Any word about client and cores for Apple Silicon?

[belloq]

Since Apple had test hardware available since summer, and now released officially this week, I was hoping to have heard something about folding@home running on the Apple M1 and future chips. Has there been any official news?

[MeeLee]

Before such hardware would hit the stores, is still at least 6 months away.
It'll all depend on if they're basing their M1 CPU on the Cortex A70 series CPUs, or doing their own proprietary design (most likely).
If it's based off a Cortex A70-line of CPUs, it means that there might be compatibility with FAH for ARM.
If it's based off Neonverse or anything newer, it might or might not.
Come back in 6 to 9 months. We'll possibly have a better answer by then.

[Joe_H]

No official news so far. Apple did release developer platforms that used one of their recent ARM based processors months ago, no idea if any developer connected with F@h obtained one.

As for hardware release, the models just announced on Tuesday can be ordered now and are supposed to ship within one to two weeks.

[belloq]

...Apple did release developer platforms that used one of their recent ARM based processors months ago, no idea if any developer connected with F@h obtained one.

Right, that's what I was hoping, that one or more dev folks associated with F@H would have obtained the hardware Apple made available over the summer.

Before such hardware would hit the stores, is still at least 6 months away.

Available in stores Nov 17, this coming week. Delivery from online orders is now later in the month.

It'll all depend on if they're basing their M1 CPU on the Cortex A70 series CPUs, or doing their own proprietary design (most likely).

Pretty sure it's proprietary. The hardware they released over the summer, which devs were supposed to use to start modifying code, was using the exact same CPU as the iPad Pro.

[JimboPalmer]

I am not raining on your parade but let me suggest another timeline.

When Intel helped F@H, we got multi threaded Cores running on SIMD.
When AMD helped F@H, we got GPU Cores.
When NVidia helped F@H, we got GPU Cores.
When Sony Helped F2H, we got PS3 Cores.

You want a Apple ARM core, one company in the world can offer expertise on that.

IF F@H is left to develop it on their own, expect a long development cycle. (F@H has never managed to make an Apple GPU Core)

I could be wrong. I have been wrong before, I will be wrong again in the future.

[gunnarre]

Part of the reason why we don't have an Apple GPU core is that there was a long-standing bug in OpenCL in MacOS. MacOS is now using Metal for GPU compute, and anyone who wants can write applications for it if they buy a developer license. Due to the very limited developer resources that FAH has available, and the limited number of Macs with discrete GPUs in them, it has so far not been decided to convert OpenMM to Metal.

With the new Macs coming with ARM cores and Metal GPU cores, that calculation might change. Instead of just a few Macs being available, it will likely be every Mac going forward. So it doesn't necessarily have to be Apple who provides the developer time to support ARM cores and Metal GPU - it just has to be someone with the time and know-how.

[Neil-B]

It doesn't have to ... but past experience might indicate that it is only with the vendor stakeholders engagement and support that this is likely to happen? ... I think what people are trying to say is "don't hold your breath" ... Yes it would be nice if someone came forward with access to the kit, the time and skills required, and the incentive to develop what is being asked for, but at this time personally I don't see it happening soon

[MeeLee]

Most apple computers wouldn't be very efficient at folding anyway; since most of their designs are low power designs (phones, tablets, laptops).
They only have few servers/tower desktops, and only a very small elect few use those.

I doubt Apple will do much effort to support folding, although, one generous donation from them wouldn't be completely out of the question (with Covid and all).
It'd probably be too early to know.
The problem is, that this is a completely new design from Apple.
Not sure if they're going to upgrade that CPU anytime soon; if bugs will be found (since it's relatively new).
If Apple chooses to continue using their M1 (or compatible M1 upgraded chips in the future), I'd say the chances are greater.
However, if they go towards an incompatible architecture in the future (for ARM, a jump from A30 to A50, or A50 to A70 cores means incompatibility), it's probably not likely that they will spend resources on creating software for the M1 CPU.
MS did that for a while, with Windows for ARM. And haven't seen much profit from it at all (on none of their programs).

Like, Apple probably will think in the lines of:
"No one would buy an Apple device with M1 CPU just because it supports a folding client..."

On the other hand, if they have several devices with that chipset, the chances are greater, someone creates a compatible client for it.

