Should F@H abandon CPU folding?

[ifolder]

If I understand well this: https://stats.foldingathome.org/os and conservatively assume that in average a CPU requires 30W of power and a GPU 250W, it looks like CPU folding consumes 25% of total CPU+GPU electricity to provide 1% of total x86 TFlops. This is quite a huge waste of energy!

So shouldn't F@H start thinking about abandonning CPU folding and rather than the "I am one in a million" campaign that aims to increase the number of folders in general, try to target gamers who own GPUs (often hi-end GPUs moreover)?

[bruce]

No.

Projects which run efficiently with 1 to, say, 16 parallel CPU threads will run extremely inefficiently on GPUs with thousands of shaders. In fact, projects which run efficiently on GPUs with 400 shaders run every inefficiently on a GPU with 4000 shaders (and vice-versa).

[foldy]

CPU folding may be an entry level for everyone while GPU folding is for more advanced users.

[ifolder]

No.

Projects which run efficiently with 1 to, say, 16 parallel CPU threads will run extremely inefficiently on GPUs with thousands of shaders. In fact, projects which run efficiently on GPUs with 400 shaders run every inefficiently on a GPU with 4000 shaders (and vice-versa).

Does this mean that today the severs assign GPU WUs depending on the number of shaders of the client to maximize efficiency?

Also when you say that a project which run efficiently with 1 to, say, 16 parallel CPU threads will run extremely inefficiently on GPUs with thousands of shaders does it mean that to complete that project on a GPU it will use more energy (Wh) than completing it on a CPU with 1 to 16 threads?

[ifolder]

CPU folding may be an entry level for everyone while GPU folding is for more advanced users.

Sure.
Especially if it is proven that CPU projects work better (ie take less energy to complete) on CPU than on GPU.

But if not, 80000 CPUs * 30 W * 24h = 57.6 MWh per day spent to deliver 1% of the total F@H computing power is quite a lot of energy spent just to offer an entry point, when you consider that the other 99% are delivered with about 170 MWh per day, isn't it?

[ifolder]

Has F@H ever tried to approach gaming platforms to have them displaying ad banners like "When you're not gaming, use your computer to cure Cancer"?

[JimboPalmer]

Has F@H ever tried to approach gaming platforms to have them displaying ad banners like "When you're not gaming, use your computer to cure Cancer"?

If they had money for advertising, they would spend it on Science.

[bruce]

No.

Projects which run efficiently with 1 to, say, 16 parallel CPU threads will run extremely inefficiently on GPUs with thousands of shaders. In fact, projects which run efficiently on GPUs with 400 shaders run every inefficiently on a GPU with 4000 shaders (and vice-versa).

Does this mean that today the severs assign GPU WUs depending on the number of shaders of the client to maximize efficiency?

Also when you say that a project which run efficiently with 1 to, say, 16 parallel CPU threads will run extremely inefficiently on GPUs with thousands of shaders does it mean that to complete that project on a GPU it will use more energy (Wh) than completing it on a CPU with 1 to 16 threads?

The assignment process is complex and involves many pieces of information. The GPU species and subspecies (as well as the CPU thread count) are used to include/exclude certain projects. Watts drawn does not figure directly in those calculations. Donors have a self-interest in adjusting their individual PPD/W values.

[toTOW]

To answer the question in this thread topic, no. CPUs can fold some things that GPUs can't. So CPU and GPU folding are still complementary.

[Jonazz]

To answer the question in this thread topic, no. CPUs can fold some things that GPUs can't. So CPU and GPU folding are still complementary.

This, basically. FAH has made great progress with GPU projects. But it might take a long time before a GPU can fold the same things as a CPU, if ever.

[bruce]

There are a lot of computers that don't support GPUs and they do make useful contributions to science. ... e.g. many laptops, most computers running Mac OX-X, etc.

I can't think of any good reason to abandon any donor who doesn't live up to some higher standard on a whim.

FAH has announced that they're abandoning the nacl client, but that's because Google is dropping support and the FAH support costs continue to increase. The CPU folding platoform does produce a LOT more science and the FAH support costs are minimal.

Then, too, just count the number of new projects that have been released recently for FAHCore_a7 which would have to be migrated to projects and because of their smaller atom counts, they would end up runing inefficiently on high-end GPUs.

[MeeLee]

The current point system, is pushing people to fold on GPU, as it gets a lot more points (10-1000x more) per day than CPU folding.
If CPU folding is very important to science, why not increase the points for CPU folding?

I hear Intel's new 10th generation, 10nm CPUs should perform admirably for folding! (they're said to perform as fast as a 5+Ghz CPU, at only 3,7Ghz), and do this at ridiculously low wattages (9, 15 and 25W TDP, which is about 2.5-5x lower than most CPUs today)!

[bruce]

The points are calibrated to represent the value of the scientific work done, which depends both on the size of the protein and the processing time. Both CPUs and GPUs do, in fact produce useful science but speed depends greatly on the number of floating point operations that can be done concurrently. Newer CPUs with more cores and support for higher level instructions (e.g. AVX) that increase internal parallelism do enable the hardware to process WUs faster, but the CPU parallelism doesn't approach that of modern GPUs.

Then, too, higher levels of parallelism increase the need for more watts, which is generally then reduced by better chip manufacturing technologies.

Bottom line: Look up the single precision GFLOPS numbers of the various technologies. It's not a perfect measure of FAH performance, but as a single number, it's a pretty good method to compare results.

[timberwolf]

I'm very sorry for resurrecting an old thread.

As someone who has previously folded with the CPU alone, it's truly disheartening to see posts like this and, if I had viewed this post now, it would certainly make me question remaining an active folder.
As previously stated, there are things a CPU, at least right now, is best delegated to do - this is likely why other distributed computer projects like WCG and Rosetta are still using the CPU.
It certainly is a nice incentive to see 700k plus points a day. I wouldn't say it's worth $1000, though, and I still feel like a tiny fish in a gigantic ocean, but crunch on .
Again, I'm sorry for bringing back an old thread. But to the CPU crunchers out there - keep crunching .

[JimboPalmer]

Understand that this was NOT someone at Folding@Home considering dropping CPU folding.

Some user using their own data was trying to interpolate for the rest of us.

You need not be dismayed that one person has a question. Especially as the informed answers supported your choice anyway.

Core_a4 supports all x86 CPUs back to Willamette and works reliably.

Core_a7 seems stable, and can support CPUs almost as old, but prefers CPUs starting at Haswell. A7 has issues with older OSs, more than older CPUs.

Core_21 is doing well on Nvidia from Keplar, and AMD from GCNv2, there is some issue with the new DRNA chip, but I think that is 'just' a driver update.

Folding@Home is working on a new Core_22, but I doubt it will abandon any GPUs. I think it is just better science.