From what I've read so far, volunteers have reported that (sometimes after reserving a CPU core for the GPU work), the GPU load on the Intel HD GPUs is close to 95% with just one task, so it doesn't seem necessary to run more than one task in parallel to saturate these tiny internal GPUs.
Unfotunatelly it doesn't work on my Win 7 machine. But the cause is the Intel driver that doesn't work on Win 7 x32 in patched mode (when it uses over 4 GBs of RAM) though NVidia does. It takes about 20 minutes to make a dummy plug. Have to notice that some Mobos have no VGA jack and have a DVI jack where the plug can not be installed on due to lack of analog pins and VGA jack.
Yes. I used this resource to build the plug. But again. This will not work on my mobo because it doesn't have analog lines in the DVI jack, so I can not even put this plug on the jack - no such holes in the jack.
Alternative way is to switch to DVI monitor and to connect it to an Intel GPU, while NVidia will run even without monitor. But this way I can not play games normally on NVidia or have to switch the monitor to NVidia. BTW, Nvidia mobile drivers allow to use internal GPU in usual work and to switch to NVidia when it is necessary using only one output for both adapters. Guess, usual drivers also have this feature, but it turned off by default and can be switched on by some keys in registry.
But the most frustrating thing is that Intel drivers don't work on my machine in 32-bit patched mode, though all other drivers for all other devices in my rig do (I have to use Win 7 x32 but I want to use more then 4 Gigs of RAM).
That's the point I have to forget to use Intel GPU yet.
Bernd commented over at Albert that he corrected the scheduler issue. It seems to be allocating the correct amount of work now. Thanks Bernd and thanks for the app, I can finally use the iGPU for something other than running a screen
It seems it is still over-fetching. When I set to .75 of a day it downloaded approx 300 tasks (each and on 3 different machines). They have been running through for the last 2 days now and it still hasn't cleared the backlog yet.
Anyway apart from that issue they are going through ranging from 12 to 14 minutes a work unit, when I leave a CPU core free.
Bernd commented over at Albert that he corrected the scheduler issue. It seems to be allocating the correct amount of work now. Thanks Bernd and thanks for the app, I can finally use the iGPU for something other than running a screen
It seems it is still over-fetching. When I set to .75 of a day it downloaded approx 300 tasks (each and on 3 different machines). They have been running through for the last 2 days now and it still hasn't cleared the backlog yet.
Anyway apart from that issue they are going through ranging from 12 to 14 minutes a work unit, when I leave a CPU core free.
You might want to try Boinc 7.2.4 on that host, it seems stable enough, at least better than 7.2.1
But the most frustrating thing is that Intel drivers don't work on my machine in 32-bit patched mode, though all other drivers for all other devices in my rig do (I have to use Win 7 x32 but I want to use more then 4 Gigs of RAM).
This sounds more like "hacked" than "patched" and seems to be a rather special case. I think I'd rather run an app in a virtual 32 bit machine than to go for complete 32 bit. But you may already have tried that.
It seems it is still over-fetching. When I set to .75 of a day it downloaded approx 300 tasks (each and on 3 different machines). They have been running through for the last 2 days now and it still hasn't cleared the backlog yet.
Anyway apart from that issue they are going through ranging from 12 to 14 minutes a work unit, when I leave a CPU core free.
You might want to try Boinc 7.2.4 on that host, it seems stable enough, at least better than 7.2.1
BOINC v7.2.4 overfetches for me, 7.1.18 does not. I had to revert all but 1 of my 7.2.4 boxes back to 7.1.18 to stop the overfetching (among other problems).
Edit: v7.2.1 was an absolute disaster.
Edit 2: Correction, The last 7.2.4 box started overfetching, had to back that one off to 7.1.18 to fix it. What's also irritating is that every time we go from 7.2.4 to 7.1.18, we have to run BOINC setup twice: setup for 7.1.18 and then a repair. Update: 7.2.4 is a disaster too. Sometimes when rebooting with 7.2.4, BOINC fails to detect any GPUs even though Afterburner and Windows see them correctly. This was supposed to be fixed from 7.2.1. It's not.
I was surprised that you guys didn't report any benefit of running 2 WUs in parallel. In one of the 1st posts I did report a substantial improvement for my main machine. Now I've collected some more data on another machine, an i7 3770 @ 4.0 GHz, HT enabled, HD4000 @ 1250 MHz, DDR3-1866 9-10-9, Win7 64:
1 WU: 6954 RAC (averaged over 19 WUs)
2 WU: 8009 RAC (averaged over 31 WUs)
Sure the GPU utilization only inncreased from "95% with occasional drops" to "98-100%", but utilization is not everything (we don't really know what counts towards being busy and what does not, right?). By now I've seen a 23% improvement on host 1 (only averaged over 5 WUs) and 15% improvement on host 2. That's enough for me to keep it running like this and I encourage others to try it out themslves.
