We just released the first app versions of a new BRP4 v1.28 app release. Over the next days, we will release this app for all supported platforms, today we released

- OpenCL (ATI/AMD), 32 bit, for Linux,
- OpenCL (ATI/AMD), 32 bit, Windows and
- OpenCL (ATI/AMD) OSX (Lion and later).

All versions require BOINC 7.0.25 or later, but we encourage volunteers to actually use the latest recommended BOINC versions, currently 7.0.28 or 7.0.31.

Versions for CPU and NVIDIA CUDA GPU will follow soon, plus some app versions that are native 64 bit apps.

New in this app version:
- Various performance improvements
- cross-validation improvements for HD 69xx GPUs

Cheers
HBE

---

Hmm. My ATI OSX/ML machine can't get work. 7.0.31.

---

Dublin, California
Team: SETI.USA

Hmm. My ATI OSX/ML machine can't get work. 7.0.31.

We had a temporary problem on Albert, BRP4 work should now be available again.

Cheers
HBE

---

Got some. Thanks!

---

Dublin, California
Team: SETI.USA

I did 1 task.

http://albert.phys.uwm.edu/results.php?hostid=3755

Was faster then the older one.

Good work.

http://albert.phys.uwm.edu/workunit.php?wuid=98764
Looking good so far. Great work.
Stephan

---

Crunched roughly 50 so far with no problems; using an HD 6850 with Boinc v.7.28 and OC'd i5 2500K (4.5GHz).

Time reduced a lot from average 3,800secs to 2,100secs doing 1xWU at a time...now if you could reduce the CPU Test WUs from 16,000+ secs I wouldn't mind. :)

- OpenCL (ATI/AMD) OSX (Lion and later).

I have two pretty much identical hosts - 1859 and 1868 that now have this app. They are both running the latest BOINC (7.0.31) and both have AMD 5750 1GB GPUs. The only hardware difference is that the first one has 4GB RAM whilst the second has 8GB.

The one with smaller RAM has run times around 7Ksecs whilst the other is taking 18Ksecs????

Both machines run EAH tasks on all 4 CPUs and both are doing S6LV1 tasks in aroung 17-18Ksecs. All figures quoted were for tasks crunched over the weekend when there was no user sitting at the keyboard. These machines are used during office hours on weekdays. When idle, both machines power down the screen quite quickly (maybe 3 mins) so the only thing I can now think of is whether or not the screensaver process is still running when the screen turns off. I'll have to check that next time I have access to those machines.

Anyone got any other suggestions? At some point I will release a CPU core from crunching to see what difference that makes and later I may attempt to run two tasks simultaneously - maybe not when you think of the extra heat that might be produced. The aluminium cases are running pretty hot already :-).

---

Cheers,
Gary.

Hi Gary!

Freeing a CPU core is probably a good thing to try.

Another idea: What I'm always a bit suspicious about (BOINC-performance-wise) is OSX's power saving feature to automatically switch between a CPU build-in GPU (e.g. the Intel stuff in Sandy/Ivy Bridge CPUs) and the dedicated graphics card. Could this be the case that this feature is set differently on those two hosts perhaps?

Cheers
HB

---

Congratulations!

My pc did a BRP 1.28 in ~1980 sec. HD6900 Series GPU, i7 with free cpu, 7.0.31, CCC 12.6
On HD5850 a wu takes ~3450sec.
AMD-apps are grown up now.

---

Phantastic!



The amd is a HD6950 running a single wu @ 82% gpu load, the nVidia is a GTX680, unknown settings.

edit: uups, picture is missing...


wu 997334
task 279100 runtime 1,981.83 Binary Radio Pulsar Search v1.28 (opencl-ati)
task 279099 runtime 11,162.70 Binary Radio Pulsar Search v1.25 (BRP3cuda32)

---

Congratulations!

My pc did a BRP 1.28 in ~1980 sec. HD6900 Series GPU, i7 with free cpu, 7.0.31, CCC 12.6
On HD5850 a wu takes ~3450sec.
AMD-apps are grown up now.

Hi,

actually the app should do even faster, did you free a CPU core on the machine with the HD 5850 as well? I have a HD 5850 that is doing a single unit in ca 1800 sec (!!) (although that is on a comparatively fast dual core 3 GHZ Core2 (no hyperthreading!) with only one other task running on the CPU.

