications done to it.

 

An A6000 (Ada) should Yield about 12.9MPPD on average versus 4.8MPPD for a 3080. Not too surprising given the 10,752 versus the 8,704 Shader Cores, the Quick Return Bonus (QRB) and having similar Base Clocks and being of the same generation.

 

As for VRAM, Folding at Home only uses 1-2GB of VRAM when running and most cards see little to no decrease in performance running in PCIe slots at x8 versus x16. Currently F@H only leverages the Shader Cores so Tensor and RT Cores have no impact.

 

Typically the FP32 performance of a card within the same generation (or at least since NVIDIA changed their definition of this in Turing) will sort of (due to the QRB) directly translate to the performance difference. Better FP64 performance will make a lesser impact in some work units (WUs) but both AMD and NVIDIA have "nerfed" their FP64 performance in recent generations.

[Sanedish]

12 hours ago, Gorgon said:

 

An A6000 (Ada) should Yield about 12.9MPPD on average versus 4.8MPPD for a 3080. Not too surprising given the 10,752 versus the 8,704 Shader Cores, the Quick Return Bonus (QRB) and having similar Base Clocks and being of the same generation.

 

As for VRAM, Folding at Home only uses 1-2GB of VRAM when running and most cards see little to no decrease in performance running in PCIe slots at x8 versus x16. Currently F@H only leverages the Shader Cores so Tensor and RT Cores have no impact.

 

Typically the FP32 performance of a card within the same generation (or at least since NVIDIA changed their definition of this in Turing) will sort of (due to the QRB) directly translate to the performance difference. Better FP64 performance will make a lesser impact in some work units (WUs) but both AMD and NVIDIA have "nerfed" their FP64 performance in recent generations.

Mind we're talking about an Quadro RTX 6000

[Gorgon]

8 hours ago, Sanedish said:

Mind we're talking about an Quadro RTX 6000

Then it likely you either have your numbers backwards or you are comparing Apples to Oranges (i.e. NOT the same Work Units). It is unlikely a Quadro RTX 6000 with 4608 Turing CUDA Cores is double the performance of a RTX 3080 with 8704 Ampere Cores. The LAR Database has it at 3.5 versus 4.8MPPD for the 3080 and 16.31 versus 29.77 TFLOPS (FP32).

 

Without knowing which Projects (Work Units) were running on each card when you observed this it's difficult to say what the cause might be but it is likely that you observed the 3080 running a poorly performing Work Unit while the RTX 6000 was running a WU that performed adequately or above average.

 

If you want to do a meaningful comparison then you'd have to use a tool like Harlam's Folding Monitor (HfM.net) to monitor the GPUs for a several days then compare the performance of the two cards running the same Projects and discounting projects with small (<50,000) Atom Sizes.

 

For example, one of my 3080s, running between January 1st and the end of February, when it was decommissioned and replaced with a 4070 Ti Super, had PPDs ranging from 4.3 (p12446) and 7.7MPPD (p12264) Clock-Limited to 1440MHz so a 3080 running at Stock should be 15-20% more performant.