Win7-64bit & ATI HD7790 : computation errors
18 Nov 2013 21:58:24 UTC
Topic 197277
(moderation:
Hi,
I just install a new ASUS HD7790 with the latest drivers (from ASUS CD-ROM - catalyst 13.11) in a fresh Win7 64bit SP1 installation.
BOINC (7.2.28) start processing ATI (BRP4G/BRP5-opencl-ati) WUs but after 5-10 seconds reports "computation error".
The stderr output is :
[ERROR] Error during OpenCL host->device original time series data transfer (error: -4) Demodulation failed (error: 2014)!
The GPU is woking fine (windows, games, etc) and I run 2 CPU WUs in a quad-core AMD PhenomII. Is there any extra configuration for the opencl client of the EaH ??
Thank you,
Win7-64bit & ATI HD7790 : computation errors
)
What does the Boinc startup show? Does it show OpenCL support, and what CAL and OpenCL versions.
Does the Catalyst 13.11 version from AMD work?
Claggy
RE: Hi, I just install a
)
Are you sure it is 13.11? That isn't released yet.
I would try the latest released version (13.9), which work fine with my HD 7790s on Einstein:
http://support.amd.com/en-us/download/desktop?os=Windows%207%20-%2064
But be sure to do a clean uninstall of the old drivers; first remove them using the Control Panel/Programs and Features, reboot, and then run a driver cleaner (I use Driver Fusion).
I tested the latest 13.11
)
I tested the latest 13.11 beta drivers (the latest intended for Battlefield 4) because of performance gains some users reported.
While it didn't do much in terms of performance for me, they did
a) work, but
b) turned out invalid results (Perseus Workunits) now and then (low percentage, but still notable)
That's why I reverted to the 13.9 WHQL Drivers again after about a week. For now, no more invalid results.
Uninstalling all AMD Drivers via the Program manager and installing the freshly downloaded Driver from the AMD Website & a reboot for good measure should do the trick.
After uninstalling the
)
After uninstalling the drivers from ASUS CDROM and install Catalyst 13.11_beta6 from AMD site BOINC is works fine (until now).
I will try the Catalyst 13.9 drivers to see if there is any performance issue.
With 13.11_beta6 my results are:
BRP4G-opencl-ati : average of 6000 sec/WU's
BRP5-opencl-ati : 1 wu in 22917 sec
Thanks,
Hm, these times seem
)
Hm, these times seem slow...
You should get better than these runtimes even when 2 WorkUnits run parallel.
My HIS HD7790 had a tad higher clockspeeds (GPU/VRAM) and did 2 Perseus parallel in ~16500s (16x PCIe 2.0 slot at 16x speed)...
So although the Card on Einstein@Home is by default running heavily bottlenecked by its 128bit VRAM (which holds back the quite powerful GPU alot), you should see better performance than that. If not already done, load it with 2 parallel tasks to assure it gets full CPU time alotted and doesn't run into the "GPU task gets no CPU time" bug.
And personally, I'd stay away from Beta drivers - there likely won't be any improvements for older cards like the HD7790.
Every driver development is usually focussed only on the latest gen cards that need to fight the front benchmark battles against the competition (and can actually cause slight performance decreases on older models).
RE: Hm, these times seem
)
I forgot to mention that these times are with 2 WUs in parallel.
I will test the stable drivers tonight and I will post the results...
Thanks,
Reading these post I
)
Reading these post I rechecked my HD7790. Really, performance dropped definitely.
So I removed any amd software and drivers, cleaned the registry and reinstalled the latest non-beta driver.
The result is: the card has only 10% of the performance of the HD7850 XT. Definitely bad.
Then I checked clock speeds with gpu-z. They are OK. But the card reports only 640 shaders, should be 896 !
I rechecked it with TRIXX, same result.
How can it happen that some shaders disappear?
That can't happen AFAIK, the
)
That can't happen AFAIK, the GPU has no logic to "disable" shaders (even if they're faulty, then the GPU will likely stop operating or fail otherwise).
I can't find any such host (Bonaire GPU equipped) listed for you on Einstein, so I can't see any of the numbers you're reporting.
10% of a HD7850 strikingly looks like the "GPU task gets no CPU time" bug (in case you only run 1 GPU workunit at a time).
The HD7790 should run at roughly (plus/minus 15% to allow for all those stock OC models) 60%-70% of the performance of a HD7850 on Einstein@Home with current tasks.
