Dear all,
I would not imagine I would come to ask you... but I have looked around, tried everything and I cannot solve the problem.

In short: I have switched to WIN7 after XP end of maintenance.
I have copied the data folder of SETI (better said, I can seamlessly use without other itches the same data folder from the former XP partition and from the new WIN7 one, starting from where a partition stopped the work) and everything seems to be working al-right, but fact is that I cannot get more than 75% GPU load.
With XP I was (and still am when I use that system) at 95%+, except under heavy use, where it dropped as it should, since it was set at belownormal in mbcuda.cfg

I have reinstalled from scratch several Nvidia drivers (I am now on 280.26) with no avail.

Data:
System: WIN7 64
I7 930 HT
BOINC 6.10.55
GPU 9800 GT
1 core free (but no changes respect to all cores working).
Now set to "high" in mbcuda.cfg. It works more stable respect to "belownormal", with which load oscillates a lot at very low levels even with idle system, which did not happen in XP to this extent, but seems to now cap at 75%, though I can see short overshootings at 90%.

I thank you for your help!
Maybe it is something stupid, but I think I have no more resources...

Sleepy

Are you using the same drivers? That may be a factor. Also, how much RAM do you have? WIN7 uses more than XP. Although if you have >=4GB, RAM isn't your issue.

---

Also: how many WUs are you running at once? On many GPUs, one runs at < 90-100%, so you have headroom to run > 1. For example, on my GTX 680s, I run 3 at a time to get GPU utilization to 95% or so.

Looking at your info, you have a 9800GT. I don't know if that is capable of running > 1 WU at a time. Can anyone let us know who is reading this?

---

I have used at first the same drivers, the latest.
Then, since it was not working right, I tried to downgrade gradually and now I am on the first one to recognise the GT9800.
Of course, this is not advisable in the long term, it was just an experiment and I can revert back to the newest. Nothing against this.
Always made a clean installation.
I have 6 GB RAM.
All settings are the same of the XP64 installation (actually, I am using the very same directories and files, not even a copy of them).

9800GT cannot run more than 1 task at a time.

Only BOINC has been reinstalled in the new system (I have also uniformed BOINC preferences, but errors are possible everywhere...).

Thank you to everybody!

Sleepy

I am not certain about any problems you were having, but you are running an older version of BOINC. Just wondering if a later one would work better for you.

I can see on my systems that even though the GPU is not working 100% I can never seem to get that no matter how many WUs I run on them. I have dedicated cores on the CPU when running APs but when the APs are running, that is when my GPU shows a very wavy usage graph (and it is a very uniform up and down.)

---

SETI@home classic workunits 4,019
SETI@home classic CPU time 34,348 hours

The basic difference between XP and Vista/Win7/8/8.1 is a much more sophisticated driver architecture (WDDM replacing XPDM). That allows the OS to use some of the GPU memory and other resources for OS functionality.

To achieve this on Pre-DirectX 10 hardware (XP was up to DirectX 9c), where certain facilities to do with memory are not the same, they do a kindof emulation in the drivers. This costs some CPU time, Latency, Video Memory and PCI express bus usage when before you might have seen nearly none. These latencies then in turn reduce (bottleneck) some resources you need for the app, reducing apparent utilisation.

There's a certain amount I can do for future applications to hide these increasing latencies, though they are proving non-trivial so taking time.

Other than waiting (for yet more advanced applications) You have 2 key options to improve utilisation with Win7 onwards. Either
- systematically upgrade/replace anything that could be bottlenecking, including using a newer GPU that has the extended DirectX10/11 support in hardware, along with any supporting hardware that's connected such as chipset drivers, and minimising services etc, OR
- Go back to a minimalist 'low tech' setup (which doesn't use the hardware you want for compute)

The first option is obviously the costly & challenging one, but depends on your use of the machine, finances etc. The second option could be better for some people.

It's a situation where I think there is no one-size-fits-all answer, even though I know for myself I tend to sit in the (sometimes frustrating) first camp. One thing that might help decide for you could be comparing the power [, initial cost and] & heat of the old GPU, against new entry-mid options. From where I sit, if I were still stuck with my old 9600 GSOs, I could easily justify replacement with a single new 750ti from power consumption and heat considerations alone (without even thinking about performance yet)

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

The 8500 GT I am using in this machine. According to GPUz it stays in the range of 90-95% load. I have not really been keeping up on the NVIDIA drivers so the ones I am using are kind of old. They might not even be the best ones to use for this old hardware.

