GTS450 too old for folding?

[Luno]

Hi,

I've been using some old hardware to do some folding. The GTS450 and a Dual-Core Celeron weren't exactly churning out a lot of points but why not use the older and still working hardware for something useful instead of throwing it in the trash .

However, for some time now the GPU has stopped folding and I'm getting this message on startup:

Code: Select all

11:22:02:Adding folding slot 01: READY gpu:0:GF106 [GeForce GTS 450]
11:22:02:Removing old file 'configs/config-20190613-155153.xml'
11:22:02:Saving configuration to /etc/fahclient/config.xml
11:22:02:<config>
11:22:02:  <!-- Client Control -->
11:22:02:  <fold-anon v='true'/>
11:22:02:
11:22:02:  <!-- Folding Slot Configuration -->
11:22:02:  <cause v='COVID_19'/>
11:22:02:  <gpu v='false'/>
11:22:02:
11:22:02:  <!-- Network -->
11:22:02:  <proxy v=':8080'/>
11:22:02:
11:22:02:  <!-- Slot Control -->
11:22:02:  <power v='full'/>
11:22:02:
11:22:02:  <!-- User Information -->
11:22:02:  <team v='233109'/>
11:22:02:  <user v='BIT1'/>
11:22:02:
11:22:02:  <!-- Folding Slots -->
11:22:02:  <slot id='0' type='CPU'/>
11:22:02:  <slot id='1' type='GPU'/>
11:22:02:</config>
11:22:03:WU00:FS01:Connecting to assign1.foldingathome.org:80
11:22:03:WU00:FS01:Assigned to work server 192.0.2.1
11:22:03:WU00:FS01:Requesting new work unit for slot 01: READY gpu:0:GF106 [GeForce GTS 450] from 192.0.2.1
11:22:03:WU00:FS01:Connecting to 192.0.2.1:8080
11:22:03:WARNING:WU00:FS01:WorkServer connection failed on port 8080 trying 80
11:22:03:WU00:FS01:Connecting to 192.0.2.1:80
11:22:03:ERROR:WU00:FS01:Exception: Failed to connect to 192.0.2.1:80: Network is unreachable
11:22:03:WU00:FS01:Connecting to assign1.foldingathome.org:80
11:22:04:WU00:FS01:Assigned to work server 192.0.2.1
11:22:04:WU00:FS01:Requesting new work unit for slot 01: READY gpu:0:GF106 [GeForce GTS 450] from 192.0.2.1
[... this repeats forever]

That sounds like network issues, but the cpu is receiving its jobs just fine - Some googling revealed that this might be a very misleading error message and that it actually means that the GPU is no longer supported. I wouldn't be surprised if that was the case with a 10-year-old GPU, but according to the Whitelist it should still work?

Everything was working fine until a few weeks ago (no exact date as the machine mostly runs without a monitor and I only check every few months or so), I didn't change software, hardware or even rebooted in between. After discovering the error, I installed all current packages for my Linux distro (Linux Mint 18.3), removed and reinstated the gpu folding slot, checked nvidia drivers (already the newest available version, some 390 LTS for older cards), updated FAH* to the newest version, rebooted the machine, played with a lot of settings (project preference etc.) - all to no avail.

tl;dr: GTS 450 still supported? If yes, what am I doing wrong?

[JimboPalmer]

Welcome to Folding@Home!

The GTS 450 is too old to fold. To minimally be able to fold a GPU needs to support OpenCL 1.2 and be able to do Double Precision floating point math. (FP64)

https://www.techpowerup.com/gpu-specs/g ... -450.c1778

The GTS 450 does not support OpenCL 1.2

If it could fold, it would have difficulty meeting deadlines, it would be twice as slow as a GTX1030, the slowest GPU Nividia still makes.

[paulmd199]

Gotta say this is a huge UI problem. Connecting to a private network because your card isn't supported is strange behavior. Client should be explicit in saying that the card isn't supported.

[JimboPalmer]

I am not disagreeing, I just read the clues and give my best help. (and guesses) The better the clues, the better my help.

Above I give the minimum requirements, but I have no documents that say that. (I blame the fact that the project is run by Biochemists, not Documentalists)

[Neil-B]

Not sure if there is a GitHub Issue for this https://github.com/FoldingAtHome/fah-issues/issues but if there isn't and you wish to raise one it will get this onto the "investigate/to do" list.

Someone has actually just raised a similar query to you own as an issue in that they felt there was an issue with the FAHClient as their GPU was in the GPUs.txt list and logged this as a "Failed to connect to 192.0.2.1" technical issue … as I commented there GPUs.txt is both a White and a Black list and GPUs that are listed as not being capable of folding as sent to this IP Address - I can't remember precisely why but fairly sure I have seen a reasonably detailed one somewhere in the forums.

What I have done is edit my comment on that issue to reflect the view that it would be more helpful if the FAHClient should explicitly state that the GPU is not capable of folding.

[PantherX]

There is an issue already created: https://github.com/FoldingAtHome/fah-issues/issues/1309

Regarding the IP address, it belongs to a test network so highly unlikely that you would have it in normal home environments: https://tools.ietf.org/html/rfc5737

[bruce]

The GTS 450 is still supported. Please be patient and you'll probably get an assignment soon.

We've been given new information. It turns out that OpenCL 1.2 is not, in fact, a requirement. OpenCL 1.1 apparently works, too. DoublePrecision (FP64) is still a firm requirement.

