AMD RX Vega

[monkeyclaw]

So hey! I might (just maybe) actually have someone willing to try folding on a Vega64 LCE. Only thing is that it is I'm not sure if it is added to the whitelist yet... does anyone know where to check the current whitelist? And, this topic: viewtopic.php?f=83&t=26208 is still the fastest/quickest way to get a card whitelisted, right? How long does it normally take for the whitelist to become updated after a request has been submitted?

Thanks!

[monkeyclaw]

Sorry to butt in, but would this ID also work for other Vega SKUs, specifically the liquid-cooled edition, or is that not added in yet?

[Joe_H]

If AMD used the same ID number for the other models, then yes. The Information topic here in this subforum gives some methods of determining what device ID an installed card has.

That topic also gives a link to the current GPUs.txt file. The few persons who can add to that file usually post here once a new device ID and card description have been added.

[Senture]

Thanks very much for adding the card in. I have the water cooled RX Vega 64 card. I would Imagine the air cooled RX Vega cards will have the same device ID string. The only thing different between them is the BIOS which has an increased clock speed and power limit on the water cooled edition.

There is a grammatical error in the device string. The line reads "0x1002:0x687f:1:5:[Radeon Rx vega" but it should be "0x1002:0x687f:1:5:[Radeon RX Vega]". Nothing important, but looks a little out of place

[Joe_H]

I just looked at the latest GPUs.txt file, there are two different device ID entries for Vega at this point. The first is for the FE version released back in June, once the Vega 56 is released next week and someone supplies a device ID, that can be added.

Possibly the entry for the RX Vega would be corrected at that time, I suspect it was added from a mobile device. That portion of the entry is just "cosmetic", does not affect getting WU's.

[monkeyclaw]

Has anyone added/corrected the Vega 56/64 IDs in yet? Still haven't heard of any Vega GPUs in action for folding.

[Joe_H]

The Vega 64 entry is still the same, no one has yet provided a device ID for a Vega 56. There is one person posting here who has run into driver problems, GPU still consuming power after folding paused.

[rhavern]

Hope it is not my chart... I check it. No it says 115k PPD for gtx 950 but that may not be perfect accurate either.
https://docs.google.com/spreadsheets/d/ ... edit#gid=0

Nice spreadsheet, I do something similar but instead of PPD, I use the output of FAHbench as it eliminates the variation in projects and you can get useful comparison numbers before gpus.txt gets updated. Note that most projects use explicit single precision, so I do exp1/watt and exp1/<unit of currency> for comparison.

[foldy]

@rhavern: Is your sheet public?

[rhavern]

@foldy: Just for you, it is now.

https://docs.google.com/spreadsheets/d/ ... sp=sharing

Note: I haven't been able to find any FAHbench data on the RX vega 64 liquid model.

[foldy]

FahBench for vega liquid could be estimated just by multiply with the increased GPU clock factor.

[bruce]

I use the output of FAHbench as it eliminates the variation in projects and you can get useful comparison numbers before gpus.txt gets updated. Note that most projects use explicit single precision, so I do exp1/watt and exp1/<unit of currency> for comparison.

Question: Does anybody know if single/double precision applies to the CPU and/or the GPU?

Fact: The recent GPU cores are compiled to use what's called "mixed precision" by the OpenMM folks. A very large percentage of the calculations use single precision but a limited number of specific calculations may require double. Early GPUs did not support double although almost all recent GPUs do. That means the core must somehow choose whether to use the GPU to do those double precision calculations or to somehow off-load that calculation to the CPU -- if you're assigned one of the projects that manages to do everything in single.

(I'm not sure how the logic of that choice works or how it might affect FAHBench. If somebody wants to dig into the OpenMM source and tell us, it would be appreciated.)

[rhavern]

That's some useful and interesting data, Bruce. It'll be interesting to see if anyone knows the answer.

[JiiPee]

Based on the FoldingBench tests, the Vega 64 looks like it's going to fall in the 650k balllpark range on average, between the 600k of the GTX 1070 and the 700k of the GTX 1080 (again, those are BALLPARK figures, specific model of card will change that quite a bit due to "OEM factory overclocking" and higher TDP on many OEM cards).

Given the MUCH higher power usage of the VEGA cards, I'd say "don't bother getting them to fold on, but use them if you have other reasons to get one" - kinda like the GTX 980ti by current standards.

My Vega 56 is doing currently 575k with under 150W.
Powerlimit +50%
HBM at 900MHz
P6 volt at 1050mV
P7 volt at 1100mV

I think that not too bad performance. When I flash Vega 64 bios, then I can push HBM up to 1100MHz. Not sure if it increase performance, but I'm hoping to hit over 600k then.

[Holdolin]

I seem to be a bit late to this party, but I have had a pair of Vega64's folding for a couple weeks now. They seem to average 450-600kPPD.