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[

The 8500 GT I am using in this machine. According to GPUz it stays in the range of 90-95% load. I have not really been keeping up on the NVIDIA drivers so the ones I am using are kind of old. They might not even be the best ones to use for this old hardware.

hmm, that suggests the OP might need to check DPC latencies of motherboard chipset drivers.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

It's a situation where I think there is no one-size-fits-all answer, even though I know for myself I tend to sit in the (sometimes frustrating) first camp. One thing that might help decide for you could be comparing the power [, initial cost and] & heat of the old GPU, against new entry-mid options. From where I sit, if I were still stuck with my old 9600 GSOs, I could easily justify replacement with a single new 750ti from power consumption and heat considerations alone (without even thinking about performance yet)

I'd certainly concur with that. At the beginning of the month, I decided the time had come to retire an old 9800GT and replace it with a GTX 750Ti



Before: Zotac 9800GT, TDP 125 watts, purchase price (Jan 2009) GBP 77.47 before sales tax.
After: Gainward 'golden sample' GTX 750Ti, TDP 60 watts, purchase price GBP 99.75

And the performance in figures - running MultiBeam/GPU only:

SETI MB_v6, 9800GT
RAC: ~7,500
APR: 133.85

SETI MB_v7, 9800GT
RAC: 3614
APR: 78.65

SETI MB_v7, GTX 750Ti
RAC: 11,141 and still rising
APR: 120.81 x2

The only remaining question - what do I do with a very well used 9800GT? Free to a good home (UK delivery only) if anybody still wants it after those figures.

Thank you Jason and all,
lately I have had some moments of 95%+ load. But then without any apparent reason load goes back toward the 75% level (in some ways graphics appear as if 75% is like 95%, with some ripples up and down this level).

Anyway, Jason can be right that the hardware is now too old for the software.

But then HAL9000 can get consistent results with an even older GPU.

I will make some new checks, possibly doing some tests in a fresh new partition starting from scratch (well, the OS will come from a precooked one I keep in case of emergencies, so it is part of the same strain of installations).

Also, why still using a 9800? Well, in my main crunching machine I have a 650TI installed. This replaced a 9800, which in turn replaced a 9500 in my home PC.
Therefore, just a bit of turnover. Not all my PCs have the best GPU on the market, but the best I can afford overall.
This way I try to maximise my total RAC, even if unfortunately not minimising electricity consumption.

I am keeping an eye on eBay for good occasions for a better GPU, but for the moment I foresee no quick changes.

:-)

Sleepy

While exploring the existing hardware/OS for reasons, you'll definitely want to explore the DPC latencies. Google 'DPC latency Checker' for a holistic sanity checker, or LatencyMon for a more in depth analysis.

The general gist is that as driver models become more complex, but time-to-market pressures stay the same or increase, then quality has to drop... so selection of drivers for every single device in the system becomes more prominent.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Thank you Jason and all,
lately I have had some moments of 95%+ load. But then without any apparent reason load goes back toward the 75% level (in some ways graphics appear as if 75% is like 95%, with some ripples up and down this level).

Anyway, Jason can be right that the hardware is now too old for the software.

But then HAL9000 can get consistent results with an even older GPU.

I will make some new checks, possibly doing some tests in a fresh new partition starting from scratch (well, the OS will come from a precooked one I keep in case of emergencies, so it is part of the same strain of installations).

Also, why still using a 9800? Well, in my main crunching machine I have a 650TI installed. This replaced a 9800, which in turn replaced a 9500 in my home PC.
Therefore, just a bit of turnover. Not all my PCs have the best GPU on the market, but the best I can afford overall.
This way I try to maximise my total RAC, even if unfortunately not minimising electricity consumption.

I am keeping an eye on eBay for good occasions for a better GPU, but for the moment I foresee no quick changes.

:-)

Sleepy

I think the only thing I can point out on my 8500 machine is that I am using the Classic theme instead of the default Windows 7 one. With a 256MB card I have to free up whatever I can to even have it run.

On the bright side even at 75% your 9800 GT is still about 5 times faster than an 8500 GT.

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[

Before: Zotac 9800GT, TDP 125 watts, purchase price (Jan 2009) GBP 77.47 before sales tax.
After: Gainward 'golden sample' GTX 750Ti, TDP 60 watts, purchase price GBP 99.75

Richard, what GFLOP rating does BOINC give for your 750Ti? I replaced a GT 440 with an EVGA "Super-clocked" 750Ti today (because my proprietary [(hp)] rig doesn't have any PCI-e power connectors). BOINC reported 311 GFLOP for the 440, which agrees with the figure on the Wiki NVIDIA GPU page, but that gives 1306 for the 750Ti while BOINC reports 2406.

---

Before: Zotac 9800GT, TDP 125 watts, purchase price (Jan 2009) GBP 77.47 before sales tax.
After: Gainward 'golden sample' GTX 750Ti, TDP 60 watts, purchase price GBP 99.75

Richard, what GFLOP rating does BOINC give for your 750Ti? I replaced a GT 440 with an EVGA "Super-clocked" 750Ti today (because my proprietary [(hp)] rig doesn't have any PCI-e power connectors). BOINC reported 311 GFLOP for the 440, which agrees with the figure on the Wiki NVIDIA GPU page, but that gives 1306 for the 750Ti while BOINC reports 2406.

