Things are looking up. I just installed catalysts 11.11 on my Windows 7 system and where before the CPU usage would be maxed out at a full core from the beginning of the task, this time it'll only be at max ~16% for a while, before going back down to 2-5%. At least, for the first 30 minutes that the task is running.

What did I do?
I had 11.6 with SDK 2.5 installed before and I just installed the display driver and the updated SDK over the 11.6 ones (no uninstallation). Then, despite nothing saying I needed that, I rebooted. Just to be sure about that, k? ;-)

---

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.

One hour, 40 minutes. CPU usage, 1-2%.

---

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.

Huh. I have 11.11 and 2.5 installed, but I get the full CPU usage.

The only difference is I am on XP64. When I get back home next week, I will try it on a win7 machine.

---

Dublin, California
Team: SETI.USA

[quote]Things are looking up. I just installed catalysts 11.11 on my Windows 7 system and where before the CPU usage would be maxed out at a full core from the beginning of the task, this time it'll only be at max ~16% for a while, before going back down to 2-5%. At least, for the first 30 minutes that the task is running.

What did I do?
I had 11.6 with SDK 2.5 installed before and I just installed the display driver and the updated SDK over the 11.6 ones (no uninstallation). Then, despite nothing saying I needed that, I rebooted. Just to be sure about that, k? ;-)

12/9/2011 9:54:31 AM | | ATI GPU 0: AMD Radeon HD 6900 series (Cayman) (CAL version 1.4.1607, 2048MB, 2016MB available, 6758 GFLOPS peak)
12/9/2011 9:54:31 AM | | ATI GPU 1: AMD Radeon HD 6900 series (Cayman) (CAL version 1.4.1607, 2048MB, 2016MB available, 6758 GFLOPS peak)
12/9/2011 9:54:31 AM | | OpenCL: ATI GPU 0: Cayman (driver version CAL 1.4.1607 (VM), device version OpenCL 1.1 AMD-APP-SDK-v2.5 (793.1), 4096MB)
12/9/2011 9:54:31 AM | | OpenCL: ATI GPU 1: Cayman (driver version CAL 1.4.1607 (VM), device version OpenCL 1.1 AMD-APP-SDK-v2.5 (793.1), 4096MB)
12/9/2011 9:54:31 AM | | ATI GPU is OpenCL-capable

I have Win7-64b and 11.11 drivers for HD6990 and the latest BOINC 7.0 Beta client and still get 1.5 hr tasks with 95% CPU usage on 2 cores that support both GPU's.

Haven't seen any improvements yet and that is funny because the ATI GPU's are capable of TeraFlops compute power...

Well, I can only state what I see. 3 hours run time, still only max 2% CPU usage.
I even played some Skyrim in between (with BOINC exited).

Perhaps:
1. It's because it's a HD4850 - 1GB, which only does OpenCL 1.0
Perhaps that OpenCL 1.1 dictates the use of a full core in the newer drivers.
2. It's because I didn't clean up/uninstall the old drivers, but just installed over them for the first time in many moons.
3. It's because of irregularities on "your" system. (no flame)
4. It'll kick in to higher gear at a later stage (50% done or so).

---

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.

Well, I can only state what I see. 3 hours run time, still only max 2% CPU usage.
I even played some Skyrim in between (with BOINC exited).

Perhaps:
1. It's because it's a HD4850 - 1GB, which only does OpenCL 1.0
Perhaps that OpenCL 1.1 dictates the use of a full core in the newer drivers.
2. It's because I didn't clean up/uninstall the old drivers, but just installed over them for the first time in many moons.
3. It's because of irregularities on "your" system. (no flame)
4. It'll kick in to higher gear at a later stage (50% done or so).

Yup, could be specific to the HD6990...

One thing I wish, for whatever reason they seem to limit the amount of GPU power to 75% more or less and it doesn't matter if it's CUDA or OpenCL... Maybe there is some delay thing or other compute limiting thing in use, I have no idea. But, I wish they would allow the option to open things up to full maximum speed because all but one of my systems are headless dedicated BOINC crunchers..

As it is now, seems like the HD6990 is used as a math coprocessor or something, so why not use it to it's fullest capability? Heck, each GPU has 2Gig DDR5, that should be enough to load everything inside I would speculate. In any case, it's nice to have an ATI task at hand... but it is far from optimized...

However, to compare, DnetC and Moo! Wrapper are the same way with the HD6990... they always need one free core per GPU to run full speed... Milkyway uses no real CPU power and collatz does okay as well.

And yes, I hear comments here and there about a bug and such... ATI Stream drivers aren't as mature as Nvidia and not as well supported I suppose... and the two boards (6990 vs. 4850) are not the same animal anyway.

Sooo, it's funny the 4850 does as well as the 6990 since the 6990 is supposed to have a much more powerful GPU on it... It simply doesn't make sense that a lower end GPU chip does as well as the newer high end GPU chip... therefore I think about CPU software delays or limitations of some sort...

Sigh...

Crunch on!

8-)

Heck, each GPU has 2Gig DDR5, that should be enough to load everything inside I would speculate.

I don't know how big the kernels are that run these tasks, but the developers at Einstein first try to get everything a bit stable and working, before they think about optimizing.

However, to compare, DnetC and Moo! Wrapper are the same way with the HD6990... they always need one free core per GPU to run full speed... Milkyway uses no real CPU power and collatz does okay as well.

