Einstein: Einstein@Home GPU Application for ATI/AMD Graphics Cards

[spingadus]

Hmm, my only attempt on a 5850 lead to failure after 0.94 hours and if not mis-remembers around 80% done, so close to 8 hours seems much too high.

But granted, computer crashing multiple times and taking the checkpoint-file with it in the process can also have screwed-up the cpu-time, so...

I've run 3 tasks and all took about 28,000 seconds. It's possible that if I freed up a thread or 2 it may speed up the task. I'll probably not test it any time soon though. Perhaps a third opinion?

[kmanley57]

I just looked at one of my Wu's that had been run by a NVidia and a ATI card. The Cuda ran it in 2,326 seconds and the ATI in 41,311 seconds, with the ATI card being the faster of the two in floating-point calculation speed! So the 2096 GFlop ATI to the 1405 GFlop CUDA looked bad! Both machines ran with a 4 core cpu, and near the same clock speed.

[DrPop]

This does not mean the CUDA card you have is faster / better than the ATI card. What it means is in this case, the code you are running on that project does not use the ATI card to its full potential. It is either poorly coded on their end, or using only a percentage of what the ATI card can do. Hope that makes sense? Some projects are really lousy on one brand and much faster on the other. For example, if you were on Collatz, you would see just the opposite happen.

Sent from my MB860 using Tapatalk 2

[Rattledagger]

I just looked at one of my Wu's that had been run by a NVidia and a ATI card. The Cuda ran it in 2,326 seconds and the ATI in 41,311 seconds, with the ATI card being the faster of the two in floating-point calculation speed! So the 2096 GFlop ATI to the 1405 GFlop CUDA looked bad! Both machines ran with a 4 core cpu, and near the same clock speed.

Well, once-upon-a-time running SETI with a full set of cpu-tasks on my Ati-5850 wasn't a problem, but after a OS re-install the GPU-utilization often dropped to zero, and since didn't manage to hit the "correct" combination of drivers the only option to speed-up SETI again was to sacrifice a cpu-core for this.

So, wouldn't be surprised if Einstein@home also wants a cpu-core to keep the GPU busy...

[kmanley57]

The difference I would expect is a coding issue! It wants less than many other projects for core utilization, but I will try to half the cores used and see if that helps any.

[Rattledagger]

After swapping-around things that seems to have solved the crashes and finally having time to install a different OS is up and running on the Ati-5850 again. For the moment under win8-preview, so many things probably won't work, but atleast v7.0.28 BOINC finally installed without any problems.

Configured Einstein@home to run 2 tasks at once on the GPU, by setting "GPU utilization factor of BRP apps" to 0.5 in the Einstein-preferences. On client this equals 0.5 GPU, 0.5 CPU per task.

Started-up GPU-Z, verified OpenCL was installed, and started downloading Einstein-GPU-work. Then the 1st. task started it according to GPU-Z used roughly zero of the GPU, and both "Elapsed" and "Remaining" was counting-up.

Then the 2nd. task started on the other hand, GPU-usage jumped to roughly 75%, and one of the cpu-only tasks was paused. While "Elapsed" for both GPU-tasks counted up, "Remaining" now started counting-down, but granted not as fast as estimated.

Both finished after roughly 2h10m, meaning on average 65 minutes per task. Looking on actual cpu-time, this was roughy 30 minutes/task, so by running 2 at once this means 1h cpu-time for 2h10m run-time or Einstein@home uses 45% of a cpu-core.

Now no idea if so short run-time was just a fluke and neither if my GPU just computed garbage and will fail validation, but atleast these results indicates Ati is much faster than either spingadus or kmanley57 results indicated, as long as allows Ati the usage of a cpu-core.

[kmanley57]

Well I reserved a core for the ATI card and it did the WU in ~14000 seconds, but my NVidia still does a WU in ~7000 for the same amount of credits(500). So I see a CPU taking ~22000(~250 credits), the ATI ~14000(500 credits), and the NVidia about ~7000(500 credits). But with faster cards I am sure others will get better returns on time used.

[Al]

Configured Einstein@home to run 2 tasks at once on the GPU, by setting "GPU utilization factor of BRP apps" to 0.5 in the Einstein-preferences. On client this equals 0.5 GPU, 0.5 CPU per task.

Is an app_info file needed in addition to the preference change? My NVIDIA card shows .2 CPU and .5 Gpu, but only one task runs at 77% Gpu utilization.

I suck at app_info files so I'm hoping it isn't necessary.


Sent from my ADR6300

[zombie67]

No app_info.xml needed. Just change the setting:

1 = 1 task per GPU
.5 = 2 tasks per GPU
.33 = 3 tasks per GPU, etc.

FYI, the upper limit is the memory of the GPU. Per the admin, you can run 2-3 tasks per 1gb of VRAM.

[Al]

My bad. Had to wait until it completed one before it would do 2 at a time. Thanks