Apologies for the late response, after my last post I disappeared to enjoy a long weekend with family.
tbret wrote (very diplomatically!):
Quote:
I've been a spectator while this thread has been lengthening.
I know you are completely competent. I know you are very experienced.
Your results are "bothering me," though, so I thought I would ask this one question:
How sure are you that your driver installation is "completely, absolutely, clean?"
I'm not suggesting you've done it wrong. I just know how much trouble I had getting a good driver installation on a Win 10 installation of a 390X (way more than I ever had before - multiple cleanups, multiple directory deletions, multiple Registry "fixes" etc.).
The Vulkan runtimes just really messed me up and they did *not* want to go-away and leave a clean "path". I have no idea what got installed where. I just know that whatever it was took a lot of effort to make it go away.
I am as certain as one can be that the driver installation is good but it should be noted that I usually take the 'upgrade' path with new drivers and again in this case I have not yet done a 'clean' install.
Dropping task concurrency to x2 improved my situation and the 480 has been consistently chugging away during my break and its runtimes have been within reasonable expectation. At x2 concurrency the performance is roughly mid way between that of a 280X and 380 (with the 280X being fastest). In my opinion the x2 runtimes are not great at roughly 46 minutes per task...
I have now removed the 280X card and moved the 480 into my only available PCI-e3 x16 slot. I'm not banking on there being much improvement from PCI-e3 and the full x16 bandwidth but you never know!
Whatever happens over the next couple of days I will stop testing for "my own ends" and maybe start testing the proper way...
Single card @ x1, x2, x3. Clean driver install etc. and attempt to take some proper power measurements.
I have already founded mixed feelings about my 480 because it's not doing what I hoped it would do, its almost a love / hate relationship, on the one hand I think it should be better than the old 280 but on the other its better than a 380. Certainly draws much less power than a 280X.
A member of SETI.Germany tried POEM on his RX 480. Interestingly (or shockingly?) the runtimes increased by roughly a factor of 5 when he went from 1x concurrency to 2x. I stronlgy suggest to check for such behaviour at Einstein.. this could explain it all.
A member of SETI.Germany tried POEM on his RX 480. Interestingly (or shockingly?) the runtimes increased by roughly a factor of 5 when he went from 1x concurrency to 2x. I stronlgy suggest to check for such behaviour at Einstein.. this could explain it all.
A member of SETI.Germany tried POEM on his RX 480. Interestingly (or shockingly?) the runtimes increased by roughly a factor of 5 when he went from 1x concurrency to 2x. I stronlgy suggest to check for such behaviour at Einstein.. this could explain it all.
MrS
Similar shock with some Fermi cards running OpenCL programs on a different project. The cards had always done 3x or at least 2x, but really suffered
trying that with the newer program. (all validated after "forever" run times)
Eventually I chalked-it-up to the CPU being "too busy" and everything smoothed-out nicely at 1x each.
1x: 3000 s/WU, 220 W
3x: 2587 s/WU, 235 W
4x: 2575 s/WU, 235 W
8x: 2912 s/WU, 235 W
So 4x or 5x seems to be best for maximum throughput. I'm not sure how applicable these findings are for other systems than Linux. The driver certainly differs in many ways.
1x: 3000 s/WU, 220 W
2x: 2660 s/WU, 230 W
3x: 2587 s/WU, 235 W
4x: 2575 s/WU, 235 W
8x: 2912 s/WU, 235 W
Lets do 5x - because we can..
I'm not having as much success as AgentB. My Windows 10 x1 runtimes were very much in line with the above chart, my x2 runtimes were closer to the agent's x3 displayed. All my attempts to run at x3 have ended in the card 'locking up', yet according to AMD settings and GPU-Z the 480 was running at full steam whilst only a random single task would make the effort to progress a tiny fraction every few seconds.
I took to heart what tbret had written earlier about ensuring a 'clean' driver install and used DDU to remove the old driver (repeatedly), spent what felt like a lifetime looking for remnants to clean manually before installing a fresh download of the latest driver (albeit beta)...
Sadly for me, attempting to run x3 still results in lock ups so I will be RMA'ing and requesting replacement this weekend in the hope that I've just been unlucky with my current card.
In the meantime I wonder if AgentB will fire up his secret Windows VM and re-run his testing ;-)
In the meantime I wonder if AgentB will fire up his secret Windows VM and re-run his testing ;-)
There's a scene in Pulp Fiction with the gimp which springs to mind...
It's now worse than a VM - i have recently taken ownership of an occulus rift, and sadly a win10 partition to support it (yes it's easy to do this with UEFI, once you work out Win10 tries to take control of everything including the power off functionality). I may write about the rift in another thread, another day.
All i'll say is, i'm glad i don't have to pay for a Win 10 licence. It reminded me of the same old ex-girlfriend but with an even more expensive handbag, somewhat fatter and even more demanding, than she was several years ago. And yes the handbag looks nice, but filling it full of cobblestones makes me feel uncomfortable.
Meanwhile 5x results are in
1x: 3000 s/WU, 220 W
2x: 2660 s/WU, 230 W
3x: 2587 s/WU, 235 W
4x: 2575 s/WU, 235 W
5x: 2520 s/WU, 235 W
8x: 2912 s/WU, 235 W
Invalids so far ... 0
I may run 6x - course Gav might fire up his Linux VM and give that a go... ;-)
I'll think about letting the gimp out over the weekend.
course Gav might fire up his Linux VM and give that a go... ;-)
I suppose I ought to give it a go myself and I apologise in advance if my shouting and swearing disturbs anyone whilst I try and get Linux to do what I need it to. Expect many calls for help later :-)
Quote:
I'll think about letting the gimp out over the weekend.
