Folding Forum

Overclockers posted some impressive results from ryzen CPUs for folding at ~400k PPD, though a non overclocked ryzen 5 should probably hit closer to 100kPPD.
https://www.overclockers.com/forums/sho ... -Ryzen-PPD
https://www.overclock.net/forum/55-over ... 100?page=3

I'm really surprised, those ryzens run a lot better than basically all Intel CPUs,
Xeons, core i, Celeron, Pentium D, N, and G processors, and core2duos I had all were folding between 500-1000PPD per core, per Ghz!

I guess it all depends on the lithography of the processor. 7 to 14nm processors are extremely efficient. Be it for graphics cards, or CPUs.

The cost of electricity is the real cost for folding, not the hardware, because it's continuous, and it compounds in summer when you combine it with added AC cost in some areas.
If you live in some cold places, it makes more sense warming up your home with a rack of graphics cards, than using an electric heater.

There still are projects that are CPU only, but the majority of the projects are now done on GPU, because of the increased efficiency and speed in most cases (apparently ryzen and perhaps core i9 CPUs excluded).
CPU folding makes most sense on the native client, and Macs, where GPU folding is either inaccessible, or no support available for GPU folding.
Though I would say with HTML5 there should be a possibility to fold on GPUs as well for the native client, but it doesn't exist yet.

Because Macs and native client (from browser) focus on CPU folding, it makes most sense to primarily choose GPU folding when you can, and add CPU folding only when you want to. The return of CPU folding PPD wise, is less than GPU wise, for the same amount of power used. Though ryzen decreased that gap it seems.
Instead it makes more sense running 2 graphics cards for folding than 1 with CPU.

in theory the floating output of ryzen should be half of what intel does, due to 128 bit AVX unit, but its indeed impressive especially in watt/ppd, if next version of compute core supports avx2 or even avx512, the results should be even more impressive
im looking forward to see the ppd of 9700k but i can't find anywhere, maybe CPU folding are irreverent at this point

ProDigit wrote:Overclockers posted some impressive results from ryzen CPUs for folding at ~400k PPD, though a non overclocked ryzen 5 should probably hit closer to 100kPPD.
https://www.overclockers.com/forums/sho ... -Ryzen-PPD
https://www.overclock.net/forum/55-over ... 100?page=3

I don't see the 400k PPD for the Ryzens (and I wouldn't believe it if I did). I think that is for the Intel.

My own Ryzen results (1700, 2700 and now a 2600, all on Ubuntu 18.04.1) indicate that you don't want to use all 16 cores; probably 12 is better.
And I think the Intels do a little better; at least my i7-8700 seems to thus far. You should be able to get a little over 130 k PPD on a core_A7 (not overclocked).

@ProDigit: In the linked article they talk about 4x Ryzen CPU make 400k PPD and 1x Ryzen CPU makes 100k PPD on Core_A7 (AVX).

@Frisa: Ryzen CPU uses 2x AVX 128bits only for AVX2 256bits which is not supported by FAH yet. So with AVX(non 2) 128bits it has full performance.

@JimF: It has only 8 cores but 16 threads. A core is shared between its threads.

foldy wrote:@ProDigit: In the linked article they talk about 4x Ryzen CPU make 400k PPD and 1x Ryzen CPU makes 100k PPD max.

@Frisa: Ryzen CPU uses 2x AVX 128bits only for AVX2 256bits which is not supported by FAH yet. So with AVX(non 2) 128bits it has full performance.

@JimF: It has only 8 cores but 16 threads. A core is shared between its threads.

avx1 is 256bit too, although theres 128 bit option for compatibility with SSE. if u checked WU's log(md.log), it would mention avx_256

foldy wrote:@JimF: It has only 8 cores but 16 threads. A core is shared between its threads.

I am using the term "core" to include virtual cores, which is how BOINC and the FAH Client sees it.

To be precise, we need to specify threads or virtual-cores if that's what we mean or specify real-core or maybe hardware-core or even pair of threads, if that's what we mean. Both Intel and AMD like to make their advertised hardware seem more powerful than it actually is. In fact, two threads share the same floating-point hardware, whether you call them hyperthreaded like Intel does or simply cores like AMD does, leaving you to guess.

If you're not running FAH or similar scientific code (including many games) then your browser/Word-processor/etc. really can do what they need to do on a single thread (fixed-point arithmetic) so performance does increase whn you turn on HT but not when you're doing 3D calculations. The FAHCores for CPU contain a mixture of AVX loats, DP floats, SP floats, and integer operations. Thus there is maybe 10% - 20% performance increase because the integer code can run while the other thread is using the shared float hardware, but when they both happen to be competing for the same float hardware, only one can be producing results.

When a GPU FAHCore is moving data around in main ram or sending/receiving on the PCIe bus, writing checkpoints, etc. that code won't be competing with whatever thread is sharing the FPU, but there also will be some percentage of 3D float operations. I've never profiled the GPU FAHCore. If somebody has the tools to do so, I'd be interested in the results.

I have a working Xeon rig, 10 cores, 20 threads, and even on that machine there's no real be edit running all cores. I expect the same on ryzen, which shares even more between cores than Intel (hyperthreading).
From a perspective of GPU folding,
I'm currently running some tests on my system, but if I disable cores to 3 cores 6 threads (HT), or 4 cores 4 threads (no HT), feeding my GPUs, I get about the same PPDs and CPU load.

Once I set it to 2 cores, 4 threads on 3 graphics cards (98% CPU load), my PPDs drop.
I have a feeling that hyperthreading cores have to be counted like 1/2 core, both in CPU and GPU folding.
Both the NaCl as the desktop client treat them as full cores, and you can actually see a performance dip trying to utilize all of them.

@Foldy, thanks for clearing that up, I was surprised a ryzen would get that high, especially considering my 10C 20T Xeon only gets 35k PPD.

I have list of at least 10 donors (no-namers but byteball most off them) that have extreme numbers.. I could list them since there are no names or nationality other then teams and CPU ID's.

Yes, if you're using a CPU slot (or the NaCl client), then a pair of hyperthreaded (virtual-)cores are very nearly equal to one (real) core, so the hyperthreaded core is very nearly equal to half a hardware core. (actually, it's closer to 55% to 60% each, giving about a 10% to 20% performance increase. In my experience, that's not the same when they're supporting GPU folding. YMMV.

My testing was done on various Intel GPUs. I have no solid test results for AMD CPUs.