Interesting, as we (Richard and I) looked at that formula not so long ago and pointed out some issues. The superclocked model should be around 1500, so will probably have to update my Boinc tree and look what's been committed since.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Before: Zotac 9800GT, TDP 125 watts, purchase price (Jan 2009) GBP 77.47 before sales tax.
After: Gainward 'golden sample' GTX 750Ti, TDP 60 watts, purchase price GBP 99.75

Richard, what GFLOP rating does BOINC give for your 750Ti? I replaced a GT 440 with an EVGA "Super-clocked" 750Ti today (because my proprietary [(hp)] rig doesn't have any PCI-e power connectors). BOINC reported 311 GFLOP for the 440, which agrees with the figure on the Wiki NVIDIA GPU page, but that gives 1306 for the 750Ti while BOINC reports 2406.

Boinc reports 2082 GFLOP on my 750TI and that is not overclocked.

---

SETI@home classic workunits 4,019
SETI@home classic CPU time 34,348 hours

Before: Zotac 9800GT, TDP 125 watts, purchase price (Jan 2009) GBP 77.47 before sales tax.
After: Gainward 'golden sample' GTX 750Ti, TDP 60 watts, purchase price GBP 99.75

Richard, what GFLOP rating does BOINC give for your 750Ti? I replaced a GT 440 with an EVGA "Super-clocked" 750Ti today (because my proprietary [(hp)] rig doesn't have any PCI-e power connectors). BOINC reported 311 GFLOP for the 440, which agrees with the figure on the Wiki NVIDIA GPU page, but that gives 1306 for the 750Ti while BOINC reports 2406.

Boinc reports 2082 GFLOP on my 750TI and that is not overclocked.

Found what I was looking for. Our fix doesn't come until 7.3.5 tag committed 25th February:

SHA-1: 3edb124ab4b16492d58ce5a6f6e40c2244c97ed6

* client/scheduler: NVIDIA compute capability 5 GPUS have 128 cores, not 192

NVIDIA has no plans to provide an API that tells you how many cores; I've asked

No acknowledgement that he got that info from Richard, which made it hard [for me] to find in the logs.

--->
Just for now mentally multiply your flops by 2/3 :D

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Before: Zotac 9800GT, TDP 125 watts, purchase price (Jan 2009) GBP 77.47 before sales tax.
After: Gainward 'golden sample' GTX 750Ti, TDP 60 watts, purchase price GBP 99.75

Richard, what GFLOP rating does BOINC give for your 750Ti? I replaced a GT 440 with an EVGA "Super-clocked" 750Ti today (because my proprietary [(hp)] rig doesn't have any PCI-e power connectors). BOINC reported 311 GFLOP for the 440, which agrees with the figure on the Wiki NVIDIA GPU page, but that gives 1306 for the 750Ti while BOINC reports 2406.

Interesting, as we (Richard and I) looked at that formula not so long ago and pointed out some issues. The superclocked model should be around 1500, so will probably have to update my Boinc tree and look what's been committed since.

Quite probably not the same as everyone else is getting :P

18-Apr-2014 11:14:56 [---] Starting BOINC client version 6.12.34 for windows_intelx86
18-Apr-2014 11:14:56 [---] Running as a daemon
18-Apr-2014 11:14:56 [---] NVIDIA GPU 0: GeForce GTX 750 Ti (driver version 33528, CUDA version 6000, compute capability 5.0, 2048MB, 410 GFLOPS peak)

Jason still calls it a Kepler:

Kepler GPU current clockRate = 1332 MHz

BOINC versions 7.3.8 and later will reduce the FLOPs over-statement by 50%:

client/scheduler: NVIDIA compute capability 5 GPUS have 128 cores, not 192. NVIDIA has no plans to provide an API that tells you how many cores; I've asked.

Jason still calls it a Kepler:

Technically it's a Kepler/Maxwell Hybrid :P, but I'll remove the reference in future builds.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Ivan,

I have a EVGA GTX 750 Ti SC also. Boinc says it has 2409 GFLOPS peak

Zalster

18-Apr-2014 11:14:56 [---] Starting BOINC client version 6.12.34 for windows_intelx86
18-Apr-2014 11:14:56 [---] Running as a daemon
18-Apr-2014 11:14:56 [---] NVIDIA GPU 0: GeForce GTX 750 Ti (driver version 33528, CUDA version 6000, compute capability 5.0, 2048MB, 410 GFLOPS peak)

I think this version of BOINC knows nothing beyond the Fermi standard of 32 cores, so read that as 1640 GFLOPS peak.