Folding Forum

muziqaz wrote:
If you have only fairly modern integrated GPU from AMD, it is always worth a try to set that going, providing you have good cooling. In most ases modern AMD integrated GPUs are theoretically a lot more powerful than CPUs next to them.

That is what I'm doing now, and your statements are correct. Using the Vega 11 only (not the CPU slot) gets me 16-17K PPD. Using the Vega 11 and CPU slot only bumps it up to about 20K PPD. Actually even with stock cooling the on board graphics don't generate much heat. I have my CPU fan profile set to ramp up quickly, but run quiet until 50C or so. With only the Vega 11 (I've been calling it the GPU slot though not a dedicated card) my temps stay in the 42-45C range. Adding any CPU cores naturally ramps up heat, fan speed, and power.

muziqaz wrote:
I'm not following you here
but
3900x@24threads produces 390k PPD
3900x@21Threads produces 300-350k PPD + 2thread slot of 16-17k + Radeon 7 with 1m-1.5m-2m PPD

When 21 Thread slot is running alongside 2 Thread one, boost is holding at 4075Mhz (which is probably upper limit for AVX256 workloads for the CPU), but AMD boost technology has been known to get the most out of CPUs, so if you pause 21 Thread slot and leave 2 Threads to run on itself, that 1c can boost to max ~4.4ghz.
The main 3000 series advantage over 2000 series Ryzens is full AVX256 support, instead of 2x128bit AVX on 2000 series. FAH still sees that as AVX256, but it kinda runs at half the speed of full AVX256.
So lets say my 3900x@21Threads is roughly twice as fast (in PPD produced) than my 2700x@15Threads which is pretty amazing.
With current influx of COVID projects which are quite small, my 3900x is not scaling to well, but still

Interesting stuff, and now I see what you were doing regarding the cores/threads. In my case the CPU adds so little that I may as well quit folding with it. For the work it does the power and heat factor aren't worth it really. I could buy a really cheap dedicated GPU and accomplish more work for the power used.

Being I'm not a gamer and have minimal use for my computer these days, I just did the 2400G upgrade to keep life simple and cheap. But if I keep folding I might have to upgrade and buy a GPU so I don't feel like my system is in the stone age.


I'm glad to see that so many such as yourself are dedicating a lot of resources to FOH.

bignerdguy wrote:Well there are a lot of reasons why thew 3000 series CPU's are faster: memory pipes, Threads, etc but i only have the 2000 series in the Ryzen 5 class. I leave the setting on 4 (1 thread for GPU, and the rest for F@H) which means it uses all but one thread for the CPU and the last is for the GPU (the Ryzen 5 2400G has 4 CPU cores w/ 8 threads). Remember the setting in the config is for CPU's not threads as in my system setting it to 4 maxes out all 4 Cores/8 threads totally. I should also mention i was wrong about how long the GPU (Vega) was taking, it took about 8-9 hours, not 22. It was also maxed out as well the entire time. The CPU would also take about as long with the Vega turned on. Now with the Vega disabled the CPU and 590 GPU complete WU's in a little over 3 hours. I have also just ordered a Ryzen 7 3800x and am waiting on amazon to deliver it. Once i get that i will install it and see how fast it runs things. I will let you all know, but for now i am leaving the Vega disabled since it isn't nearly as fast as the 590.

From my understanding FOH settings are actually looking at cores on the AMD chips, but I could be wrong about that. If I use default settings to run both the Vega 11 and CPU, the setting set by the client is 7, but displays as 6 in the status on the advanced control, since it is still linking available threads to cores. And though I haven't done many CPU WU's, it seems to run CPU WU's quicker than you are stating. It's possible memory is the difference.

In my case I bumped the CPU number down to find that happy medium of where I wanted temps, fan speed, and power. Since the CPU isn't contributing much anyway I can just keep it cool and let it work a little longer.

As for throughput, your correction for how fast the onboard Vega is puts it in line with the other 2400G numbers I've seen, so that now makes sense.

With that new CPU and memory, both that 590 and the CPU should put out some decent numbers.