Notes:
- I had one task failure upon switching host 2 into running 2 WUs, but otherwise things appear fine
- In both cases 3 POEM WUs and 4 CPU WUs (WCG and/or SIMAP) were running along the Einstein(s)
- The POEMs reduce the Einstein throughput, since both long for memory bandwidth, but for comparing 1 WU vs. 2 WUs that's no problem
RE: From what I've read so
)
Agree with that. 95% load on my 4600 with 1 task:
I've tested running x2 on my
)
I've tested running x2 on my HD 4000 and the completion times doubled, so at least for me there was no gain in running more than one task at a time.
Times running x1: ~900-1200s
Times running x2: ~1900-2200s
And a few results with times outside the above stated.
RE: RE: Unfotunatelly it
)
Yes. I used this resource to build the plug. But again. This will not work on my mobo because it doesn't have analog lines in the DVI jack, so I can not even put this plug on the jack - no such holes in the jack.
Alternative way is to switch to DVI monitor and to connect it to an Intel GPU, while NVidia will run even without monitor. But this way I can not play games normally on NVidia or have to switch the monitor to NVidia. BTW, Nvidia mobile drivers allow to use internal GPU in usual work and to switch to NVidia when it is necessary using only one output for both adapters. Guess, usual drivers also have this feature, but it turned off by default and can be switched on by some keys in registry.
But the most frustrating thing is that Intel drivers don't work on my machine in 32-bit patched mode, though all other drivers for all other devices in my rig do (I have to use Win 7 x32 but I want to use more then 4 Gigs of RAM).
That's the point I have to forget to use Intel GPU yet.
RE: Bernd commented over at
)
It seems it is still over-fetching. When I set to .75 of a day it downloaded approx 300 tasks (each and on 3 different machines). They have been running through for the last 2 days now and it still hasn't cleared the backlog yet.
Anyway apart from that issue they are going through ranging from 12 to 14 minutes a work unit, when I leave a CPU core free.
BOINC blog
RE: RE: Bernd commented
)
You might want to try Boinc 7.2.4 on that host, it seems stable enough, at least better than 7.2.1
Claggy
RE: But the most
)
This sounds more like "hacked" than "patched" and seems to be a rather special case. I think I'd rather run an app in a virtual 32 bit machine than to go for complete 32 bit. But you may already have tried that.
Work fetch seems to be fine for me now.
MrS
Scanning for our furry friends since Jan 2002
RE: But you may already
)
Exactly. I need pure 32-bit and FoxPro 2.5 for DOS that is also 32-bit, but starts up through the DOS 16-bit environment, that is absent in Win x64.
RE: You might want to try
)
Upgraded one of the 4 machines to 7.2.4 and running down its Einstein cache at the moment. Once done I will let it loose and see how much it fetches.
BOINC blog
RE: RE: It seems it is
)
BOINC v7.2.4 overfetches for me, 7.1.18 does not. I had to revert all but 1 of my 7.2.4 boxes back to 7.1.18 to stop the overfetching (among other problems).
Edit: v7.2.1 was an absolute disaster.
Edit 2: Correction, The last 7.2.4 box started overfetching, had to back that one off to 7.1.18 to fix it. What's also irritating is that every time we go from 7.2.4 to 7.1.18, we have to run BOINC setup twice: setup for 7.1.18 and then a repair. Update: 7.2.4 is a disaster too. Sometimes when rebooting with 7.2.4, BOINC fails to detect any GPUs even though Afterburner and Windows see them correctly. This was supposed to be fixed from 7.2.1. It's not.
Running 2 WUs on an
)
Running 2 WUs on an HD4000
I was surprised that you guys didn't report any benefit of running 2 WUs in parallel. In one of the 1st posts I did report a substantial improvement for my main machine. Now I've collected some more data on another machine, an i7 3770 @ 4.0 GHz, HT enabled, HD4000 @ 1250 MHz, DDR3-1866 9-10-9, Win7 64:
1 WU: 6954 RAC (averaged over 19 WUs)
2 WU: 8009 RAC (averaged over 31 WUs)
Sure the GPU utilization only inncreased from "95% with occasional drops" to "98-100%", but utilization is not everything (we don't really know what counts towards being busy and what does not, right?). By now I've seen a 23% improvement on host 1 (only averaged over 5 WUs) and 15% improvement on host 2. That's enough for me to keep it running like this and I encourage others to try it out themslves.
Notes:
- I had one task failure upon switching host 2 into running 2 WUs, but otherwise things appear fine
- In both cases 3 POEM WUs and 4 CPU WUs (WCG and/or SIMAP) were running along the Einstein(s)
- The POEMs reduce the Einstein throughput, since both long for memory bandwidth, but for comparing 1 WU vs. 2 WUs that's no problem
MrS
Scanning for our furry friends since Jan 2002