13.4 AMD Drivers?

[Jesse_V]

And since I consider myself pretty "naive" on the core-17-topic: why does it take so much time to test the OpenMM-code? That's not meant to be disrespectful in any way, but if it pushes ATI-contributors _that_ far (like you said), I just can't find a reason why it's not top priority "#1 AAA+++".
So, if you (or any other member of this great folding experience) can explain it to me: please do so - seriously, I'm begging for information about this, so that I can understand it with no question(s) left unanswered.

That's ok, core 17 is new, so we're all learning. OpenMM itself has been around for years now, but version 5.0 was a pretty significant change. From this blog post, core 17 (based on OpenMM 5.0) includes changes like

We have deprecated several layers of GROMACs and other wrappers as the old architecture severely limited the types of simulations that can be run. Much of the work on the new core has been to replace existing features. The resulting code is now more streamlined and integrated. We also anticipate that this major re-design will allow us to introduce new features into the Folding@Home much faster.

We have introduced a new serialization mechanism that allows Pande Group researchers to setup significantly more diverse simulations. Pande Group researchers can now easily setup jobs and projects using Python (with a much richer and easier to use set of libraries), while the core maintains its speed by virtue of being written in C++.

Before we had two development branches for NVIDIA and AMD cards. It was a difficult and cumbersome task to debug and maintain. We couldn’t easily mix runs and gens produced by different GPU types. Now, using OpenCL, a single core supports not only AMD and NVIDIA, but theoretically any OpenCL-capable device.

Testing these changes took time, and there were many stability issues, many of which you can see in viewtopic.php?f=66&t=23868&start=60
Core 17 is a high priority, but it's taking time to work out all the glitches. Here's a post about OpenMM 5.1: http://folding.typepad.com/news/2013/03 ... re-17.html

But we're getting off-topic from the original question about AMD 13.4 drivers...

[bruce]

But we're getting off-topic from the original question about AMD 13.4 drivers...

... and the bottom line is that none of the drivers which are currently available work dependably with FAH. We're all at the mercy of ATI creating a complete set of drivers that are dependable.

[mdk777]

. and the bottom line is that none of the drivers which are currently available work dependably with the FAHcore that was written for the previous generation of cards (two generation past)We're all at the mercy of FAH to create a Core that uses the current generation of cards. (The current AMD drivers, sdk and programing tools available, work wonderfully on open cl in many other applications and have for some years.)

Fixed that for you.

http://developer.amd.com/tools-and-sdks ... ng/codexl/

There is a significant difference between saying that there has not been time or manpower to utilize something, and saying that thing does not exist.

AMD has supplied the hardware and the software.

FAH is indeed in the process of utilizing those tools.

[bruce]

But we're getting off-topic from the original question about AMD 13.4 drivers...

... and the bottom line is that none of the drivers which are currently available work dependably with FAH. We're all at the mercy of ATI creating a complete set of drivers that are dependable.

In the interest of fairness, the problem is not unique to ATI (though this topic is). NV has had similar problems in the past and Intel has GPUs which, if supplied with dependable drivers, could conceivably add some useful resources to FAH. It should also be noted that FAH often uses portions of the developer library and portions of the drivers which may never be used by the game developer and vice versa.

[Eagle]

Before I reply to each one of you, here are my results:

Code: Select all

Explicit Solvent: 7.85392 ns/day
Implicit Solvent: 23.2074 ns/day

Again, this was done using Catalyst 13.1 with OpenCL of Catalyst 12.8 (manually uninstalled the one from 13.1 and reinstalled the 12.8 one).
I should upgrade to 13.5 Beta 2, shouldn't I? Thought so and did accordingly by
- uninstalling OpenCL version 12.8
- installed OpenCL of 13.1
- updated the whole setup to Catalyst 13.4 (though, I omitted the "AMD Accelerated Video Transcoding"-option) and
- finally installed 13.5 Beta 2.
Now the benchmark runs again - let's see, when the results are in..

On to your posts:

Well, my opinion is that the path of least resistance was always taken in the past. (Due logically to constraints of limited resources and getting the most done quickest) Since CUDA had depth of use and support, it was simply easier to support and for many years NV cards simply worked better.

It was not top priority until very recently when open CL established an Eco-system...a critical mass of users and suppliers, and code.

Once it became obvious that there were significant economies of supporting one core...which could be used on any graphics card, GPU, CPU, or any other computing device that supported the opencl language..it has now become a priority.

Makes all sense - thanks for that!

Because this is a combination of software, and physics, and mathematics, and research...all at the same time.
There is always a lag of some years between when something is available, and when it is really utilized to its complete potential.(take sony and the ps3...games were really starting to utilize the hardware well after a year late)

The great advantage of having a standard, is that even when hardware advances rapidly, since the standard existed, it should continue to run...In theory.

Hopefully that gets "repeated" in reality, too.

My point :

When the core 17 launches with open cl 1.2 support...even if open cl 1.3, or 1.4 launches in the meantime, and AMD 8xxx series cards, or 9xxx series cards launch...they should still run the open cl 1.2 flawlessly.

