CPU vs GPU work units

[mwroggenbuck]

Do CPU work units perform computations that GPU cannot do, or do CPU work units just perform a slower implementation of GPU work units? Is there any value in CPU work units that cannot be obtained faster with GPU work units?

[JimboPalmer]

I know that in the past, GPUs were limited to implicit solvation while CPUs could do explicit solvation, I am told this is no longer the case. (either from Core_17 to Core_21, or from Core_21 to Core_22)
https://en.wikipedia.org/wiki/Implicit_solvation
https://en.wikipedia.org/wiki/Water_model
https://en.wikipedia.org/wiki/List_of_F ... home_cores

(this is backwards of how most people think of computers, but F@H spends most of its time doing SIMD)
In any case, the SIMD engine in a CPU has a fully functional CPU to do the logical work, while the much larger equivalent of the SIMD engine in a GPU has a weak logic unit designed primarily for graphics.
https://en.wikipedia.org/wiki/SIMD

There are three compelling reasons to continue CPU folding

1) in some scenarios, you want that better logic of a CPU.

2) some proteins are 'too small' for modern GPUs, the volunteers would complain if their GPU was only partially used since their Points Per Day would be reduced. Meanwhile there are now 256 thread CPUs from AMD that approach GPU speed.
viewtopic.php?f=38&t=35286

3) there are a great many PCs with only CPUs, or with GPUs too simple to Fold. CPU folding works back to 2001 hardware, while GPU Folding from 2013 is really iffy to meet minimum specs.

[mwroggenbuck]

I appreciate the reply, but it really does not answer my question. Let me try to ask in another way. In the context of Folding@Home, do CPU work units do something that is actually different than GPU work units? Are the CPU work units providing information that cannot be obtained from GPU work units? I realize it may be more appropriate for small work units to be assigned to CPUs, but are you getting different information? Again, answer this in the context of FAH.

You correctly inferred my reason for asking this question, which is does it make sense to do CPU work for FAH if you have a GPU already doing work. If you do get different information, then it does make sense. If not, then I think it becomes a personal decision if you want to spend electricity on the CPU. Of course, if all you have is CPU, then it makes sense to do CPU. Every little bit helps.

Just curious.

[HaloJones]

If you're happy to use your cpu for FAH, do so. If you're not, don't.

CPU units are different and absolutely have value. Whether that value is sufficient for you to dedicate cpu time to the project is something only you can answer. Modern high-core count cpus do very significant work.

[JimboPalmer]

F@H is ALWAYS a personal decision. You are volunteering your money to pay electric bills and the wear and tear of operating your equipment.

F@H obviously finds value in CPU WUs or it would not issue them. (I fold on 3 GPUs and 10 CPUs, the GPUs make 1.2 million PPD total and the CPUs only add 50,000 PPD)

I can't tell you what you value and what you don't.

I know that until recently, CPUs did operations that GPUs could not. I suspect that now CPUs are just more efficient at some Operations than a GPU. From a PR point of view, both CPU folders and GPU folders will be unhappy if CPU folding dies. (The researchers are currently working on a Core_a8 for CPUs, so the researchers think CPUs are valuable. Rumor has it that 50% speedups are possible on CPUs newer than 2013 using AVX2 and FMA
https://en.wikipedia.org/wiki/Advanced_ ... tensions_2

[Neil-B]

@JimboPalmer beat me to it but since I've type it ...

Everything is a personal decision, irrespective of any differences to be honest.

I'll add that from my understanding the GPUs are serviced at the moment by cores based on OpenMM and the CPU cores uses Gromacs ... over the years theses have developed towards each other and are maybe more similar now than they used to be ... some researchers prefer one over the other - others less so but it mean there is demand for both from the researchers point of view (again, repeat, as I understand it - I may be wrong).

Another factor CPU folding cores exist is simply is that there are some 2.1M CPUs folding and some 0.4M GPUs ... Whilst GPUs folding cores may progress science quicker there are 5x more CPUs being donated.

Hopefully some researchers will confirm with the precision you are looking for the precise (if any) differences in capability of the CPU cores and the GPU cores but for many people their personal choice more comes down to what they have available and what they have chosen to donate.

[mwroggenbuck]

I may be posting this twice, but I think my last post was dropped for some reason.

To my knowledge (which may be wrong), a CPU can do anything a GPU can do, although it might be at a much slower pace. I don't believe the reverse is true. I think a CPU can do more complex logic (like multiple branches of if statements) that are simply not possible (or at least efficient) to do on a GPU.

I guess my question boils down to this: Does FAH get different information from CPU work units than from GPU work units? Or are CPU just generating the same information? I could think of a CPU doing something like "pre-screening" for GPU processing, but I don't know. Which is why I am asking.

[JimboPalmer]

(while I was a programmer, I have never worked in molecular dynamics, I wrote mostly communication code until the internet and inventory/payroll code after. Lots of networking, lots of database work)
https://www.youtube.com/watch?v=4-DN8-USf74

When I worked at a open pit mine, we used huge 320 to 380 ton trucks when moving waste and much smaller 30 to 180 ton trucks when moving ore. I suspect F@H does something similar, those 512 to 4096 core GPUs are discovering interesting sites on proteins to attack, and the 2 to 128 core CPUs are testing drugs to see how well they attack the discovered sites. (this is just my understanding, again I have not seen this code, I certainly have not written this code)

Our issue with helping, is that F@H is designed as a hobby for your PC, not for you. Since you don't need to know what each WU is doing, they do not put the explanation in layman's terms.

