Did you swap that 7770 back in just to do 2 tasks for my benefit?
yes )
As I said - I'm embarrassed but appreciative of your generosity!!
Quote:
Quote:
What happens with the GPU utilisation factor when you change cards like this? You're running x5 on the 7970 aren't you?
raised to 16-17 with 1 7970 running and about 27-28% with 2 7970 running. yes, x5
I'm sorry, I should have phrased it better. In my mind I was visualising your machine with 10 BRP4 tasks in full flight with your preference setting for GPU utilisation factor set at 0.2. I was wondering what would happen if you shut down that machine and replaced one (or both??) 7970s (each 3GB RAM cards) with a 7770 card with 1GB. I imagine you wouldn't want 5 tasks restarting with only 1GB RAM. So what skilful trickery did you use to get just two tasks to complete on the 7770? Did you somehow change from x5 to x2 or perhaps did you suspend other tasks so that only 2 could run? I was just trying to imagine how you pulled off that impressive feat :-).
Quote:
Quote:
but I think the biggest factor may be PCIe3.
no, i have performed test with 7770@pci-e 2.0, same 3600 for x2.
OK, that is interesting. So both nvidia and AMD cards don't seem to benefit from PCIe3. I'm really surprised that going from SB to IB and from 2.8GHZ to 4.2GHz could drop the CPU time component from 800s to 300s. I'm also surprised about the 20% improvement (confirmed here by Jeroen) from the latest 12.11 beta drivers. Sounds like good news for AMD owners!!
Many thanks for your excellent work and for telling us all about your findings!
OK, that is interesting. So both nvidia and AMD cards don't seem to benefit from PCIe3. I'm really surprised that going from SB to IB and from 2.8GHZ to 4.2GHz could drop the CPU time component from 800s to 300s. I'm also surprised about the 20% improvement (confirmed here by Jeroen) from the latest 12.11 beta drivers. Sounds like good news for AMD owners!!
I started out with the latest stable driver 12.10 in Linux. My overall processing time was 830 seconds. No matter what I did, I could not get the time down much below that. I then upgraded to 12.11 beta and the processing time is down to 623 seconds for a single task. I am not sure what was changed with OpenCL in 12.11 compared to 12.10 but the improvement appears to be fairly significant with this driver and project.
Regarding PCI-E 3.0, I am seeing very little difference in performance with my AMD card between x16 2.0 and x16 3.0. The difference is perhaps 2%. My experience with NVIDIA Kepler has been different however. Here were some tests I ran a while back via the GTX 680 in Linux with driver 295.33:
I am not sure what was changed with OpenCL in 12.11 compared to 12.10 but the improvement appears to be fairly significant with this driver and project.
Yes, indeed. When I got home and had a look at my host with the 7770, I found it was using 12.8. I've now downloaded, built and installed 12.11-beta8 and have restarted crunching. It's just finished the first pair which had been partially crunched under 12.8. Around 11 mins was shaved off the run time and close to 200s off the CPU time. So I'll be enjoying close to 20% performance increase (for a complete run) too, thanks to you guys.
Quote:
My experience with NVIDIA Kepler has been different however. Here were some tests I ran a while back via the GTX 680 in Linux with driver 295.33:
Although I said "Kepler series", I was really only talking about 650 and 650Ti since I don't have anything higher. I should have been more careful to indicate I had no experience with higher end cards and was only talking about those two.
I am not sure what was changed with OpenCL in 12.11 compared to 12.10 but the improvement appears to be fairly significant with this driver and project.
I am not sure what was changed with OpenCL in 12.11 compared to 12.10 but the improvement appears to be fairly significant with this driver and project.
Yes, indeed. When I got home and had a look at my host with the 7770, I found it was using 12.8. I've now downloaded, built and installed 12.11-beta8 and have restarted crunching. It's just finished the first pair which had been partially crunched under 12.8. Around 11 mins was shaved off the run time and close to 200s off the CPU time. So I'll be enjoying close to 20% performance increase (for a complete run) too, thanks to you guys.
I wonder whether the speedup only occurs on newer (GCN perhaps?) cards, because I didn't see any performance improvement on my ole HD 5850 under Linux with the 12.11 beta driver. Can anyone confirm this ?
I am not sure what was changed with OpenCL in 12.11 compared to 12.10 but the improvement appears to be fairly significant with this driver and project.
Release notes say .NET 4.0
.Net is only used for AMD's GPU control application; it doesn't actually run anything on the GPU.
RE: RE: Did you swap that
)
As I said - I'm embarrassed but appreciative of your generosity!!
I'm sorry, I should have phrased it better. In my mind I was visualising your machine with 10 BRP4 tasks in full flight with your preference setting for GPU utilisation factor set at 0.2. I was wondering what would happen if you shut down that machine and replaced one (or both??) 7970s (each 3GB RAM cards) with a 7770 card with 1GB. I imagine you wouldn't want 5 tasks restarting with only 1GB RAM. So what skilful trickery did you use to get just two tasks to complete on the 7770? Did you somehow change from x5 to x2 or perhaps did you suspend other tasks so that only 2 could run? I was just trying to imagine how you pulled off that impressive feat :-).
OK, that is interesting. So both nvidia and AMD cards don't seem to benefit from PCIe3. I'm really surprised that going from SB to IB and from 2.8GHZ to 4.2GHz could drop the CPU time component from 800s to 300s. I'm also surprised about the 20% improvement (confirmed here by Jeroen) from the latest 12.11 beta drivers. Sounds like good news for AMD owners!!
Many thanks for your excellent work and for telling us all about your findings!
Cheers,
Gary.
RE: suspend other tasks so
)
yes, this case.
gpu utilization was about 96%. so, i think 3600s for x2 it is the best of 7770 with fresh 12.11beta.
RE: OK, that is
)
I started out with the latest stable driver 12.10 in Linux. My overall processing time was 830 seconds. No matter what I did, I could not get the time down much below that. I then upgraded to 12.11 beta and the processing time is down to 623 seconds for a single task. I am not sure what was changed with OpenCL in 12.11 compared to 12.10 but the improvement appears to be fairly significant with this driver and project.
Regarding PCI-E 3.0, I am seeing very little difference in performance with my AMD card between x16 2.0 and x16 3.0. The difference is perhaps 2%. My experience with NVIDIA Kepler has been different however. Here were some tests I ran a while back via the GTX 680 in Linux with driver 295.33:
PCI-E 3.0 x16: 2-tasks - 1220-seconds
PCI-E 2.0 x16: 2-tasks - 1380-seconds
PCI-E 2.0 x8: 2-tasks - 1784 seconds
RE: I am not sure what was
)
Yes, indeed. When I got home and had a look at my host with the 7770, I found it was using 12.8. I've now downloaded, built and installed 12.11-beta8 and have restarted crunching. It's just finished the first pair which had been partially crunched under 12.8. Around 11 mins was shaved off the run time and close to 200s off the CPU time. So I'll be enjoying close to 20% performance increase (for a complete run) too, thanks to you guys.
Although I said "Kepler series", I was really only talking about 650 and 650Ti since I don't have anything higher. I should have been more careful to indicate I had no experience with higher end cards and was only talking about those two.
Cheers,
Gary.
RE: I am not sure what was
)
Release notes say .NET 4.0
RE: RE: I am not sure
)
I wonder whether the speedup only occurs on newer (GCN perhaps?) cards, because I didn't see any performance improvement on my ole HD 5850 under Linux with the 12.11 beta driver. Can anyone confirm this ?
Cheers
HB
RE: RE: I am not sure
)
.Net is only used for AMD's GPU control application; it doesn't actually run anything on the GPU.