Well, we haven't clearly identified a 4xxx specific problem so far. We're still investigating the overall numerical accuracy/correctness of our OpenCL apps in general though. Given AMD's driver quality that's no trivial, straight-forward task unfortunately.

Please be patient with us.

Oliver

Q9550 and 6950 here, driver 11.11, average GPU use 60%-70%

Hi,
I'm getting the following message:

22.01.2012 14:54:19 | Albert@Home | Nachricht vom Server: Neuere BOINC Version benötigt. Bitte die neueste Software installieren.

My system is:
22.01.2012 09:31:02 | | Starting BOINC client version 7.0.8 for windows_x86_64
22.01.2012 09:31:02 | | log flags: file_xfer, sched_ops, task
22.01.2012 09:31:02 | | Libraries: libcurl/7.21.6 OpenSSL/1.0.0d zlib/1.2.5
22.01.2012 09:31:02 | | Data directory: C:\ProgramData\BOINC
22.01.2012 09:31:02 | | Running under account Service
22.01.2012 09:31:02 | | Processor: 5 AuthenticAMD AMD Phenom(tm) II X4 840 Processor [Family 16 Model 5 Stepping 3]
22.01.2012 09:31:02 | | Processor: 512.00 KB cache
22.01.2012 09:31:02 | | Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 htt pni cx16 syscall nx lm svm sse4a osvw ibs skinit wdt page1gb rdtscp 3dnowext 3dnow
22.01.2012 09:31:02 | | OS: Microsoft Windows 7: Home Premium x64 Edition, Service Pack 1, (06.01.7601.00)
22.01.2012 09:31:02 | | Memory: 3.75 GB physical, 7.50 GB virtual
22.01.2012 09:31:02 | | Disk: 55.80 GB total, 22.67 GB free
22.01.2012 09:31:02 | | Local time is UTC +1 hours
22.01.2012 09:31:02 | | VirtualBox version: 4.1.8
22.01.2012 09:31:02 | | NVIDIA GPU 0: GeForce GT 430 (driver version 285.62, CUDA version 4.10, compute capability 2.1, 1024MB, 955MB available, 269 GFLOPS peak)
22.01.2012 09:31:02 | | ATI GPU 0: ATI Radeon HD 2300/2400/3200 (RV610) (CAL version 1.4.1016, 368MB, 336MB available, 112 GFLOPS peak)
22.01.2012 09:31:02 | | ATI GPU 1: ATI Radeon HD 2300/2400/3200 (RV610) (CAL version 1.4.1016, 368MB, 336MB available, 112 GFLOPS peak)

Any ideas?

Alexander

---

OK, I found a hint in the HD7900 thread.
BM 7.0.11 is latest version, but the boinc page lists only up to version 7.0.8

Has anyone a link to the newest version?

---

Well, we haven't clearly identified a 4xxx specific problem so far. We're still investigating the overall numerical accuracy/correctness of our OpenCL apps in general though. Given AMD's driver quality that's no trivial, straight-forward task unfortunately.

Please be patient with us.

Oliver

I posted earlier that my 4850 card had 512MB of memory so you said it would be lowered to 490MB
But ageless showed me how to chech for free mem in Boincmanager.
My 512MB card only has 480MB free

Is there any chance that you can lower again to 450MB?

---

Found all I need to know.

Need a better AMD-card.

---

Found all I need to know.

Need a better AMD-card.

I think you could have use of ATI/AMD Catalyst driver 11.12 or 12.1 for your 69xx card.

---

I think you could have use of ATI/AMD Catalyst driver 11.12 or 12.1 for your 69xx card.

Hi,
not really, there are big problems coming up with the new drivers.
On my AMD-System I've upgraded to newest Catalyst driver and BM 7.0.11 with the effect, that collatz is no longer running. Seti has rarely Astropulse wu's for AMD and the memory on my AMD-card is too small for Albert ATI open-cl apps.

On my other system I run older drivers and older BM to be able to run MW, this project needs drivers < 11.10

An intresting fact is, that with the new HD7970 one can run the nVidia OpenCL-app, faster than the CAL-app on HD6970, without the need of latest BM. 6.12.34 is good for that ...

So, after upgrading to latest available versions I have a GPU running dry. It's really a mess with these versions and requirements.

---

I think you could have use of ATI/AMD Catalyst driver 11.12 or 12.1 for your 69xx card.

Hi,
not really, there are big problems coming up with the new drivers.
On my AMD-System I've upgraded to newest Catalyst driver and BM 7.0.11 with the effect, that collatz is no longer running. Seti has rarely Astropulse wu's for AMD and the memory on my AMD-card is too small for Albert ATI open-cl apps.

On my other system I run older drivers and older BM to be able to run MW, this project needs drivers < 11.10

An intresting fact is, that with the new HD7970 one can run the nVidia OpenCL-app, faster than the CAL-app on HD6970, without the need of latest BM. 6.12.34 is good for that ...

So, after upgrading to latest available versions I have a GPU running dry. It's really a mess with these versions and requirements.

Yes, it is a mess with the drivers, I agree on that.
And Collatz is thinking of do an Open-Cl app too, but it will be slower then the old app they say.

---

OK, I found a hint in the HD7900 thread.
BM 7.0.11 is latest version, but the boinc page lists only up to version 7.0.8

Has anyone a link to the newest version?

