I was running CPU Benchmarks (to test apps elsewhere) and noticed an Einstein WU kept running on the intel_gpu, with LAIM either on or off.
I thought this might skew the actual benchmarks, but it didn't make much difference:
06/09/2013 19:19:00 | | 4116 floating point MIPS (Whetstone) per CPU
06/09/2013 19:19:00 | | 13908 integer MIPS (Dhrystone) per CPU
Einstein suspended,
06/09/2013 19:20:23 | | 4118 floating point MIPS (Whetstone) per CPU
06/09/2013 19:20:23 | | 14334 integer MIPS (Dhrystone) per CPU
I was running CPU Benchmarks (to test apps elsewhere) and noticed an Einstein WU kept running on the intel_gpu, with LAIM either on or off.
Correct operation when Boinc does a Benchmark is for the app to stop Crunching, but remain in memory, and resume when the Benchmark has completed, if it doesn't stop crunching then there is an app or api problem.
Another oberservation regarding the main memory bandwdith requirements for the Intel GPU: I switched from Einstein on the other CPU cores to SIMAP, which appears less demanding, and thus decreased my time per 2 WUs from ~22 min to ~20 min - that's a 10% decrease!
Setup: HD4000 @ 1.25 GHz, i7 3770K @ 4.1 GHz with HT on, DDR3-2400 9-11-11. So crunching on the moderately OC'ed CPU will require a bit more bandwidth than non-OC'ed, but the main memory is also faster than usual.
Forgot.. overall I'm seeing this now:
- 2 x E@H @iGPU, 7 x POEM on CPU + nVidia: ~24 mins / 2 Einsteins
- 2 x E@H @iGPU, 1 x GPU-Grid on CPU + nVidia, 6 E@H on CPU: ~22 mins / 2 Einsteins
- 2 x E@H @iGPU, 1 x GPU-Grid on CPU + nVidia, 6 Simap on CPU: ~20 mins / 2 Einsteins
It took me quite some time to get the HD4600 in my new haswell working, despite a monitor connected to the VGA Port as "Dummy Plug" and the multidisplay option enabled in the BIOS of my ASUS H87M Plus board. Windows and GPU-Z didn't recognize the HD4600 next to the GT650Ti in the PCIex16 slot.
After a BIOS-Update to Version 0603 from 08/06/2013 it suddenly worked. Now I'm happily crunching 6 jobs at any time ( 2 x cuda, 2 x intel, 2 x CPU). The system consumes a 130-140W at this load, the GPUs are loaded around 95%.
So saving energy is possible, my Athlon 1000XP System around the turn of the century consumed 170W idle.
Despite being a Linux-Lover since 1993 I'll boot my system more often into Windows to see where this combination gets me.
You could likely improve power efficiency further by lowering the iGPU voltage. Mine had lot's of headroom, -0.19 V at +100 MHz! Would be nice to see other peoples numbers. Of, and the CPU also has quite some undervolting headroom, if you did not yet touch that. And you could put the power savings straight into running a few more threads ;)
My iGPU now runs at 1200MHz and I allowed a third CPU thread. Surprisingly it does not starve the GPU threads. Adding more GPU threads IMHO makes no sense as the GPUs are essentially fully loaded (>95%).
- 2 x E@H @iGPU, 7 x POEM on CPU + nVidia: ~24 mins / 2 Einsteins
- 2 x E@H @iGPU, 1 x GPU-Grid on CPU + nVidia, 6 E@H on CPU: ~22 mins / 2 Einsteins
- 2 x E@H @iGPU, 1 x GPU-Grid on CPU + nVidia, 6 Simap on CPU: ~20 mins / 2 Einsteins
With a UEFI Update I can finally set memory timings manually again. Changed Command Rate from 2T to 1T and get runtime reductions from 24 to 23 mins under heavy load! I guess I should have taken more precise measurements, but the improvement is >35s for sure. Not bad for a single timing parameter!
In the other scenarios the iGPU is less starved for bandwidth and hence the benefit drops, but I'm regularly seeing up to 10s below 20 mins not, whereas it was a few seconds over 20 mins before in the last scenario.
Oh, POEM runtimes also decreased significantly.. around 2.5%, which translates to ~20k credits/day more.
Hi I tried running the Intel GPU app but it seems like it doesn't like my computer; I have a Mid 2012 MacBook Pro on Windows 7 64bit. CPU is Ivy Bridge i5 3210M, and the OpenCL Driver is 10.18.10.3345 for the HD 4000.
Also, within seconds of the WU starting, my gfx driver stops, the screen goes black, then flashes back with the icon on the bottom right saying the driver crashed and has recovered.
Hi I tried running the Intel GPU app but it seems like it doesn't like my computer; I have a Mid 2012 MacBook Pro on Windows 7 64bit. CPU is Ivy Bridge i5 3210M, and the OpenCL Driver is 10.18.10.3345 for the HD 4000.
Also, within seconds of the WU starting, my gfx driver stops, the screen goes black, then flashes back with the icon on the bottom right saying the driver crashed and has recovered.
I was running CPU Benchmarks
)
I was running CPU Benchmarks (to test apps elsewhere) and noticed an Einstein WU kept running on the intel_gpu, with LAIM either on or off.
