ARM CPUs (not yet)

[MeeLee]

ARM might become more interesting in the future, if X86 CPUs will start to hit their physical limitations, at ~5nm.
Then there's no room to optimize x86 CPUs anymore with current silicon technology.
That's when either ARM at 7nm could bounce head to head with x86 at 5nm, or x86 CPUs will start implementing ARM/RISC cores in their design, in an attempt to do light calculation on those CPUs, while the heavy x86 oriented workload is distributed over a few x86 cores on the package.

More than likely, ARM might slowly develop their core count, address their high idle TDP problem, and make a chip that's totally competitive in the market.
But until then, AMD has a few generations more experience. And ARM chips of today, would have made great CPUs 5 years ago!

[bruce]

To support both FP32 projects and FP64 projects, changes must be made at the server level so those who can't support FP64 won't get assignments they can't process. Then the science team will also mange their projects in separate streams.

[_r2w_ben]

Hardware without a FP64 FPU can still do double-precision math in software mode. It will be more than 2x slower but it's still possible. Just like there is an SSE2 and an AVX_256 version of the CPU core compiled with different flags, there would need to be an ARM core with and another without native FP64.

[MeeLee]

I'm sure software mode can 'borrow' 1 core for FP64 computations, while doing the rest in 32bit

[bruce]

Hardware without a FP64 FPU can still do double-precision math in software mode. It will be more than 2x slower but it's still possible. Just like there is an SSE2 and an AVX_256 version of the CPU core compiled with different flags, there would need to be an ARM core with and another without native FP64.

The real question is whether a programmer using OpenCL 1.2 needs to revise the API calls to OCL code or whether the drivers recognize that they need to manually inject an extra macro into the processing.

In fact, it is MORE THAN 2x slower. Even when the hardware supports the operations, some GPUs benchmark FP64 results as much as 12x slower that FP32 but FAH doesn't need to build a new FAHCore for that. The FAHCore works with MIXED precision or it works with SINGLE precision. The science team needs to segregate the distribution of their projects into SINGLE or MIXED precision paths.

[toTOW]

ARM CPUs support might be closer than we think :

On April 16, Neocortix completed a build of FaHCore with the new Gromacs 2020.1 engine, for both x86_64 and aarch64 targets, with SIMD optimization (AVX2_256 and ARM_NEON_ASIMD, respectively). On April 20, Neocortix completed a build of FaHClient. With these milestones reached, the project is now transfered to the Folding@Home team for review, verification, testing and publication.

From : https://www.neocortix.com/coronavirus

[MeeLee]

I think x86 has a longstanding record of optimizing performance;
While ARM has a long standing record of keeping power limits low.
When ARM is going to tune for performance, it will more and more become like x86.
Less power efficient than ARM, and less performance than x86.
Even with their excellent power efficiency, ARM is no match for x86, just like x86 is no match for GPUs.
It's just that ARM has not grown as much as x86 architecture. They don't have power saving C-states, turbo boost frequencies (or very few have them), etc...

I want to see an 80 core ARM cpu crunch, but I think it'll be slower than a 64 threaded Threadripper.

[oreggin]

ARM64 client is done on beta lane but I can't get any WUs for my ARM64 machines:

Code: Select all

ERROR:WU00:FS00:Exception: Could not get an assignment

[Joe_H]

ARM64 client is done on beta lane but I can't get any WUs for my ARM64 machines:

Code: Select all

ERROR:WU00:FS00:Exception: Could not get an assignment

The ARM client has not been released. The persons who made announcements elsewhere were premature and not part of the F@h Consortium administration group. There are bugs that need to be fixed before work for ARM will be released.

When released ARM processing will be Beta. Membership in the Beta Team is strongly recommended - viewtopic.php?f=66&t=18031

[Foliant]

[...]ARM processing [...] Membership in the Beta Team is strongly recommended - viewtopic.php?f=66&t=18031

Is there any *minimum* Hardware requirement to be part of this?
I only could contribute a Raspberry Pi3 B but im not familiar with performance on ARM Hardware.

Regards,
Patrick

[PantherX]

...Is there any *minimum* Hardware requirement to be part of this?
I only could contribute a Raspberry Pi3 B but im not familiar with performance on ARM Hardware...

There isn't any minimum hardware, you can look at the requirements here: viewtopic.php?f=16&t=8

[Joe_H]

Once they are ready to announce the ARM beta test, minimum hardware requirements should also be part of hat announcement.

[bruce]

Part of FAH's problem is the wide variety of hardware that people want to donate. Either they have to establish minimum requirements so everybody can meet them while maintaining a reasonable range of PPDs (given the way the bonus points work on fast hardware) or they somehow have to maintain various classes of projects and an assignment logic that gives reasonable assignments to both fast and slow hardware (ugh).

They're working on a plan to benchmark every different hardware platform on a variety of projects. Then the work out a complex assignment process that sends work to the fastest GPUs and different work to the slowest iGP or slowest Arm device and never interchanges those assignments. BIG JOB.

I want a client that will run on my Android phone without overheating it, but it's really, really slow compared to the top hardware platforms. My desires are probably unrealistic.

[oreggin]

Thanks for all. I mean ARM64 beta is available: https://foldingathome.org/beta/ Not others announcements elsewhere
I have two SBCs, OdroidC2 and a Pine Rockpro64, I set client-type=beta on both and now I get a8 core and 16810 project WUs. Started with light folding power to prevent overheating.
What is not clear to me: "Beta Team Membership". Its mean I should change team number ID in Fahclient config? Or it is a privileged membership in Beta section of this forum?

[Neil-B]

the links on that page are for invited devs and testers ... If no formal announcement has been made then they aren't intended for full public beta testing.

Membership of the Beta Team allows post privileges for the Beta Forum ... This allows folders who wish to commit to helping test and reporting issues with new Projects and Cores to assist the FaH Team with focused Beta Testing ... The use of Beta Flag in the slot configuration allows the Beta Team to access Projects under test before they are released to Advanced and then full access ... see viewtopic.php?f=66&t=18031

You continue to use your username/team/passkey as currently - you simply get to commit to help/report .. You stand a greater risk of getting issues and it isn't necessarily the best thing to do if Points are what drives you ... and it really helps if you know your kit is stable and has no inherent weaknesses.

At the moment with no ARM Client/Cores released there isn't a dedicated pool of ARM testers (though no doubt some existing testers probably have ARM kit ... If you look at the Beta Forum and want to commit the instructions are there on how to apply to join (iirc).

If I have anything wrong here one of the mods will correct me