Ever wanted to BURP with your graphics card? Well, soon you can, over the next few months we're starting occasional random tiny tests with GPU-compatible workunits.
If you are really lucky you'll spot them as they fly by - they will go by the application name "Blender (GPU)". Quite likely the first few of them will be errors but eventually we will be able to move on to actually rendering something.

GPU workunits are going to be a bit different from other workunits in that they will not use all the CPU cores like the new multithreaded clients; rather they will use a full GPU and less than a single CPU core to manage that GPU.
GPUs are made slightly differently than CPUs and one of the things they have trouble with is switching between doing one thing and doing another - so they always let computations finish on one thing before moving on to the next. When using the same GFX card to compute and to draw the desktop this means that there can be a detrimental effect on the performance of the desktop. For this reason it is recommended to select "Keep apps in memory" and "Only do computation when computer is not in use" in the project prefs - more GPU-related options as well as the option to opt in/out of the testing will be available as soon as more of the new site is in the air.

Only certain machines will be able to take part in the tests so far:

• Must be a Windows 7/8/8.1 64-bit machine with at least 4GB of RAM
  
• Must have at least one NVIDIA graphics card with compute capability 2.0 or higher and CUDA 3.0 or higher with driver version 190.38 or higher
  
• Around 512MB or more of free, direct, non-shared GPU memory on each card
  
• Recommended for later tests: The cards should be connected via a PCI-Express 2.0x16 or 3.0 slot for optimal performance. A lot of traffic goes over this link, so making sure that the card is actually running at proper PCI-Express speed is important. Cards with slow PCI-Express connections may sometimes get less than 100% utilization.
  

Other platforms (Linux and OSX) may be able to run CUDA apps too, but right now focus is on getting it working at all, one step at a time - sorry!

If your machine is up for the challenge then be sure to check if you have the newest graphics drivers and that BOINC recognizes your cards correctly. It should print out the card name and info when it starts:

28/07/2014 13:05:39 | | Starting BOINC client version 7.2.42 for windows_x86_64
28/07/2014 13:05:39 | | log flags: file_xfer, sched_ops, task
28/07/2014 13:05:39 | | Libraries: libcurl/7.25.0 OpenSSL/1.0.1 zlib/1.2.6
28/07/2014 13:05:39 | | Data directory: C:\ProgramData\BOINC
28/07/2014 13:05:39 | | CUDA: NVIDIA GPU 0: GeForce GTX 560 Ti (driver version 340.43, CUDA version 5.0, compute capability 2.0, 1280MB, 1173MB available, 1312 GFLOPS peak)
28/07/2014 13:05:39 | | OpenCL: NVIDIA GPU 0: GeForce GTX 560 Ti (driver version 340.43, device version OpenCL 1.1 CUDA, 1280MB, 1173MB available, 1312 GFLOPS peak)
28/07/2014 13:05:39 | | Processor: 8 GenuineIntel Intel(R) Core(TM) i7 CPU 950 @ 3.07GHz [Family 6 Model 26 Stepping 5]
28/07/2014 13:05:39 | | Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss htt tm pni ssse3 cx16 sse4_1 sse4_2 popcnt syscall lm vmx tm2 pbe
28/07/2014 13:05:39 | | OS: Microsoft Windows 7: Professional x64 Edition, Service Pack 1, (06.01.7601.00)
28/07/2014 13:05:39 | | Memory: 12.00 GB physical, 96.00 GB virtual
28/07/2014 13:05:39 | | Disk: 148.96 GB total, 17.56 GB free
28/07/2014 13:05:39 | | Local time is UTC +2 hours
28/07/2014 13:05:39 | | Config: use all coprocessors

Too bad my CUDA machines are both Macs :*(

Windows machine is currently a Radeon.
____________

There has been some development on an OpenCL version but it was slower than CUDA when running on NVIDIA cards, didn't compile/run properly on AMD cards due to some limitations and was around the same speed as CPU when running as OpenCL on Intel CPUs.

Give it a couple of years then both the drivers, compilers, Blender, the hardware and OpenCL will have matured a lot!

Hopefully it won't take years for those of us with ATI cards to come to the party... D'oh, guess I should have gone with the GTX 760 instead of the R9 280X.

Ditto with OS X, the one CUDA system I have is my MacBook Pro.
____________

Picked up a Quadro FX M in a laptop that I've been using for over a month, but haven't seen any GPU work as of yet... Will also be getting a beefy gaming laptop soon with a GeForce 870 M... Bring on the GPU work!
____________

Bring on the GPU work!

The upcoming Linux CUDA tests have some additional requirements as the client requires the CUDA runtime and toolkit to be installed along with an X11-compatible graphical terminal with a working OpenGL environment. Apart from that the requirements from the earlier Windows test are the same: You still need to be able to see the card in BOINC etc.

It is very likely that initial workunits may fail on a specific machine and you may have to install one or more missing libraries to participate in this test. If so, the libraries are listed in the workunit output of the failed workunit.

Surprisingly the first 100 Linux CUDA workunit instances went extremely well with zero errors in the test group. The results validate against their Windows counterparts. The test is now extended to the entire Linux CUDA capable segment of the farm.

The next test unfortunately will not work with the very newest CUDA cards (GTX1060, 1070 and 1080) because the compiler for these cards was not ready when the current client was prepared for the farm.
Attemps to render with those cards will result in an error along the lines of "No kernel available for 6.1 cards yet".

The next version of the client will contain rendering kernels for the new cards as well if everything goes as planned.