I thought this might skew the actual benchmarks, but it didn't make much difference:
06/09/2013 19:19:00 | | 4116 floating point MIPS (Whetstone) per CPU 06/09/2013 19:19:00 | | 13908 integer MIPS (Dhrystone) per CPU
Einstein suspended,
06/09/2013 19:20:23 | | 4118 floating point MIPS (Whetstone) per CPU
06/09/2013 19:20:23 | | 14334 integer MIPS (Dhrystone) per CPU
Hardly a show stopper!
RE: I was running CPU
)
Correct operation when Boinc does a Benchmark is for the app to stop Crunching, but remain in memory, and resume when the Benchmark has completed, if it doesn't stop crunching then there is an app or api problem.
Claggy
Another oberservation
)
Another oberservation regarding the main memory bandwdith requirements for the Intel GPU: I switched from Einstein on the other CPU cores to SIMAP, which appears less demanding, and thus decreased my time per 2 WUs from ~22 min to ~20 min - that's a 10% decrease!
Setup: HD4000 @ 1.25 GHz, i7 3770K @ 4.1 GHz with HT on, DDR3-2400 9-11-11. So crunching on the moderately OC'ed CPU will require a bit more bandwidth than non-OC'ed, but the main memory is also faster than usual.
MrS
Scanning for our furry friends since Jan 2002
Forgot.. overall I'm seeing
)
Forgot.. overall I'm seeing this now:
- 2 x E@H @iGPU, 7 x POEM on CPU + nVidia: ~24 mins / 2 Einsteins
- 2 x E@H @iGPU, 1 x GPU-Grid on CPU + nVidia, 6 E@H on CPU: ~22 mins / 2 Einsteins
- 2 x E@H @iGPU, 1 x GPU-Grid on CPU + nVidia, 6 Simap on CPU: ~20 mins / 2 Einsteins
So overall already a 20% difference!
MrS
Scanning for our furry friends since Jan 2002
Here my Intel IGP
)
Here my Intel IGP report:
It took me quite some time to get the HD4600 in my new haswell working, despite a monitor connected to the VGA Port as "Dummy Plug" and the multidisplay option enabled in the BIOS of my ASUS H87M Plus board. Windows and GPU-Z didn't recognize the HD4600 next to the GT650Ti in the PCIex16 slot.
After a BIOS-Update to Version 0603 from 08/06/2013 it suddenly worked. Now I'm happily crunching 6 jobs at any time ( 2 x cuda, 2 x intel, 2 x CPU). The system consumes a 130-140W at this load, the GPUs are loaded around 95%.
So saving energy is possible, my Athlon 1000XP System around the turn of the century consumed 170W idle.
Despite being a Linux-Lover since 1993 I'll boot my system more often into Windows to see where this combination gets me.
This is the Windows personality of the machine:
http://einsteinathome.org/host/5197474
and this the Penguin identity
http://einsteinathome.org/host/4284636
Turnaround times for 2 Intel GPU E@H jobs are ~23 min.
Armin
Thanks for sharing! You
)
Thanks for sharing!
You could likely improve power efficiency further by lowering the iGPU voltage. Mine had lot's of headroom, -0.19 V at +100 MHz! Would be nice to see other peoples numbers. Of, and the CPU also has quite some undervolting headroom, if you did not yet touch that. And you could put the power savings straight into running a few more threads ;)
MrS
Scanning for our furry friends since Jan 2002
Thank you for the hints
)
Thank you for the hints MrS!
My iGPU now runs at 1200MHz and I allowed a third CPU thread. Surprisingly it does not starve the GPU threads. Adding more GPU threads IMHO makes no sense as the GPUs are essentially fully loaded (>95%).
Undervolting experiments follow tomorrow, I'll report back.
Good night and good luck ;-)
Armin
RE: - 2 x E@H @iGPU, 7 x
)
With a UEFI Update I can finally set memory timings manually again. Changed Command Rate from 2T to 1T and get runtime reductions from 24 to 23 mins under heavy load! I guess I should have taken more precise measurements, but the improvement is >35s for sure. Not bad for a single timing parameter!
In the other scenarios the iGPU is less starved for bandwidth and hence the benefit drops, but I'm regularly seeing up to 10s below 20 mins not, whereas it was a few seconds over 20 mins before in the last scenario.
Oh, POEM runtimes also decreased significantly.. around 2.5%, which translates to ~20k credits/day more.
MrS
Scanning for our furry friends since Jan 2002
Hi I tried running the Intel
)
Hi I tried running the Intel GPU app but it seems like it doesn't like my computer; I have a Mid 2012 MacBook Pro on Windows 7 64bit. CPU is Ivy Bridge i5 3210M, and the OpenCL Driver is 10.18.10.3345 for the HD 4000.
The error seems to be a driver issue, but I'm not sure.
Here's two of the results that errored out:
http://einsteinathome.org/task/413691887
http://einsteinathome.org/task/413473803
Also, within seconds of the WU starting, my gfx driver stops, the screen goes black, then flashes back with the icon on the bottom right saying the driver crashed and has recovered.
RE: Hi I tried running the
)
http://einsteinathome.org/node/197300&nowrap=true#128258