Cheers
HB

EDIT: Ah I see....2 cards in one box I guess. That is distorting the numbers a bit of course, but I'm glad to see that you see a substantial performance increase.

HB

---

Hi,

actually the app should do even faster, did you free a CPU core on the machine with the HD 5850 as well? I have a HD 5850 that is doing a single unit in ca 1800 sec (!!) (although that is on a comparatively fast dual core 3 GHZ Core2 (no hyperthreading!) with only one other task running on the CPU.

Cheers
HB

OK, its the same machine which has both gpu's installed. CPU is a i7, it's a machine for normal daily business (office and controlsystem design), which sometimes runs 100% for BOINC, actually exclusively for Albert (I'm working here right now).
CPU-load for all 4 cores/8 threads are all well below 30%.
The 'smaller' gpu is in an x8 pcie-slot, maybe this also makes a difference.
And since two generations of gpu's are installed, I never can be shure that I have the optimal drivers installed.
Anyway, I'm happy that the amd gpu's are much more usable now for DC jobs than in earlier days when just Milkyway had a pretty fast amd app.

Alexander

---

Seeing as there are few projects that will crunch WUs on my MacBook's GPU, I'm always interested to see how it pans on a new release.

1. the time estimates are way off. initial estimate was >50 hours. task crunched in about 7 hours. IIRC this is down from 12 hours.
2. the CPU utilization was 0.564 of a core. I think this is high (Collatz grabs just 0.02 of a core). Even so I think this is a reduction
3. i still only got 500 credits. for 7 hours crunching Collatz gives me ~6,500

I'm not whining about the credits - I read, understood and accepted the notice. But this is a feedback thread after all.

I enjoy the credit game as much as the next cruncher but I don't always pick the highest paying project. But i'd like a decent whack even so. What i'd like to know is if and when these apps move to production, what will be the going rate? I have a very slight preference for astronomy projects over mathematical projects but 500 credits won't get my GPU. I know, I'm just one and there's plenty more dedicated fish in the sea. nonetheless, this is my feedback.

I'll keep coming back and crunching a WU each time there's a new app. After all, you went to the trouble of making an OSX app, the least I can do is test it.

Thanks!

I could not resist ..

I 'borrowed' my old testsystem and tried some wu's.
AMD A8 3870 APU: wu 100169 runtime 6,489.60
HD5830 runtime ~ 2,915.96

---

Hi!

Thanks for the feedback!

As for credits, our yardstick used to be SETI@Home, so we tuned the credits so that the CPU apps would grant roughly as many credits as the SETI@HOME CPU apps on the same hardware in the the same time. The GPU apps would grant as many credits per UNIT as the CPU apps (that is: all results are treated equally, whether done on CPU or GPU). With SETI@Home now using the "Credit New" BOINC crediting system, things might have diverged a bit. Still my feeling is that E@H will probably try in the future to stay in line with the bigger (in terms of nr of participants) "mainstream" projects and will not try to "out-credit" all the other projects :-). Just my personal idea, tho.

Cheers
HB

---

Thanks all for testing, the OpenCL app versions are now released on Einstein@Home.

Next we will release the CUDA apps based on the same codebase to Albert@Home. The perfmance increase will not be as high as for the OpenCL versions which still had to to a bit of catching up with the CUDA versions, so we should now have CUDA and OpenCL apps that are on par, performance-wise, so to speak.


CU

HBE

---

Are the CUDA version released?
I am seeing the BRP4 v1.28 version is running on my BOINC menager

Are the CUDA version released?
I am seeing the BRP4 v1.28 version is running on my BOINC menager

Indeed, the Windows CUDA app is out now, I'll launch the Linux and Mac OSX versions today and then write an announcement.

From what I've seen so far the CUDA Windows app should reduce the runtime of a BRP4 task in a way that scales more with the CPU speed than with GPU parameters, e.g. on modern CPUs (nd if running just one unit at a time), the overall runtime should be reduced by roughly 400-700 seconds. For slower cards that take 2 h per unit, this doesn't amount to much, but for faster cards that take less than half an hour, this should be a substantial speed-up. Any feedback on speed-up in different configurations (2 tasks t time, multiple cards,...) are welcome.

Cheers
HB

---

I am looking forward to testing out BRP4 v1.28 for Linux. Thanks for the updates!

---