During the initial beta test they were issuing batches of the same size as CPU batches. Uplinger added some larger batches and I assume those will be the ones distributed when they go production. I don't believe the batch sizes have been settled yet. I haven't been a big GPU user over the years but was able to acquire a Radeon Pro W5500 so I guess I'll install it and see what happens. I haven't tried installing GPU drivers on Centos yet so this will be a first for me. I decommissioned 4 servers and have been in discussion with Luckasz at iThena about shipping them to him. Once we get past this deep freeze that should pick up again.
I think these are the correct settings to only get GPU tasks, although I haven't gotten any yet.
Make sure Beta tasks are enabled for the desired profile(s). I'm just using School.
Set your GPU preferences under My Projects -> Device Profiles. Allow CPU tasks as well, or not in my case, and enable the appropriate GPU type for your hardware.
I didn't have much time to observe performance when GPU work was available. I am running CPU tasks while waiting for GPU Beta work, but I have an app_config.xml file and CPU work units are limited when GPU work is running. Here is the text I am using for a single GPU on a quad core CPU.
For those new to the app_config, it only controls BOINC scheduling. The CPU and GPU will use whatever resources they need regardless of what's set here. This will keep a thread free for GPU processing though.
My computers completed 166 Beta GPU work units this evening with no errors. Members are having WUs fail and I suspect video drivers on their computers and probably some overheating as well. Mine are running fine, but the work runs a GPU as hard as PrimeGrid does. Uplinger stopped testing because of so many (primarily) nVidia GPU failures.
I hadn't checked this since the initial setup days ago. I haven't seen any of these running but I see that I returned four of them on the 27th. Validation is still pending.
Overclocking isn't recommended for this type of work as it's very sensitive, even factory overclocks can crash tasks. Nvidia Linux users also need to enable coolbits so they can control fan speed or else the card will run hot and clock down severely. I don't know if there's a similar situation for AMD cards or not.
My GPU machines run Windows 10 and I use MSI Afterburner or Radeon Control Panel. No overclock, plus 10% power, fan speed adjusted if needed to run GPU temperature at 75 C or lower, VRAM set to lowest speed (150 MHz with Radeon Control Panel).
There's going to be a mess on the WCG forums when the GPU Beta testing is over. It is bad enough already with Beta testers who have never run GPU projects!
The list of cards here are all low end cards that really shouldn't be used for crunching, with the possible exception of the AMD Radeon(TM) Vega 8 Graphics which is an APU. My suspicion is those other cards don't have enough VRAM for these tasks. I didn't check every one but the ones I looked at had either one or two GB, today at least four is recommended. I think admins can restrict workflow based on Compute Capability. I wouldn't send these to anything below a 750 Ti which I think is CC 3.5.
Yes, they will have to improve the points awarded. CPU users have to understand that some of the GPU rigs have thousands of stream processors instead the cores of their computers. My mining rig has 16,384 stream processors and I will go elsewhere when the Beta tests are done.
I have given up on the WCG beta work. I was around for the development of the GPU client for Folding@home. There will be problems with inadequate points awarded for GPU work and complaints from CPU crunchers about GPU clients being awarded too many points. The message board will light up about how the client broke computers, usually due to overheating or lagging video. Actually, the WCG GPU client was running when I lost a power supply and one of five video cards, but my hardware had many hours on it already.
On a more serious note, I was unable to get 100% GPU utilization even when running three WUs per GPU with power at +10%. The WCG client is currently extremely hard on video card and power supply voltage regulation circuits because of sporadic utilization. I don't know why in the world they are stopping the client to write to disk when GPU work has been so short. Save points are for CPU work.
Some growing pains are expected with a new application. WCG in general isn't a project you should run for the credit, there are much better choices if that's what you're into. There's bound to be someone complaining about the credit, either too much or too little regardless of what they do. The real issue for me is getting the best production from their resources, which will take some time. Hopefully they can get things optimized.