Folding Forum

It seems that the OpenCL 2.0 has been provisionally released (http://www.techpowerup.com/187620/khron ... ation.html). I wonder if FahCore_17 is going to use it in the near future and how it might potentially change the GPU folding landscape.

Near future? From my experience with implementation of provisional standards that is unlikely. In any case, before Core_17 could use version 2 OpenCL the drivers from AMD-ATI and nVidia would have to be updated to support it. Since nVidia is only up to supporting version 1.1, that alone could take a while. I would put it more in the Soon-ish to Not Soon range.

The core took months and months to be written with OpenCL 1.1 in mind. I doubt it'll be easy to update it, and like Joe said the drivers would have to support it. Also, there'd have to be a need to upgrade: faster speeds or a simpler codebase, something like that. Once the drivers support it, there's a demonstratable advantage in upgrading, then FahCore 17 might be upgraded, or maybe not. Second the "soonish to not-soon" category, leaning towards "not soon".

provide feedback before specification finalization, which is expected within 6 months.

Now whether AMD and NV can write drivers concurrently is another question. Otherwise we're at least a few months out, if not 6 or more.

Based on my non-programmer perspective of the information in the provisional spec, FAH will want to produce FahCore Zeta-Prime once there are is dependable support from the hardware manufacturers. I wonder if Intel's support focused mainly on the shared memory model or if they also have pre-beta drivers for their GPUs. It would be a real coup d'état in the traditional driver development pattern if dependable support for OpenCL 2.0 came out on Intel GPUs before it came out on AMD and/or NV.

If there's vendor (NVIDIA, AMD, Intel) support for it and there are features that help us speed our code, we'll start incorporating those features into OpenMM, which will make it into core17. It wouldn't be a rewrite as much as "Hey, we can make it go 20% faster if we use this new feature" and then the code evolves to use it.

PS We are also working hard to make sure there is backward compatibility with older cards, so we at times choose not to use a new feature if it breaks FAH on older GPUs. With that said eventually, things do break on older cards and FAH has to give up support of them. It's a tricky balance.