There has been some discussion lately about running BOINC projects on Intel GPUs on Linux. Some people have tried to use Beignet but without success. I believe those people had made a mistake in their test setups so I decided to test Beignet myself.

Right now I have Beignet running happily on this host.

CPU in the box is i7-3770 and it's running Mint 17 Live CD/USB, kernel 3.13.0-24-generic, Wine 1.6.2 and Beignet versions 1.0.2 (32-bit) and 0.9.2 (64-bit, for Einstein). Beignet 1.0.2 was compiled with/against llvm 3.5, clang 3.5, gcc 4.8.2 and libdrm 2.4.56.

While Einstein BRP4 needs 0.9.2, Seti MB7 appears to be compatible with more versions, even if I have only tested 0.9.2 and 1.0.2 here (and 0.9.2 only saw quick benchmarking). I chose to run 1.0.2 here because release notes mentions some performance improvements (not that I saw anything in benchmarks) and because I could :P 1.0.2 also installs into a per-architecture lib directory by default which is handy when you need to support two architectures.

There has been no publicly released version of MB7 for Intel GPUs on Linux that I can remember so instead I'm running Windows version using Wine. The version appears to be r2489 which I found inside Lunatics installer. One downside of using Windows version is that BOINC client can't communicate with it. Progress and Remaining time will be based entirely on the client's guesses and the client doesn't track used CPU time.

So far about 100 tasks have validated. There has been one or two tasks that needed a third opinion before validating (not counting cases where wingman is turning in only overflows and such). That puts the inconclusive rate at 1-2% which I suppose isn't terribly bad. But of course the sample size is rather small.

Performance seems to be somewhere in the neighborhood of the Windows drivers although it's hard to compare with some arbitrary host when you don't know how much the CPU is loaded in that other host. I have ran this test without loading the CPU.

I also tried to run MB7 on i7-4790 but that was a complete failure. Benchmark workunits ran for 3000+ seconds and didn't even produce correct results. Beignet home page mentions a few problems with Haswell processors. However I didn't get any error messages anywhere and the kernel patch didn't seem to apply here. But now I'm having second thoughts, maybe I should have tried with an updated and patched kernel too.

Anyway, Beignet looks pretty good and in my opinion it is ready for wider beta testing. However, for anyone interested in testing it, I recommend starting with offline testing, with Lunatics benchmark tool, for example. Even if I got Beignet working on this host doesn't mean you'll find the right GPU+Beignet+kernel+anything else combination right away. If you start with live testing you'll risk trashing your entire daily quota in no time.

Since the machines I have used are ones that I have just borrowed I'll be leaving further testing to those who actually own or at least have a long term access to suitable hardware.

(Einstein side of the story is here.)