In terms of Perseus Workunits, this it what it looks to me (AMD FX-6300, stock speeds, 2 parallel GPU tasks with BRP factor = 0.5) :
HIS HD7850 (860Mhz GPU / 1200MHz 256bit GDDR5) : ~11500s (Single card @16x PCIe 2.0)
HIS HD7790 Turbo (1075MHz GPU / 1600MHz 128bit GDDR5) : ~16500s (Single card @16x PCIe 2.0)
Beware that running any of these cards at less than 16x PCIe2.0 speeds, performance will suffer somewhat even on the HD7790.
RE: I can't find any such
)
Should be this computer http://einsteinathome.org/host/6633241
but it really looks like its reporting wrong card.
Messages show :
19.11.2013 23:54:23 | | CAL: ATI GPU 0: AMD Radeon HD 7870/7950/7970 series (Tahiti) (CAL version 1.4.1848, 2048MB, 2008MB available, 7488 GFLOPS peak)
19.11.2013 23:54:23 | | CAL: ATI GPU 1: AMD Radeon HD 7700 series (Capeverde) (CAL version 1.4.1848, 1024MB, 984MB available, 960 GFLOPS peak)
19.11.2013 23:54:23 | | OpenCL: AMD/ATI GPU 0: AMD Radeon HD 7870/7950/7970 series (Tahiti) (driver version 1268.1 (VM), device version OpenCL 1.2 AMD-APP (1268.1), 2048MB, 2008MB available, 7488 GFLOPS peak)
19.11.2013 23:54:23 | | OpenCL: AMD/ATI GPU 1: AMD Radeon HD 7700 series (Capeverde) (driver version 1268.1 (VM), device version OpenCL 1.2 AMD-APP (1268.1), 1024MB, 984MB available, 960 GFLOPS peak)
19.11.2013 23:54:23 | | OpenCL: Intel GPU 0: Intel(R) HD Graphics 4000 (driver version 9.18.10.3165, device version OpenCL 1.2, 1624MB, 1624MB available, 45 GFLOPS peak)
19.11.2013 23:54:23 | | OpenCL CPU: Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz (OpenCL driver vendor: Intel(R) Corporation, driver version 1.2, device version OpenCL 1.2 (Build 63463))
19.11.2013 23:54:23 | | OpenCL CPU: Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz (OpenCL driver vendor: Advanced Micro Devices, Inc., driver version 1268.1 (sse2,avx), device version OpenCL 1.2 AMD-APP (1268.1))
which is also wrong, it's not a capverde product. Should be Bonaire.
GPU-Z reports that correct.
Card is running in a Pcie3 x 8 slot, running 2 wu's, enough cores left free.
Performance of the other gpu's is well within expected range.
I found a 'HD7790 performance driver', but this did not help; at least it is not slower after installing it.
@ Techpowerup I found a lot firmware images. But every image I tried results in a so-called 'subsystem mismatch'.
I'm giving up for today, have my next meeting in about 7hrs, it's time to go to bed.
Anyway, thx for your help!
Alexander
Screenshot of GPU-Z:
https://dl.dropboxusercontent.com/u/50246791/Bad%20HD7790.PNG
Edit: I see GPU-Z is reporting it now wrong ...
Odd, GPU-Z also reports a
)
Odd, GPU-Z also reports a CapeVerde (?)
That would make it more of a HD7770 (with some GPU OC) and not a HD7790.
Could it be that some of the firmware flashes made you try to flash a HD7750/HD7770 to a HD7790 ? Or that otherwise the Firmware you tied to flash on a HD7790 was actually intended for the Cape Verde GPU ?
Otherwise, the GPU-Z readouts and OpenCL reports don't make much sense to me, IF you have an actual HD7790 in your System.
In any case, the HD7750/HD7770 Cape Verde Video cards aren't firmware compatible with the Bonaire GPU of the HD7790.
Generally speaking, there is just about no performance to be gained here at Einstein with either Drivers or Video card Firmware/BIOS flashes. It's best to leave these things alone; nothing to be gained but invited LOTS of potential problems.
Especially if you forgot to save the factory BIOS for a rollback. Video card Firmwares are sometimes hardware revision specific, so for example flashing a V1.02 RevB onto a V1.0 RevA Board may cause issues/errors i.e. due to changed timings etc. Same is true for GDDR3 vs. GDDR5 equipped models, as well as 1GB vs. 2GB Variants.
Many people forget to check the PCB of the Video card for board revision figures.
(made that mistake once myself, won't repeat it ;) )