Max Concurrency limits (7970s)
17 Mar 2021 6:11:16 UTC
Topic 225064
(moderation:
Recently a 7970 (r9 280x) was donated to me from a teammate. I now have 2 in this machine. So i recently switched back to einstein and quickly remembered that running 1 wu at a time was leaving a lot on the table. So I quickly went up to 2 WUs at a time per gpu. After doing that for a week or two I read some posts about running much more concurrently. I can't recall what the setups were in that discussion but I've found a wall for the 7970/r9 280x/tahiti, but i'm not sure what it is. It seems 3x concurrency is best for my 7970s, on a credit/hour basis. The gains from 1 to 2 is like 20-25%, from 2-3 probably 10% but if i go to 4 concurrent things fall off huge. I've allowed 3 full cores (7 threads free) free on this 3770k for testing. I'm not sure if it's vram spillover or what, i'm curious where/what/how to check what this limiter is.
MindCrime wrote: Recently a
)
It's also got to do with the amount of onboard gpu ram they have, only 3gb, meaning you are running out of memory on the card and there is nothing you can do about that. Do you use gpu-z or MSIAfterBurner to monitor the usage of the gpu? Each will give you a more accurate usage showing than anything MS gives us. I'm assuming you are using Windows but don't actually know because your pc's are hidden.
GREAT -- hiding computers
)
GREAT -- hiding computers and letting us guess if GR or GW is meant -- please !
San-Fernando-Valley
)
Users are perfectly entitled to leave the default setting for privacy unchanged. In most cases, it's just because they don't realise that the setting is there and can be changed. No harm in advising politely of the benefits (eg. not having to list all relevant details) of having computers visible to others.
For the substance of your complaint, it should be pretty obvious that it can't be GW - for multiple reasons, like GCN 1st gen GPUs (Pitcairn, Tahiti, etc) cannot run the GW GPU app successfully. Also, a 3GB GPU probably doesn't have enough VRAM to run 2x, let alone the quoted 3x and 4x observations. The app must be FGRPB1G for those multiplicities to be possible.
It's no surprise that 4x falls off a cliff. GRP tasks need close to 1GB each so even 3x must be very close to the edge.
From my experience running a HD7950 on GRP tasks, 2x is the best compromise for long term stability. If I run 3x, there is a very minor output improvement at the cost of major increases in the rate of lockups and other instabilities. The time wasted in monitoring and handling the problems far outweighs any theoretical gain.
Cheers,
Gary.
I'm sorry for the lack of
)
I'm sorry for the lack of info. It is the "Gamma-ray pulsar binary search #1 on GPUs v1.22 (FGRPopencl1K-ati)
windows_x86_64"
And yeah it looks like a vram thing. One card (primary/display) is using 2.4gb of vram right now and the other is using 2.1gb at 3x
4x would push the first one over 3gb but the 2nd one could fit another. :)