Folding Forum

Has Nvidia got any driver which is specifically tuned for running Folding@Home?

If they can tune them for games, then why can't they do them for Folding?

.

My wild guess: More people playing games compared to spending money on folding. Just a question of priorities for NV

I actually dropped a posting in the NV forum asking why the current beta version is so awful for folding on a bit older card like 660Ti. Response: zero. As per expectation.

I don't believe there's a set of drivers that maximizes folding. As to why, there's likely not enough money in that to make it worthwhile. Based on the current stats, there's around 22,000 active Nvidia cards. That's a drop in the bucket compared to the number of cards sold to gamers. Having employees write folding-specific drivers would be a non-trivial task that would keep them from the more profitable task of improving gaming performance, and they'd have to revisit those drivers whenever new GPUs or a new folding core came out.

I'd like folding-optimized drivers, but I don't see it happening barring exponential growth in the project.

You need to consider that our idea of what's important about FAH doesn't match the world's understanding -- especially those employed by hardware marketing departments. FAH has always advertised itself as a way to donate unused computer resources to science ... not as a primary reason for buying hardware. You're expecting everybody to think like you do. When you treat FAH as the primary reason that people buy NVidia hardware (or any hardware, for that matter) you're representing a very small portion of their customers.

well, that all goes to the difference between consumer and professional cards.
Both AMD, and Nvidia make professional cards that have the exact(except in many cases not crippled)GPU that goes into consumer cards.
Often the difference between these $4000 professional grade cards and the $500-$1000 consumer cards is only the optimized software(drivers)...that and that they intentionally cripple some of the GPU compute capability inherent in identical silica....

long story short...no one pays them to do it.
Even if they did, the crippled nature of consumer cards would limit huge improvements.

Of course better support of Open Cl...improved drivers that allow certain operations on existing consumer cards...would indeed speed up folding.
However, once supported in drivers...there is no need for a specific "FOLDING" version. Those updated capabilities would help any consumer GPU-compute software. ... Maybe to the point of making consumer cards attractive alternatives to the professional grade cards...and there in lies the rub.

for example, CUDA JIT COMPILER software support would help folding...but also many other programs.

I've emailed them myself, and posted them on their forums, w/o any response. They certainly like us to use their products for Folding though, and on other Distributed Computing programs.

How about AMD doing it, especially w/their new R9 290X cards?

Who ever I see putting an effort into the Distributed Computer Projects, I'm going to buy their products.

As for the time used for writing the drivers, couldn't that be written off as donating time to charity?

.

How about AMD doing it, especially w/their new R9 290X cards?

We will find out in a few days. I have researched the changes in hardware...and it should improve GPU compute capability...but I haven't seen any benchmarks yet to confirm my presumption.

Of course I knew the 7970 would crank...and it took some years to get that up to speed...so ...you can see what my opinion is worth.

As for the time used for writing the drivers, couldn't that be written off as donating time to charity?

Well, in the case of AMD, they are supporting open CL to the best of their ability (some would argue if they have the necessary resources) So, to the extent that their compilers are bug free, and OPEN MM 5.2 optimizes the capability of OPEN CL...then FOLDING should be optimized by AMD support.

I'm not saying this will work ...(burned by predictions in the past)...but AMD doesn't need to post specific drivers for FOLDING...their OPEN CL support should be beneficial to any OPEN CL program.

That is the idea anyway.

bruce wrote:You need to consider that our idea of what's important about FAH doesn't match the world's understanding -- especially those employed by hardware marketing departments. FAH has always advertised itself as a way to donate unused computer resources to science ... not as a primary reason for buying hardware. You're expecting everybody to think like you do. When you treat FAH as the primary reason that people buy NVidia hardware (or any hardware, for that matter) you're representing a very small portion of their customers.

Whenever I upgrade my hardware, I tend to do it towards Folding, rather than towards gaming, though they both over lap. I also keep my old systems and let them run 24/7 Folding, especially in the winter, since they help provide heat to my place. I have yet to turn on my real home heater, my water cooled, i7 six core, w/four (4) GTX video cards, and a 1500 watt power supply, which I built specifically for Folding. It puts out too much heat for the summer months.

.

To answer the original question.

No.
Gamers are the primary consumer. 80/20 rule applies.


The original question also assumes that tuning for games de-tunes for fah, and that fact doesn't seem to be in evidence.

7im wrote:To answer the original question.

No.
Gamers are the primary consumer. 80/20 rule applies.


The original question also assumes that tuning for games de-tunes for fah, and that fact doesn't seem to be in evidence.

How do you know what the cards are even capable of doing then if they aren't even tuned for Folding, or any Distributed Computing program? Just because Folding works on them doesn't mean that they are running very efficiently, just like w/games, which aren't running as effiently as the ones which have tuned drivers.

You're basically asking if OpenCL is optimized for computing. That's the whole purpose of the OpenCL initiative. FAH is not unique in that regard.

If you write an OpenCL program that can be shown to be inefficient, you can submit bug reports to NVidia or ATI or the Khronos Group. All the developers that write code using the OpenCL API will be appreciative of your efforts as they will be working toward the same goals.

Just because gaming works on them doesn't mean that they are running very inefficiently for fah.

I don't personally know, but Pande Group has close ties with the development teams at both NV and AMD. So they know, and they don't seem too worried about it. Plus both NV and AMD want to be competitive on all fronts, not just gaming, so they do work to optimize the drivers and hardware for GPGPU.

While it would be nice to eek out a few extra % points of performance for fah, IF that's even possible, gaming is still their primary focus. There are bigger fish to fry.

Note that the last NV driver to perform above average in fah also caused 7xx series cards to reboot every 36 hours. So optimizing is much like overclocking. Too much is never a good thing. The fixed driver ran 10% slower in fah, but at least it doesn't crash your system every other day.

indeed the TDR bug was a huge pain. i for one am happy to have stable machines. and frankly i haven't noticed much of a points drop on gtx 780's. anyway, stability counts.

If hardware vendors focused on open standards like OpenCL, it would help F@H and other projects using OpenCL too. However, they don't have the "brand uniqueness", i.e. this hardware model A has feature X which isn't present on any other model since it is propriety technology. Moreover, with Nvidia, it has to support OpenCL (open standard) and CUDA (proprietary). If they focus on CUDA, it will benefit F@H in the future when the JIT Compiler will be released. So if JIT is released by Nvidia soon, we could eventually have a CUDA FahCore that would take the maximum advantage of Nvidia GPUs and have OpenCL FahCore that would take full advantage of AMD GPUs (or any other OpenCL devices potentially). However, this might put strain on the already limited resources by having two different FahCores to support. Is it worth it, only time can tell.

As long as NV can focus their attention on CUDA, they'll sell more GPUs since they'll probably never grant a CUDA license to their competitors. They also have to keep CUDA one or two steps ahead of their support for OpenCL to make sure that everybody who has the ability to support either one switches to CUDA. As PantherX says, the features supported by NVidia are currently inferior to the OpenCL features offered by AMD -- namely a JIT compiler.

The future is unpredictable. Statements like these can be rendered obsolete at any time with very little notice.