Page 1 of 1

Retirement of Core15

Posted: Mon Oct 10, 2016 12:27 pm
by DocJonz
Vijay posted in his blog over the weekend that Core15 is to be retired, and;
"that means that some older GPUs (that are not supported by OpenMM) will no longer be supported by FAH".

So that we can inform our respective teams, do we know what cards that means in practice? (Is there a list of non-OpenMM GPU's?)

Re: Retirement of Core15

Posted: Mon Oct 10, 2016 2:16 pm
by 7im
Tesla class GPUs are now blacklisted as there is no more work that will run on those. (G8x, G9x, GT2xx)

Fermi and above are supported. (4xx series and up)

Also note that EOL was announced years ago, and that these work units ran out months ago. This is just a formality.

Re: Retirement of Core15

Posted: Mon Oct 10, 2016 2:58 pm
by JimboPalmer
Some Tesla GPUs could only do Core 11.
Some Tesla GPUs could also do Core 15, they are now done. Nothing before GT4xx can fold. (including G 405)
Some slow Fermi GPUs could run Core 15 fast but Core 17 slowly, they may be unable to complete WUs in the time allowed, even though they fold.

Most GT 4xx card support OpenCL 1.1, a small number support 1.2

Re: Retirement of Core15

Posted: Mon Oct 10, 2016 9:48 pm
by bruce
Perhaps the real criteria is OpenCL 1.2 or greater -- but there were also some support for a FAHCore with certain versions of CUDA-only. It might also be Compute Level x.x or greater. Maybe we need to get an OpenMM expert to define it.

Defining it by model number has never been particularly accurate since the GPU manufacturers have been (intentionally) inaccurate at requiring that a model number-family all have equivalent capabilities. Unfortunately their sales staff an most customers typically think in terms of model numbers, not actual API support + hardwre capabilities.

Fortunately we figured out that even the code-name Maxwell consisted of two subcategories without having to retroactively classify different capabilities.

Re: Retirement of Core15

Posted: Tue Oct 11, 2016 5:59 pm
by 7im
Minimum folding specs for GPUs, in the current GPU cores...
OpenCL v1.1 or above, AND Compute Capability 1.2 or above.

GeForce 8, 9, and 2xx series are compute level 1.1. Half are OpenCL 1.0, and 1.1.

These minimum requirements may change with the new FAHCore based on OpenMM 7.0 mentioned in the blog post.

Re: Retirement of Core15

Posted: Tue Oct 11, 2016 7:29 pm
by JimboPalmer
According to this https://developer.nvidia.com/cuda-gpus
All the GT4xx are Compute Capability 2.1 except the GT 480M, which is 2.0. So we are not near the edge.

If the new OpenMM needs Compute Capability 3.0 or higher then more cards will be retired.

Image

Re: Retirement of Core15

Posted: Tue Oct 11, 2016 9:27 pm
by Joe_H
GT405 is a re-badged GT310, so is definitely not included. The lower end GT4xx cards may meet the OpenCL and Compute Capability minimums, but have too few shaders to process WU's currently being assigned within the deadlines. This also applies to some of the lower end cards in the higher numbered series from nVidia as well.

Re: Retirement of Core15

Posted: Wed Oct 12, 2016 1:58 am
by bruce
JimboPalmer wrote:If the new OpenMM needs Compute Capability 3.0 or higher then more cards will be retired.
Even if it is required, that doesn't mean any cards will be retired any time soon. There will still be active projects using Compute Capability 1.x or 2.x.

As long as the client/server is able to identify which capabilities you have, new projects can be restricted to hardware with the new capabilities. Older projects which currently run with lower capabilities will still be in production and the servers will be given the responsibility of limiting assignments to match what your hardware can do. (At the time that Core15 was new, it didn't immediately force the retirement of all older GPUs. There were still projects that could run on Core11 and they were still productive. Similarly, Cores 17/18/21 have been out quite some time before the last Core15 projects were completed.)

Eventually, more the GPUs with more limited capabilities will finish whatever projects they are capable of completing and only then will the older cores be retired.

You'll notice that the V4.4.15 client (and beyond) now reports Compute Capability of your GPUs.

Re: Retirement of Core15

Posted: Wed Oct 12, 2016 2:08 am
by JimboPalmer
Thanks for explaining what I should have. Even if new Cores need new capabilities, only new projects will be affected. Existing Cores and Projects will still exist until all their WUs are solved.