Hopefully we will finally get off this constant delay and under utilization treadmill.

I hope so, too. But from a developer point of view, it shouldn't interfere with running versions, if a newer branch is being integrated - most of the time, at least.

RE:

(BTW: "benchmarking....6% done, estimate: ~7.527 ns/day" right now), I was able to make a step forward in understand the situation.

Shouldn't take that long and should get better results.

See my first post.. Seems like my setup was a bit "out of specification"..

Did you note all the requirements?

http://proteneer.com/blog/?page_id=1671

Requirements
MSVC 10 Redistributables

I think that is installed, because if not, I would expect some kind of error while launching/running the benchmark..

OpenCL:
–Windows
–Latest NVIDIA/ATI/Intel Drivers. Note that Intel OpenCL SDK Drivers must be from the 2013 Beta.
CUDA:
–CUDA 5.0 SDK
–MSVS 2010 Express
–Path to cl.exe defined in PATH environment variable

- Windows 7 x64, patched to the latest version
- latest ATI & Intel drivers (see above, regarding Catalyst)
- OpenCL wasn't included in Beta 2 of 13.5, so all I got was the one of 13.1. Where can I get the newer one?!
- MS Visual Studio Express won't be installed (I own a full copy of VS 2010), the PATH-variable wasn't modified by myself in any way

I had to install the MSVC 10 redistributables seveal times and reinstall the FAHbench subsequently to get it to work on my computer(AMD cpu ..so no experience with the INtel open cl sdk requirement).

It worked right out of the box on my end.. O.o

That's ok, core 17 is new, so we're all learning.

Thank you.

OpenMM itself has been around for years now, but version 5.0 was a pretty significant change.
(...) [Listing several (good) reasons which I could all understand] (...)
Testing these changes took time, and there were many stability issues, many of which you can see in viewtopic.php?f=66&t=23868&start=60
Core 17 is a high priority, but it's taking time to work out all the glitches. Here's a post about OpenMM 5.1: http://folding.typepad.com/news/2013/03 ... re-17.html

But we're getting off-topic from the original question about AMD 13.4 drivers...

Thanks a lot for that clarification and sorry about being a bit off-topic!

... and the bottom line is that none of the drivers which are currently available work dependably with FAH. We're all at the mercy of ATI creating a complete set of drivers that are dependable.

Well, that's a pitty - to say the least!

. and the bottom line is that none of the drivers which are currently available work dependably with the FAHcore that was written for the previous generation of cards (two generation past)We're all at the mercy of FAH to create a Core that uses the current generation of cards. (The current AMD drivers, sdk and programing tools available, work wonderfully on open cl in many other applications and have for some years.)

Fixed that for you.

http://developer.amd.com/tools-and-sdks ... ng/codexl/

There is a significant difference between saying that there has not been time or manpower to utilize something, and saying that thing does not exist.

AMD has supplied the hardware and the software.

FAH is indeed in the process of utilizing those tools.

Hmm, now I'm confused.
I don't want to play "pass the buck to X", but instead I'm looking forward to core 17. Because since the benchmark completed successfully in the first place, I suspect the old cores as really being unable to run on anything from the Radeon 5000 series and newer..

In the interest of fairness, the problem is not unique to ATI (though this topic is). NV has had similar problems in the past and Intel has GPUs which, if supplied with dependable drivers, could conceivably add some useful resources to FAH. It should also be noted that FAH often uses portions of the developer library and portions of the drivers which may never be used by the game developer and vice versa.

That seems legit to me. So to sum it all up, I
- ran the benchmark well and that equals to "core 17 will bring folding back to the GPU"
- it's a fault (or task, if you prefer that wording) on both sides, because ATI didn't deliver proper drivers in the first place and then FAH didn't deliver proper cores to run on Graphics Core Next.

Right?

BTW: FAHBench/core 17 indeed seems to benefit from the beta-driver: my score for "Explicit Solvent" currently says "benchmarking....2% done, estimate: ~5.414 ns/day" - that's nearly a 100% increase compared to what it looked like yesterday with Catalyst 13.1 and OpenCL of 12.8..

[mdk777]

Well, your results are still not good on the benchmark

You should be getting something like this on an AMD 7970:

HD7970: (Explicit 18.1 ns/day | Implicit 101.3 ns/day)

I am not an expert on the software. I just use the brute force method...keep installing the required bits until it works.

Here is a thread on debugging the install:

viewtopic.php?f=16&t=23434#p233613

[Eagle]

That's too much for me as I only wanted to know if core 17 runs well (and it did as explained above).
More or less power is not of any interest for me when it comes to the actual benchmarking..

[Eagle]

Just for the record - running with 13.4, I got this:

Code: Select all

Explicit Solvent: 5.60736 ns/day
Implicit Solvent: 17.2936 ns/day

So, the latest stable driver lessens the the scores (7.85392 and 23.2074 ns/day).

