Page 1 of 1
Question about GPUSpecies
Posted: Wed Oct 14, 2020 6:32 am
by Jupis_folder
Hello all together and a nice Wednesday!
Yesterday I read the announcement
https://foldingforum.org/viewtopic.php?f=24&t=36252 and I wondered if there is somewhere something to read about the parameter GPUSpecies. Is there something?
Re: Question about GPUSpecies
Posted: Thu Oct 15, 2020 5:45 am
by NormalDiffusion
17304 is excluded from amd GPUs (PPD extremely low) and not for very fast Nvidia GPUs (low ppd).
17305 is limited to fast amd GPUs and very fast nvidia GPUs
17306 is limited to fast amd and very fast nvidia GPUs.
Re: Question about GPUSpecies
Posted: Thu Oct 15, 2020 7:08 am
by Jupis_folder
Is there a detailed list, which GPUs are which species?
Re: Question about GPUSpecies
Posted: Thu Oct 15, 2020 7:18 am
by PantherX
If you look at your GPUs.txt file, it will have all those details per GPU.
Off the top of my head:
8 -> Top of the line Nvidia GPUs from 1000, 2000 and 3000 Series
7 -> Second best Nvidia GPUs
6 -> mid-range Nvidia GPUs
5 -> something...
4 -> something...
3 -> something...
2 -> low-end Nvidia GPUs
1 -> Nvidia GPUs that support Double Precision but only have OpenCL 1.1
0 -> Nvidia GPUs that don't support Double Precision
For AMD, they have a completely different classification which I don't know since it is a bit more convoluted than Nvidia's classification.
This will completely change in the future when we move towards automatic benchmarking and classification of GPUs.
Re: Question about GPUSpecies
Posted: Thu Oct 15, 2020 2:21 pm
by Jupis_folder
Thanks.
Is it the 4th parameter?
Code: Select all
0x10de:0x1c81:2:7:GP107 [GeForce GTX 1050 LP] 1862
^ GPUSpecies?
Re: Question about GPUSpecies
Posted: Thu Oct 15, 2020 2:42 pm
by Joe_H
Yes. The first two fields are the PCI device ID (manufacturer:device), third field indicates AMD (1), nVidia (2) or Intel (3). The fifth field is a description string, only used for display purposes.
Re: Question about GPUSpecies
Posted: Thu Oct 15, 2020 3:03 pm
by aetch
The purpose of the GPU Species is to try to direct work units to graphics cards suitable for that project.
The GPU Species is a band representing the speed and capabilities of the graphics card you're offering for use.
We donors/folders/whateveryouwanttocallus have a wide variety of graphics cards ranging from old/slow to new/fast and every combination in between.
The work units come in different sizes and complexity as well, some big, others small. A big unit will swamp a small card and likely fail the deadline. Equally a small unit wastes the resources of a big fast card as it may only use the card to 40%.
Occasionally we have received work units unsuitable for the hardware we're running which is probably part of the reason there is a project to give a greater level of granularity to the GPU species list to better distribute the work units to the hardware we have offered up.
Re: Question about GPUSpecies
Posted: Thu Oct 15, 2020 7:56 pm
by Jupis_folder
Thank you all for your explanations!
Re: Question about GPUSpecies
Posted: Fri Oct 16, 2020 11:53 pm
by MeeLee
They may need to extend the classification, as Nvidia RTX 3000 GPUs are a bit faster than 2000 GPUs, but are significantly faster (with more many more cores) than the GTX 1000 series.
I think the RTX 3000 and upcoming 4000 series, as well as AMD BIG NAVI will need a totally new category.
Re: Question about GPUSpecies
Posted: Sun Oct 18, 2020 6:29 am
by PantherX
IMO, I won't worry about reclassification of the GPU Species at this point since the new system is being worked on. I would personally prefer that the new system is developed and leave the existing system as is until the new one is on-line and functional.
Re: Question about GPUSpecies
Posted: Sun Oct 18, 2020 6:25 pm
by bruce
The active system was based primarily on the hardware classification given my to the actual chip in the GPU. They were NOT split up based on the number of execution units (aka shaders). The future system will account for actual measured FAH throughput speeds; not on advertised features.