As in GPU performance. I think it'll be close to what we get on our phones.
Not really worthy of OpenCL compute.
Think 500Gflops at best. With only 8 unified pipelines, I think we're closer to 250Gflops max. That's worse than the IGP from an intel 7th gen CPU.

[calxalot]

Apple says 2.6 TFLOPS.
Early benchmarking looks good.
https://www.tomshardware.com/news/apple ... erformance

[bruce]

IMHO, it's really a thermal issue. I doubt Apple is going to ever design their GPUs to run close to 100% compute. Apple expects the GPU to do a good job at rendering the desktop images and videos but as far as continuously computing geometry without converting it to pixels, that takes a lot of Watts and a lot of active heat dissipation in the cooling subsystem. Even Windows laptops are limited in their ability of running FAH because they can't dissipate the heat.

NVidia and AMD do a pretty good job of designing GPUs that blow a lot of hot air out the backplane and they can handle continuous scientific computing. Phones/tablets/Macs just aren't designed with dissipating that heat in mind. Laptops typically down-clock the CPU and iGPU to reduce the power enough for it to cope.

[gunnarre]

The new Mac Mini, which has an M1 SoC, has pretty substantial cooling fan. The MacBook Pro has a small laptop fan, while the new MacBook Air is fanless. If the GPU performance pans out, it might be better to run GPU compute only on the notebooks, while the Mac Mini might be OK with continuous CPU and GPU compute.

[MeeLee]

Apple says 2.6 TFLOPS.
Early benchmarking looks good.
https://www.tomshardware.com/news/apple ... erformance

I would take that number with a grain of salt.
If we look at their prospect of 3.5x faster,
New benchmarks just came in,
Its just as fast as a core i5/i7 quad core (with HT) intel 8th gen cpu, running at 3Ghz base/4Ghz boost.
Just your average desktop, that mainly pushes a good gpu, like a 2060 to 2080 for gaming.

As far as the gpu numbers go, with 8 compute units they'll reach 2.5tflops if their gpu frequency reaches 8-9Ghz (or something; highly unlikely).
More than likely it runs at 2.6Tops (8 or 16 bit shaders), as I do believe Apple wants to step in the ai bandwagon. But those quarter precision, or half precision calculations aren't really used in fah.

[bruce]

.... But those quarter precision, or half precision calculations aren't really used in fah.

I would suggest that that sentence might be more accurate if it ended with "...aren't really used for scientific calculations." (Fah isn't the only scientific endeavor that depends mostly on FP32 calculations with a smattering bit of FP64.)

It seems that the apple is not the only company with a sales department that has redefined what we used to call shaders/cores/compute_units to also count half precision (FP16) and quarter precision (FP8), all of which can be used for other things but not for science.

[MeeLee]

I read an article comparing the M1's gaming performance to a 1060 3GB.
Same fps, also claiming 2 to 3 Tflops.
It totally makes no sense how such a chip would be able to hit the same fps a 1060 does, unless the benchmarks are mobile oriented, and the 1060 is just being capped (by ram or thread count or something), essentially not running at full potential.
The M1 has like what, 8 cores and 128 shaders (or EUs), running at an unknown frequency (probably somewhere between 2 and 3Ghz) bouncing heads with a GPU having almost 10x the shaders?
Unless Apple somehow magically did a technological leap resulting in 10x performance... I think some of their numbers don't make any sense...

[Joe_H]

For now this discussion on the M1 GPU is moot. Unless support for Metal is added to OpenMM and a GPU folding core created it will not be usable for folding.

As for the technical aspects, little is available so far on how the 8 GPU cores are organized internally in the form of shader processors. What has been said by Apple is up to 2.6 TFLOPs of FP32 processing and nearly 25,000 simultaneous threads. The FP8 and FP16 processing appears to be provided by a separate Neural Engine core. Early testing appears to support the 2.6 TFLOPs figure, graphics performance is on par with other GPUs of the same approximate FLOP rating - midrange cards from about 2 years ago.

Someone will get better performance figures once some compute applications or benchmarks are available. Then some idea what the sustained processing rate will be can be tested with the chip running at its thermal limit instead of peaks needed for game graphics.