But none of those are OpenCL. They're all BrookGPU/CAL. So I'm not sure you can really compare them. It's like comparing a 1980s Ford and a present one. Both are cars, both drive miles, both had all kinds of new things, but you can't truly compare them.

ATI Stream drivers aren't as mature as Nvidia and not as well supported I suppose...

Again, you can't compare these. Nvidia has neigh on no OpenCL drivers, ATI is far ahead when it comes to those drivers, since they first had Close to metal, followed by ATI Stream programming, via BrookGPU to Stream SDK, followed by now OpenCL. ATI have been at it since 2004, compared with Nvidia who only started in 2007.

Yes, CUDA is more mature than OpenCL, but that's not weird as it's had more time to mature now than OpenCL has had. As said, Nvidia's OpenCL drivers are either nowhere to be seen or when they are available, very minute and full of bugs.

Sooo, it's funny the 4850 does as well as the 6990 since the 6990 is supposed to have a much more powerful GPU on it... It simply doesn't make sense that a lower end GPU chip does as well as the newer high end GPU chip... therefore I think about CPU software delays or limitations of some sort...

The difference is that my GPU runs work in 9 hours, yours does it in one hour. Doesn't matter how much CPU is used then. ;-)

---

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.

At almost 90% and still no more than 2% CPU being used.
Here, Windows Resource Monitor even says 1.35% on average. (Warning: Screen shot at 1920px x 1080px)

---

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.

Interesting... Well, I'm sure things will improve as time goes by... for some reason I got the impression the 4850 was doing the tasks the same speed as the HD6990... my bad I suppose.

Also, maybe someday, we will see this project show up on BOINCStats or something... or maybe I negelected to read something somewhere too...

8-)

Huh. I have 11.11 and 2.5 installed, but I get the full CPU usage.

The only difference is I am on XP64. When I get back home next week, I will try it on a win7 machine.

No difference on win7 64. 100% CPU required per GPU (5870).

I had 11.6 with SDK 2.5 installed before and I just installed the display driver and the updated SDK over the 11.6 ones (no uninstallation).

Can you explain more, please? What updated SDK? I can't find anything higher than 2.5, which I already have. Which display driver?

---

Dublin, California
Team: SETI.USA

In the Catalysts installer there's this option where one can choose to update separate parts, like the display driver, the HDMI driver, the Catalysts Control Center, and the SDK.

Going from 11.6 -> 11.11 I only chose to update the display driver and the SDK, all the rest of the options I had unchecked. Like so:


(This is the screen shot for the 11.12 drivers)

I upgraded to 11.12, chose all available options in the installer this time.
Upon BOINC start-up, Albert uses one core to full for around 15-20 seconds, then usage goes down to the previous 1-2%. I'll keep an eye on 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.

Did you check your GPU-utilization with GPU-Z?

I have a 5750 card on a nehalem system. If 8 Einstein threads are running on the CPU, then the GPU utilization for the additional Albert application is only 0%.

Now I changed the settings so that only 90% of the CPU could be used for boinc. Thus the GPU thread gets more time it seems, the GPU load is back to ~90-91%.

CPU-load is still only 2-3%, as u encounter it. However, I thought this should be ok, it is an OpenCL-app, so why should it use the CPU?

I am worried more about the GPU. Especially, because the progress bar in boinc was increased as if there would be normal calculations ...

I also tried to increase the priority of the OpenCL.exe, but it didnt help.

Driver version: 8.902 (Cat 11.10), Windows Server 2008R2 (like Win7)

However, I thought this should be ok, it is an OpenCL-app, so why should it use the CPU?

How do you think that data gets from the hard drive into the videocard's memory? The CPU will do the translation of the data in the task to something the GPU understands, transport this translated data over to the GPU, then wait for the GPU to be ready with it, transport it back to PC memory, translate it back into something the humans understand and write it to disk.

No GPU application can run without CPU usage.

By the way, I since changed videocard, went to a HD6850 2GB and am very happy with that one. :)

---

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.

However, I thought this should be ok, it is an OpenCL-app, so why should it use the CPU?

How do you think that data gets from the hard drive into the videocard's memory? The CPU will do the translation of the data in the task to something the GPU understands, transport this translated data over to the GPU, then wait for the GPU to be ready with it, transport it back to PC memory, translate it back into something the humans understand and write it to disk.

No GPU application can run without CPU usage.

By the way, I since changed videocard, went to a HD6850 2GB and am very happy with that one. :)

Totally ok, but the point is that you don't need 100% of the CPU for that. As I said, I have 2-3% CPU load for the OpenCL client and the GPU is @90%. Seems fine to me.
I would complain, if it were 100% cpu load *g*

*double post*

However, I thought this should be ok, it is an OpenCL-app, so why should it use the CPU?

How do you think that data gets from the hard drive into the videocard's memory? The CPU will do the translation of the data in the task to something the GPU understands, transport this translated data over to the GPU, then wait for the GPU to be ready with it, transport it back to PC memory, translate it back into something the humans understand and write it to disk.

No GPU application can run without CPU usage.

By the way, I since changed videocard, went to a HD6850 2GB and am very happy with that one. :)

Totally ok, but the point is that you don't need 100% of the CPU for that. As I said, I have 2-3% CPU load for the OpenCL client and the GPU is @90%. Seems fine to me.
I would complain, if it were 100% cpu load *g*

There was the "opencl 100% cpu bug" with older drivers. You use a opencl application? A cpu core is used at 100%, no matter what.