Three different PCI info for the same GPU model

[KtC]

CPUs.txt specifies three different PCI info for one model of GPU:
- 0x10de:0x1e84:2:8:TU104 [GeForce RTX 2070 SUPER] 8218
- 0x10de:0x1ec2:2:7:TU104 [GeForce RTX 2070 SUPER]
- 0x10de:0x1ec7:2:7:TU104 [GeForce RTX 2070 SUPER]

Which one is the best for folding and why does one of them have "8218" number at the end?

[Joe_H]

The only one that matters for you is the one that has the same PCI device ID as the card installed in your system.

As for the entries that have a trailing number, that is the approximate processing power in GFLOPS, used in trying to group GPUs a bit.

[ajm]

I would say same.
The #2 and #3 are exactly the same (probably a typo in GPUs.txt). And 8218 is just the base GFLOPS of the 2070S. I suppose there is a additional entry for it to fit with some specific manufacturer's use of the ROM indications.

[KtC]

Soo, the first entry was probably added first and that is why have this number specified, or?

Is it possible not all listed devices are participating in the same projects? For example I saw on the forum people asking to assign GPUs for some project (?). Not sure how that is related to the PCI dev ID.

[uyaem]

Is it possible not all listed devices are participating in the same projects? For example I saw on the forum people asking to assign GPUs for some project (?). Not sure how that is related to the PCI dev ID.

You can't see this on the client side, and providing a list would probably not be helpful since the assignments change sometimes.

[Joe_H]

I would say same.
The #2 and #3 are exactly the same (probably a typo in GPUs.txt). And 8218 is just the base GFLOPS of the 2070S. I suppose there is a additional entry for it to fit with some specific manufacturer's use of the ROM indications.

2 and 3 are distinct entries, both are valid PCI device ID numbers for some variant 2070 Super and not a typo.

nVidia uses different ID numbers for just about every variant card even if they are based on the same chip. Whereas AMD may use the same ID number across many variants, sometimes GPUs based on the asme chip, but with some of the shaders disabled.

The client detects GPUs based on their device ID. Assignment is based on the vendor which is coded in the 3rd position, and species within that vendor's cards, the 4th field.

[ajm]

I would say same.
The #2 and #3 are exactly the same (probably a typo in GPUs.txt). And 8218 is just the base GFLOPS of the 2070S. I suppose there is a additional entry for it to fit with some specific manufacturer's use of the ROM indications.

2 and 3 are distinct entries, both are valid PCI device ID numbers for some variant 2070 Super and not a typo.

Indeed, sorry!

[KtC]

I found here on the list https://en.wikipedia.org/wiki/GeForce_20_series the "8218" is referenced as Single precision (GFLOPS). Seems that number is base "reference" only, and may reach 9062 with boost. I am guessing it is related to factory OC. But manual OC would change that too.
I also read theories the two additional Dev IDs are binned chips from different GPU models. But found in other place the third Dev ID is in fact Reference model from Nvidia.
Very confused.

[bruce]

When you run lspci, each device on your computer will return two numbers, (the first two on each line. Here's a convenient database where you can look them up.
https://devicehunt.com/search/type/pci/ ... evice/1EC2

[KtC]

It seems PCI Dev ID has an impact on project assignment, even if the GPU model is the same.
This is example based on two 2070S with different Dev IDs (projects limited to COVID-19):
GPU type #1 almost constantly assigned to:
- Projects 14445-14469 (COVID-19 NSP12)
- Projects 13416,13417 (COVID-19 Moonshot)
GPU type #2 the same model, different PCI Dev ID
- Projects 16800-16803,16805-16808,16810 (COVID-19)
- Projects 16447 (Cancer)

[JimboPalmer]

In GPUs.txt the PCIE iD is the primary index, while the Model Number is a comment field, only used to display to humans.

Code: Select all

- 0x10de:0x1e84:2:8:TU104 [GeForce RTX 2070 SUPER] 8218
- 0x10de:0x1ec2:2:7:TU104 [GeForce RTX 2070 SUPER]
- 0x10de:0x1ec7:2:7:TU104 [GeForce RTX 2070 SUPER]

It is interesting why one is listed as Species 8 while the other 2 are listed as Species 7

Some GPUs as far back as GTX 10x0 are Species 8, while many GPUs in RTX 20x0 are Species 7. That is not how I understand Species works.

[KtC]

You mean the numbers :8: and :7: specified in GPUs.txt. It looks like GPUs with performance higher than 2070 (non super) are marked as '8'. Except that two 2070 Supers.

[JimboPalmer]

[I am not an authoritative source, I am a user like you with 11 years of curiosity and a programming background. Nothing I say is true, it is just what I understand)

We have 2 fields of PCIE ID then a vendor field then a Species Number and then a comment field, which often has the card name.

Code: Select all

- 0x10de:0x1e84:2:8:TU104 [GeForce RTX 2070 SUPER] 8218
- 0x10de:0x1ec2:2:7:TU104 [GeForce RTX 2070 SUPER]
- 0x10de:0x1ec7:2:7:TU104 [GeForce RTX 2070 SUPER]

So Nvidia is Vendor 2 (AMD is Vendor 1, Intel is Vendor 3) And the first card is species 8 and the other two are species 7.

A blacklisted card may be :: or 0:0: or 2:3: (as an example) :: and 0:0: mean that this card will never be supported again while 2:3: is a card that MAY get more work, even if it does not currently get WUs.

Traditionally, the species is determined by capability, not performance. A Card that does not support OpenCL 1.2 or does not do Double Precision floating point math may have a lower species number than a card which does both. (Currently Double Precision floating point math is a hard requirement, while OpenCL 1.2 is believed to be needed, but it is not clear) On the AMD side, the newest cards need the newest Core, so 5700XTs are not allowed to use Core_21. For Nvidia all supported cards can do any WU, given time.

Sorting cards by capability however, puts the GT 1030 and the GTX 1080ti in the same species, even though their performance varies wildly. The GT 1030 can do everything the GTX 1080ti does, eventually, although the heat death of the universe may happen first. There has been some 'push' to measure performance.

While I have not heard that any Species are chosen by performance, it could be a trend I just have not noticed.

[KtC]

Thanks for more information about the species. I think it is hard to define those capabilities. See my example above, two cards theoretically with the same capabilities and yet different species. Here on the forum I heard there is also a 2060 KO card version which seems to be a better performer then a 'normal' 2060. It may have different capabilities and still the same specie. Not sure about PCI devices for those cards. In other hand some cores may fail totally and that would be an capabilities requirement.

I am guessing at some point the assignment is based on performance metrics running cores on particular hardware. The point where cards are assigned to :8: may cause some cores starving due to unavailability of hardware in the network. This is just my guess/estimation how this distributed system could be balanced with the assignment. It is also interesting if new family of cards RTX30xx in the future will get a new :9: specie after some cores will be maybe optimized for the new architecture.