The Change Log Thread has links.

---

Jord.

BOINC FAQ Service

They say most of your brain shuts down in cryo-sleep. All but the primitive side, the animal side. No wonder I'm still awake.

Today, AMD released a new beta driver, Catalyst 12.2. With this driver, Workunits are using 81% of the GPU (HD 6950@6970 full Shader). It could be even more, because the CPU limits the GPU-load. My Core i7-2600k ( 4.3GHz, only one Albert@home WU is running) is not able to deliver enough performance to bring GPU-load to 100%.
If the CPU starts crunching other WUs (one core is exclusivly for Albert@home GPU-Workunit), the GPU-load drops to 75% and lower.

Is there a chance, that CPU-load for GPU-WUs could be lowered in the future?

OK, I found a hint in the HD7900 thread.
BM 7.0.11 is latest version, but the boinc page lists only up to version 7.0.8

Has anyone a link to the newest version?

FWIW, you'll find available BOINC packages here: http://boinc.berkeley.edu/dl

Oliver

PS: 7.0.12 is out

Is there any chance that you can lower again to 450MB?

If you ask whether our app's memory requirements might be reduced to 450 MB at some point, then probably since we always strive to support as many cards as possible by keeping our requirements as low as possible. Right now we can't due to other technical constraints.

Oliver

Is there a chance, that CPU-load for GPU-WUs could be lowered in the future?

Yes, we have some ideas for that in the pipeline, it's just a matter of time to implement and test them. However, keep in mind that not all algorithms can be ported 100% to the GPU. This really depends on the problem and our current implementation is already not that bad after all.

Right now we still need to focus on getting a first stable and numerically correct release out of the door -- and apart from our own code we largely depend on BOINC's readiness WRT OpenCL and proper AMD drivers, both are not easy to come by.

Best
Oliver

Flagging a weird one.
I caught this one running at zero GPU load for more than an hour, before I aborted it. http://albert.phys.uwm.edu/result.php?resultid=107592

It had been running for well over 4 hours and had an estimated time of another 4 hours, was at 41% done. Its progress was increasing. But still, I didn't trust it.

---

Jord.

BOINC FAQ Service

They say most of your brain shuts down in cryo-sleep. All but the primitive side, the animal side. No wonder I'm still awake.

OK, I ended up aborting all my work in cache, after I found a couple of things.

a) For some unknown reason running an Albert task on my ATI GPU would make my internet connection stop working, leaving me with loads upon loads of page loading faults, even between same forum pages. Disable the Albert task and the internet would fly.

b) Checking in GPU-Z I found that all these tasks would start with some GPU load before it tapering off to no load at all, or only a small bump every 5 minutes or so. I've seen this happen before with Albert GPU tasks. A reboot of the system didn't help in this matter.

I'll test again after I found some I/O recording and debugging programs, to see if I can show you what I see.

---

Jord.

BOINC FAQ Service

They say most of your brain shuts down in cryo-sleep. All but the primitive side, the animal side. No wonder I'm still awake.

Hi,
yesterday I installed my new system with two ATI-GPU's, latest BM 7.0.11 , latest AMD drivers and an outdated nVidia GPU (9500).
I've seen that the GPU usage is very very low, so it makes sense to run more than one wu on every GPU. Is there an app_info available for that purpose?

Alexander

---

Hi Alex,

Hi,
yesterday I installed my new system with two ATI-GPU's, latest BM 7.0.11 , latest AMD drivers and an outdated nVidia GPU (9500).
I've seen that the GPU usage is very very low, so it makes sense to run more than one wu on every GPU. Is there an app_info available for that purpose?

We don't provide an app_info.xml but I'm sure people will help you prepare one yourself. Also, BOINC 7.0.12 is the latest and the GPU usage is usually between 60-80%. However, this is only the case when there is enough CPU power left to feed the GPU (see previous postings).

Best,
Oliver

Hi Oliver,
Thx for your reply.

I checked the gpu usage over a longer period now; there are longer gaps with a usage between 0 and 3% (which caused me to say 'very very low'), shorter periods with ~35% and long times with a usage between 65 and 75%.
This is the on-chip GPU of my FM1 Cpu.
I got the whole system (Cpu, Mainboard and Ram) for ~250.- Euro.
A good GPU costs about the same price and I don't want to run that one at ~30% as on my HD5830, this is a wast of resources and its expensive.
Even for my on-chip gpu I calculate:
1 wu in ~ 7 hrs @ 65% average
2 wu's in ~ 11 hrs @ 95% average --> better performance. :-))

So if anyone can compile an app_info , PLEASE HELP!

Alexander

PS: intresting, the next wu is after 1:30:06 at 56%, quite faster than the first ones!
BTW, the first two completed and validated, thats much better than the CPU wu's ...

---

This is the on-chip GPU of my FM1 Cpu.
I got the whole system (Cpu, Mainboard and Ram) for ~250.- Euro.

Now that's interesting. Can you post a link to that box? I'd like to take a look at the specs! Is this using an AMD APU or similar?

BTW, the first two completed and validated, thats much better than the CPU wu's ...

Good news :-)


Oliver


PS: we identified and fixed an issue that was causing a number of validation issues. v1.20 will be released short and should improve things.