Now onto the beta-driver..

[mdk777]

Still off by a factor of 4-5X regardless of the driver used for an AMD 7970.

Are you running the test with the CPU still folding?

Something just not right...

I get variations in the order of 17.8 to 18.3 and 110 to 118 on the test depending on the driver used.

5 and 17 are not even in the ballpark.

Consequently you do have to have something else going on here....testing the IGP on the cpu instead of the AMD7970?
testing on CUDA instead of OPEN CL?
I don't know, but I need to point out that further testing is not worth your time until you get it worked out.

ps please note: from the CAT BETA page:

NOTE! You must un-install AMD Catalyst 13.4 (if you have already installed AMD Catalyst 13.4) before installing AMD Catalyst 13.5 Beta

[Eagle]

Still off by a factor of 4-5X regardless of the driver used for an AMD 7970.

Are you running the test with the CPU still folding?

Yes.. CPU was running on 7 threads..

Something just not right...

I get variations in the order of 17.8 to 18.3 and 110 to 118 on the test depending on the driver used.

5 and 17 are not even in the ballpark.

Now fixed, because I simply paused it for the benchmark run.

Consequently you do have to have something else going on here....testing the IGP on the cpu instead of the AMD7970?
testing on CUDA instead of OPEN CL?
I don't know, but I need to point out that further testing is not worth your time until you get it worked out.

Come on, buddy. You didn't think I mistake CUDA with OpenCL or IGP instead of Tahiti, didn't you?

ps please note: from the CAT BETA page:

NOTE! You must un-install AMD Catalyst 13.4 (if you have already installed AMD Catalyst 13.4) before installing AMD Catalyst 13.5 Beta

DDamn it! I knew I had forgotten something..
Anyway, after I downloaded and ran the "AMD Catalyst Un-install Utility", I faced the next problem: 13.5 Beta 2 didn't recognize my GPU, so all it offered for installing was the Install-Manager and the Catalyst-Control-Center itself. I fixed that by renaming "W8_INF" and "W764A_INF" to "W7_INF" and "W764A_INF" (I've added "4" between "6" and "A", because all other subfolders use that naming scheme..) at first and since Setup.exe still didn't recognize my GPU, I tried "InstallManagerApp.exe" from the "Bin64"-folder.

This way, I finally got all drivers and tools listed, chose all but the "SteadyVideo"-package (which, according to Google results is only for "smoother" browser scrolling which I don't care about) and there we go:

Code: Select all

Explicit Solvent: benchmarking....4% done, estimate: ~7.013 ns/day

And just to confirm, these are my Catalyst versions:

Code: Select all

Driver Packaging Version        12.102.3-130416a-155949E-ATI
Catalyst Version                13.5
Provider                        Advanced Micro Devices, Inc.
2D Driver Version               8.01.01.1295
2D Driver File Path             /REGISTRY/MACHINE/SYSTEM/ControlSet001/Control/CLASS/{4D36E968-E325-11CE-BFC1-08002BE10318}/0000
Direct3D Version                9.14.10.0968
OpenGL Version                  6.14.10.12198
Catalyst Control Center Version 2013.0416.1036.17145
AMD Audio Driver Version        7.12.0.7714

[Eagle]

So, here are my results with the whole machine being unused, except for FAHBench:

Code: Select all

Explicit Solvent: 6.95051 ns/day
Implicit Solvent: 16.6337 ns/day

What's going on?! :-/

[teaough]

OK - I've given up. I've returned to the 12.8 driver and the 2.7 SDK. Now I'm back to getting 97-99% utilization. I'll wait and see about changing when the OpenMM 5.1 code comes out and the Core17 beta is finished and let out "into the wild." Sorry folks.

What took me over the edge was the frequent occurrence of UNKNOWN_ENUM (-1073741819 = 0xc0000005) - about every 3% of progress in Project 11293. Core shuts down and no work gets done. Enough of that. Things appear more stable now. Time will tell though.

[bruce]

As you probably know, error 0xc0000005 is a memory violation. Often that's because of errors in the memory subsystem such as overheating or overclocking but often it's essentially impossible to figure out what caused it.

[teaough]

@Bruce - nope, didn't know about the mem violation. Thanks but doesn't matter. Things ran all night w/o incident at 90+% (although I didn't monitor it ALL night . I think I'll stay with the 12.8/2.7 combination. I don't do games, just the occasional OpenCL that in my advanced years I'm aspiring to pick up.

btw - my GPU is a 7770 with stock clocks and open case. However, maybe I should check for dust in the fan. Thanks.

[foldy]

Tested 13.4 driver vs. 13.1 with FAHbench on my HD 7950 @ 1000Mhz

13.1
Explicit 29.7
Implicit 87.8

13.4
Explicit 30.1
Implicit error

So no real difference between both versions on Explicit.
And Implicit had an error due to FAHbench bug with 13.4

Also real Fahcore_17 showed no difference in performance and GPU usage is 97%.
I stay with 13.4. now