I do not have the latest NVIDIA cards to test with but the runtime of Fury X appears to be similar to what I am seeing with my 780 Ti and the latest CUDA 5.5 application. A single task runs at 2000-2100 seconds with this configuration.
Jeroen
Thanks Jeroen !
So .... the 290X was slower than a 7970 ...
and now this thing.
So .... the 290X was slower than a 7970 ...
and now this thing.
Meh !
Only in double precision, which luckily hardly any projects use. BTW: the Hawaii chip itself supports a 1:2 SP:DP ration, but it's locked away on Radeons to sell more FirePro :/
WUs are computed and send, but all of them fail on validation errors. Games are working fine, other OpenCl projects aswell (Seti, Collatz, Poem) http://einsteinathome.org/task/512578820
Since there is a definite problem with the AMD Hawaii(290/380) & Fiji(390) cards running multiple BRP6 wu's, I have a suggestion...
Perhaps, someone nearby the E@H developers could provide a card on short-term loan for the developers testing & resolution of this problem - all in the interest of science of course.
These AMD cards really do have untapped potential and something in the BRP6 openCl app is preventing this from being realized.
I think those high-end GPUs
)
I think those high-end GPUs need a lot of units in parallel to benefit from their massive amount of ALUs, regardless what the GPU usage says.
Would be nice if you could compare them with 4x to 8x.
Hi, all gpu tasks are failing
)
Hi,
all gpu tasks are failing on my Fury X
Host: http://einsteinathome.org/host/11721813
Result: http://einsteinathome.org/host/7526859/tasks
Starting BOINC client version 7.6.6 for windows_x86_64
OpenCL: AMD/ATI GPU 0: Fiji (driver version 1800.8 (VM), device version OpenCL 2.0 AMD-APP (1800.8), 4096MB, 4096MB available, 5376 GFLOPS peak)
OpenCL CPU: AMD FX(tm)-8300 Eight-Core Processor (OpenCL driver vendor: Advanced Micro Devices, Inc., driver version 1800.8 (sse2,avx,fma4), device version OpenCL 1.2 AMD-APP (1800.8))
no problem on Poem, Seti and Collatz
RE: driver version 1800.8
)
What's that? Some virtualized environment?
MrS
Scanning for our furry friends since Jan 2002
RE: I do not have the
)
Thanks Jeroen !
So .... the 290X was slower than a 7970 ...
and now this thing.
Meh !
Bill
RE: So .... the 290X was
)
Only in double precision, which luckily hardly any projects use. BTW: the Hawaii chip itself supports a 1:2 SP:DP ration, but it's locked away on Radeons to sell more FirePro :/
MrS
Scanning for our furry friends since Jan 2002
@ExtraTerrestrial Apes -
)
@ExtraTerrestrial Apes - What's that? Some virtualized environment?
No, nothing virtualized or special. Just Windows 8.1 and most recent 15.7.1 AMD drivers
http://support.amd.com/en-us/download/desktop?os=Windows+8.1+-+64
WUs are computed and send, but all of them fail on validation errors. Games are working fine, other OpenCl projects aswell (Seti, Collatz, Poem)
http://einsteinathome.org/task/512578820
Well then.. sorry, I have no
)
Well then.. sorry, I have no idea what else to suggest. This could be some serious driver bug.
MrS
Scanning for our furry friends since Jan 2002
I don't see any recent GPU
)
I don't see any recent GPU tasks.
I guess these are the invalid ones.
If you look at the "Validate error" tasks, (rather than the "Completed, marked as invalid" tasks) it seems these tasks both stopped and restarted.
Do you recall the host restarting during those tasks ?
One example
Otherwise i don't see any obvious problems, perhaps down-clocking might be worth trying.
Ahh, it looks like it's
)
Ahh, it looks like it's related to http://einsteinathome.org/node/198165
2 tasks at once will always fail, but single tasks are validating.
Since there is a definite
)
Since there is a definite problem with the AMD Hawaii(290/380) & Fiji(390) cards running multiple BRP6 wu's, I have a suggestion...
Perhaps, someone nearby the E@H developers could provide a card on short-term loan for the developers testing & resolution of this problem - all in the interest of science of course.
These AMD cards really do have untapped potential and something in the BRP6 openCl app is preventing this from being realized.
Gord