A little background information: The OpenMM code contained in the FAHCore was thoroughly tested with FP32 and OpenCL 1.1. Later it was tested with OpenCl 1.2 and FP64 but by the time FP64 became a firm requirement, the lab no longer had any of theold GPUs that supported OpenCL 1.1 so they never were able to test OpenCL 1.1 with FP64. Based on the lack of error reports from the testing that you've (unknowingly) helped with, we have decided to re-enable hardware like yours that only supports OpenCL 1.1

[Luno]

The GTS 450 is still supported. Please be patient and you'll probably get an assignment soon.

We've been given new information. It turns out that OpenCL 1.2 is not, in fact, a requirement. OpenCL 1.1 apparently works, too. DoublePrecision (FP64) is still a firm requirement.

A little background information: The OpenMM code contained in the FAHCore was thoroughly tested with FP32 and OpenCL 1.1. Later it was tested with OpenCl 1.2 and FP64 but by the time FP64 became a firm requirement, the lab no longer had any of theold GPUs that supported OpenCL 1.1 so they never were able to test OpenCL 1.1 with FP64. Based on the lack of error reports from the testing that you've (unknowingly) helped with, we have decided to re-enable hardware like yours that only supports OpenCL 1.1

Yay!

I'll check in on the assignments as soon as I get back to the office - which might not actually be that soon because of you know who / what

I've also checked some stats ( https://www.techpowerup.com/gpu-specs/m ... e-oc.b1072 ) and was a little confused by comparing the GTS 450 to the mentioned GT 1030 - while the newer card is obviously way faster in everything that matters for Gaming and has much newer and more features, the raw FP64 values are putting the GTS450 at 54.40 GFLOPS while the GT 1030 sits at 35.23 GFLOPS? Granted, it's described as "theoretical performance", still seems a bit weird?

[Joe_H]

The computing being done is "mixed" precision, mostly Single Precision (FP32) with FP64 used where needed to maintain sufficient accuracy. So you also need to compare the FP32 ratings which will dominate the mix. There are several models of GT 1030, the FP32 ratings are 1000 GFLOPS plus or minus. A GTS 450 is about 600 GFLOPS.

[MeeLee]

If it could fold, it would have difficulty meeting deadlines, it would be twice as slow as a GTX1030, the slowest GPU Nividia still makes.

The GT 730, and 710 is still around (surprisingly).

[Luno]

The GTS 450 is still supported. Please be patient and you'll probably get an assignment soon.

We've been given new information. It turns out that OpenCL 1.2 is not, in fact, a requirement. OpenCL 1.1 apparently works, too. DoublePrecision (FP64) is still a firm requirement.

A little background information: The OpenMM code contained in the FAHCore was thoroughly tested with FP32 and OpenCL 1.1. Later it was tested with OpenCl 1.2 and FP64 but by the time FP64 became a firm requirement, the lab no longer had any of theold GPUs that supported OpenCL 1.1 so they never were able to test OpenCL 1.1 with FP64. Based on the lack of error reports from the testing that you've (unknowingly) helped with, we have decided to re-enable hardware like yours that only supports OpenCL 1.1

Yay!

I'll check in on the assignments as soon as I get back to the office - which might not actually be that soon because of you know who / what

As of today, I'm still getting the same error message

[JimboPalmer]

192.0.2.1 is a sign that the GPUs.txt you have lists the GTS 450 as unsupported. You can try deleting your copy on your PC and seeing if it does not download a new copy.

[bruce]

I've talked to the head of OpenMM. At this point, there is little hope that they'll attempt to support OpenCL 1.1 in a FAHCore. The API functions that are missing are essential. while it is possible that they can be emulated with software macros, that means three things. First, the emulation produces a hit to production, running slower than the actual ocl 1.2 code would. In as much as the GPUs that don't work with OCL 1.2 are relatively old, they're already pretty slow and the added emulation code makes them just that much slower.

Second, that means they'd have to build a separate version of the FAHCore to incorporated that added slowness and then manage assigning that degraded FAHCore to only those GPUs. And, of course then future changes would then have to be compiled for both versions. At this point, the OCL 1.2 core works for both AMD and for NVidia and new features are being incorporated into that multi-purpose version specifically to do the analysis required by COVID19 so it's absolutely not a good time to deal with the added burden of making the changes twice. (Read about COVID Moonshot on FAH's website.)

Third, there really are not a lot of these OCL 1.1 limited GPUs. It's only half of the FERMI generation of GPUs. The FERMI GPUs actually represent two different generations -that I'll call FERMI-0 and Fermi-1 and unfortunately you've got a FERMI-0 that's built with a GF106 chip. What's critical is the second digit: GF10x vs. GF11x. You need a chip from the second half of the Fermi generation (or later).

[Luno]

Naw

Well, thanks for digging into the matter and providing a detailed answer.

It looks like I'll be retiring a GTX 660 soon sooo that could be my path back to GPU folding .

[Luno]

Soooo this happened 3 days ago:

Code: Select all

0:18:23:WU00:FS01:Connecting to assign1.foldingathome.org:80
20:18:23:WU00:FS01:Assigned to work server 128.252.203.10
20:18:34:WU00:FS01:Requesting new work unit for slot 01: READY gpu:0:GF106 [GeForce GTS 450] from 128.252.203.10
20:18:34:WU00:FS01:Connecting to 128.252.203.10:8080
20:18:39:WU00:FS01:Downloading 15.43MiB
20:18:45:WU00:FS01:Download 35.25%
20:18:51:WU00:FS01:Download 84.27%
20:18:52:WU00:FS01:Download complete
20:18:52:WU00:FS01:Received Unit: id:00 state:DOWNLOAD error:NO_ERROR project:11759 run:0 clone:6626 gen:52 core:0x22 unit:0x0000005a80fccb0a5e6e935a9e2a5f26
20:18:53:WU00:FS01:Starting

And it's back up to >10000 points per day as compared to the previous CPU only < 2000. Yay ?