Your post brought tears of mirth!
If you do find time to fire up the devil this weekend could you jump straight in to testing at x3 or above please? My 480 will run happily at x3 for anything up to around 30 minutes before the memory controller activity suddenly drops to zero and crunching stalls...
RE: And BTW someone else
)
Someone did for the 1070.
Apologies for the late
)
Apologies for the late response, after my last post I disappeared to enjoy a long weekend with family.
tbret wrote (very diplomatically!):
I am as certain as one can be that the driver installation is good but it should be noted that I usually take the 'upgrade' path with new drivers and again in this case I have not yet done a 'clean' install.
Dropping task concurrency to x2 improved my situation and the 480 has been consistently chugging away during my break and its runtimes have been within reasonable expectation. At x2 concurrency the performance is roughly mid way between that of a 280X and 380 (with the 280X being fastest). In my opinion the x2 runtimes are not great at roughly 46 minutes per task...
I have now removed the 280X card and moved the 480 into my only available PCI-e3 x16 slot. I'm not banking on there being much improvement from PCI-e3 and the full x16 bandwidth but you never know!
Whatever happens over the next couple of days I will stop testing for "my own ends" and maybe start testing the proper way...
Single card @ x1, x2, x3. Clean driver install etc. and attempt to take some proper power measurements.
I have already founded mixed feelings about my 480 because it's not doing what I hoped it would do, its almost a love / hate relationship, on the one hand I think it should be better than the old 280 but on the other its better than a 380. Certainly draws much less power than a 280X.
Gav.
A member of SETI.Germany
)
A member of SETI.Germany tried POEM on his RX 480. Interestingly (or shockingly?) the runtimes increased by roughly a factor of 5 when he went from 1x concurrency to 2x. I stronlgy suggest to check for such behaviour at Einstein.. this could explain it all.
MrS
Scanning for our furry friends since Jan 2002
RE: A member of
)
No such mischief with BRP6x1 here...First task at x1
GPU temp 79C-
Wall power ~220W (220V)
Tasks being validated ok.
Elapsed time ~3000s
CPU time ~430s
RE: A member of
)
Similar shock with some Fermi cards running OpenCL programs on a different project. The cards had always done 3x or at least 2x, but really suffered
trying that with the newer program. (all validated after "forever" run times)
Eventually I chalked-it-up to the CPU being "too busy" and everything smoothed-out nicely at 1x each.
Summarizing AgentB's results
)
Summarizing AgentB's results so far:
1x: 3000 s/WU, 220 W
3x: 2587 s/WU, 235 W
4x: 2575 s/WU, 235 W
8x: 2912 s/WU, 235 W
So 4x or 5x seems to be best for maximum throughput. I'm not sure how applicable these findings are for other systems than Linux. The driver certainly differs in many ways.
MrS
Scanning for our furry friends since Jan 2002
OK results for 2x are
)
OK results for 2x are in...
1x: 3000 s/WU, 220 W
2x: 2660 s/WU, 230 W
3x: 2587 s/WU, 235 W
4x: 2575 s/WU, 235 W
8x: 2912 s/WU, 235 W
Lets do 5x - because we can..
RE: OK results for 2x are
)
I'm not having as much success as AgentB. My Windows 10 x1 runtimes were very much in line with the above chart, my x2 runtimes were closer to the agent's x3 displayed. All my attempts to run at x3 have ended in the card 'locking up', yet according to AMD settings and GPU-Z the 480 was running at full steam whilst only a random single task would make the effort to progress a tiny fraction every few seconds.
I took to heart what tbret had written earlier about ensuring a 'clean' driver install and used DDU to remove the old driver (repeatedly), spent what felt like a lifetime looking for remnants to clean manually before installing a fresh download of the latest driver (albeit beta)...
Sadly for me, attempting to run x3 still results in lock ups so I will be RMA'ing and requesting replacement this weekend in the hope that I've just been unlucky with my current card.
In the meantime I wonder if AgentB will fire up his secret Windows VM and re-run his testing ;-)
Gav.
RE: In the meantime I
)
There's a scene in Pulp Fiction with the gimp which springs to mind...
It's now worse than a VM - i have recently taken ownership of an occulus rift, and sadly a win10 partition to support it (yes it's easy to do this with UEFI, once you work out Win10 tries to take control of everything including the power off functionality). I may write about the rift in another thread, another day.
All i'll say is, i'm glad i don't have to pay for a Win 10 licence. It reminded me of the same old ex-girlfriend but with an even more expensive handbag, somewhat fatter and even more demanding, than she was several years ago. And yes the handbag looks nice, but filling it full of cobblestones makes me feel uncomfortable.
Meanwhile 5x results are in
1x: 3000 s/WU, 220 W
2x: 2660 s/WU, 230 W
3x: 2587 s/WU, 235 W
4x: 2575 s/WU, 235 W
5x: 2520 s/WU, 235 W
8x: 2912 s/WU, 235 W
Invalids so far ... 0
I may run 6x - course Gav might fire up his Linux VM and give that a go... ;-)
I'll think about letting the gimp out over the weekend.
RE: course Gav might fire
)
I suppose I ought to give it a go myself and I apologise in advance if my shouting and swearing disturbs anyone whilst I try and get Linux to do what I need it to. Expect many calls for help later :-)
Your post brought tears of mirth!
If you do find time to fire up the devil this weekend could you jump straight in to testing at x3 or above please? My 480 will run happily at x3 for anything up to around 30 minutes before the memory controller activity suddenly drops to zero and crunching stalls...
Gav.