(I fire up each new PC, tune the OS and the hardware and forget it: turn off sleep when plugged in, use more than 4 gig of dual channel RAM for CPUs, etc)

[mwroggenbuck]

Ok, I am not saying CPU is more or less valuable than GPU. Every little bit helps. However, my undergraduate degree was a double major in computational mathematics and chemistry so I have an interest in both. I am just curious if different information is being extracted. That is all. I am not making a judgement on anything. I just want to acquire information.

My doctorate was in physical chemistry. I used computers to assign NMR resonance frequencies from combined multiple dimension NMR spectra of RNA to help define structure. I know for certain that algorithm I used there would not port to a GPU. Like I said multiple times, I just am interesting in knowing if FAH uses these to gather different information.

I am sorry if I offended anyone. I just always ask questions. It is part of my nature.

[JimboPalmer]

However, my undergraduate degree was a double major in computational mathematics and chemistry so I have an interest in both. I just want to acquire information.

OK the good news is that the science code is open source!

CPUs http://www.gromacs.org/

GPUs http://openmm.org/

I have not read any of this code as my experience is largely inventory/accounting/database design. You have the right back ground to understand the source code.

What I am not sure you will learn, and I think you want, is when to the researchers choose one over the other. As was mentioned above, some is what have they used in the past. "We have always done it that way" lives in many fields. Some of what is implied by Moonshot, is we will do this quickly, not most structured.

I wish you luck!

[mwroggenbuck]

You made me think of something I had not. It is possible that they are using two different algorithms to obtain the same information (protein folding trajectories). One algorithm may do better in one situation, and the other may do better in a different. The choice is up to the researcher. It is possible that both algorithms could be implemented in GPU, but gromacs has not gone that way (either because it works now and we need now, or it does not transfer easily)

In any event, every CPU/GPU cycle is trying to help solve this terrible virus, and really that is all that matters. Thanks to all who have weighed in on this.

[PantherX]

...Does FAH get different information from CPU work units than from GPU work units? Or are CPU just generating the same information? I could think of a CPU doing something like "pre-screening" for GPU processing, but I don't know...

If you're talking about the output information from CPU/GPU WUs,then it would be the same (atom/molecule positions, temperatures, velocity, etc.). However, the information from the CPU can't directly be fed to the GPU for processing. If Project A starts using CPU, then it must continue using that (including any follow-up) to ensure consistency between the generated datasets.

AFAIK, I haven't read about the CPU Projects which "screens" data and then use the GPU to process the "chosen ones" but that could be an approach. Whether or not it is used by researchers, is something that I don't know.

[MeeLee]

GPU amplified CPU processing...
Probably the next phase.
The CPU can do more complex mathematical calculations than the GPU,
In theory the CPU can do things, a GPU can not.
However, in the context of FAH, I don't know if CPU WUs are actually doing calculations that couldn't be done on GPUs.

Modern GPUs now are equipped with more Double floating point precision circuits than GPUs from before, in that such calculations in theory, should no longer form a bottleneck on GPUs.

GPUs work best in looping calculations. Input gets processed, written to output, and that output is looped back into the input.
Where GPUs don't work well (yet), is where new data needs to be merged or replaced with the output data, or where constant new data needs to be presented to the input.

Nvidia is working on incorporating ARM coprocessors inside their GPUs, to alleviate some of that '1 core per GPU' people are talking about.
They do this especially for deep learning, AI, and perhaps Folding can make use of this too in the near future.


As you may already know,
GPUs work a lot faster, because they have more cores (usually over 40) and thousands of shaders doing the work at ~2Ghz, vs CPUs which usually are limited to 8 cores / 16 threads, at 4Ghz.

[bruce]

I'll add that from my understanding the GPUs are serviced at the moment by cores based on OpenMM and the CPU cores uses Gromacs ... over the years theses have developed towards each other and are maybe more similar now than they used to be ... some researchers prefer one over the other - others less so but it mean there is demand for both from the researchers point of view (again, repeat, as I understand it - I may be wrong).

there are several analysis code bases that are used by scientists in stand-alone environments on whatever hardware is being used for their personal research. FAH has adapted two of them for cloud-sourced computing by donors like yourself.

Sometimes they develop toward each other; sometimes not. The fact is that (A) they are developed by different teams and (B) each one gets a new algorithm added from time to time. If a scientist needs to do a different type of analysis, (s)he can propose an enhancement to the internal methodology of his/her favorite analysis package. A couple of years later, that same method will likely be added to the other code.

The COVID Moonshot needed some new methods and FAH has recently been adding special algorithms to OpenMM.

[mwroggenbuck]

My original question has been answered by PantherX and Bruce and JimboPalmer (and everyone else). The outputs are the same, but the algorithms are different. Given my background, I was just curious. I was not trying to say that one method (or its hardware implementation) was better than the other--I just wanted to know what was going on. Thanks to a lot of people, I now have a much better idea of the process.

Keep those CPU/GPU cycles going.