As for me, it's the first time I've had an urge to get a graphics cards and it's not for work or gaming!

current projects due to the way they are set up will not work on 5, 7, 11, 13, 14, 17, 19, 22, 23 threads. If you have so many threads available and set up the slot to use them, Fahcore_a7 will reduce thread count to first working count. So from 5 to 4, from 7 to 6, from 14 to 13 to 12, etc. It is not because FAH is clever enough to see HT
I can see that in your cases, you worry about temps, etc.
If temps, power usage wasn't an issue, I would set 1 slot for 6 thread CPU, and another slot with GPU (vega11). That would give most PPD. If on top of that there is a discreet card (like RX590), then I would set up 1 slot with 6 thread CPU, 1 slot for Vega GPU, 1 slot for RX590. That way it would produce most PPD.
regardless of what you guys are thinking, 4core 8thread CPU was high end just couple of years ago, and is still considered very valuable in Folding@Home. That 2400G is still at the top 10% probably, considering people still fold with CPUs from 10+ years ago

Well i am no longer too worried about temps since the MSI Afterburner app took care of that with a little judicious editing of the fan profile. As for the CPU, it works but it works much better with the Vega disabled. With the Vega running it was taking the cpu anywhere from 4-8 hours for one WU to be processed. With it disabled it took no more than 3:15 Max for most and some WU's as little as an hour and a half. Since then however i managed to buy a new Ryzen7 3800x and replaced the Ryzen 5 with it. Now WU's complete in less than an hour. However i ran into a major issue with the system memory. It apparently is the source of my bottleneck as it wont run reliably past DDR-2666. I had it cranked to DDR-2800 and put the new CPU to 4.4GhZ (should be able to do 4.5) and the HDD MBR corrupted almost at once (couple of reboots actually). I just got done reinstalling the entire OS and apps this morning. Now the whole thing is running smoothly at DDR-2666 and 4.0 GHz so i am going to leave it at that setting until my new memory i needed to fix this gets here. Hopefully i can max out the performance of this beast soon. In the meantime the benchmarks are looking sweet and F@H is cranking and banking!

I would suggest leaving CPU at stock. And concentrate on memory tuning. AMD boost technology is very good at using the most of the CPU clocks. You should see 4ghz clocks regardless during folding. You would gain a lot of performance by clocking memory higher (if your RAM lets you do it, of course).
3800x has very good memory controller which lets memory clock up to 4ghz+ (providing you have memory sticks which clock that high).
3800x with 2666Mhz RAM is very bottlenecked
I would suggest DRAM Calculator to get your RAM clocks up (if possible)

Yeah i set the CPU at stock 3.9 GHz and it auto boosts to 4 GHz during heavy use. The memory is VERY bottle necked and wont run any faster than 2666 reliably. That's why i just bought 2 new sticks though i have to wait for Amazon to ship them.

bignerdguy wrote:Yeah i set the CPU at stock 3.9 GHz and it auto boosts to 4 GHz during heavy use. The memory is VERY bottle necked and wont run any faster than 2666 reliably. That's why i just bought 2 new sticks though i have to wait for Amazon to ship them.

Once you receive new RAM, download DRAM Calculator. Its extremely helpful to get the most out of your memory

OK good news is it was the RAM that was bottle-necking everything. The latewncy curve was awful! New memory has a MUCH better curve and a lot better latency, like 25ns less! Bad news is DRAM Calculator doesn't support my memory ... yet. The micron RAM SPD doesn't list what die it is and all the settings for the ones that Calc does list don't work with my memory. However i got it to work just fine by using the manufacturer suggested settings which the Mobo doesn't apparently do via SPD correctly. So now i have everything tested and running at DDR4-3200 and the CPU @ 4.1GHz with no issues. Now to get Amazon to ship me the second memory stick they forgot to send with the first one...

Thanks for the suggestions, i am adding the software to my always growing bag of Tech-tricks.

bignerdguy wrote:OK good news is it was the RAM that was bottle-necking everything. The latewncy curve was awful! New memory has a MUCH better curve and a lot better latency, like 25ns less! Bad news is DRAM Calculator doesn't support my memory ... yet. The micron RAM SPD doesn't list what die it is and all the settings for the ones that Calc does list don't work with my memory. However i got it to work just fine by using the manufacturer suggested settings which the Mobo doesn't apparently do via SPD correctly. So now i have everything tested and running at DDR4-3200 and the CPU @ 4.1GHz with no issues. Now to get Amazon to ship me the second memory stick they forgot to send with the first one...

Thanks for the suggestions, i am adding the software to my always growing bag of Tech-tricks.

Download Thaiphoon 1.6. Read SPD info of one of the chips, and look at Die Density and Count section on right hand side. That will tell you what die chips are I just bought Team Group Vulkan 3600Mhz ram with some Hynix C-Die. DRAM Calculator gave me quick and stable timings to get me started
P.S. DRAM calculator is being actively developed. Author just announced he will be releasing new version of it with more ram support

muziqaz wrote:3900x@24threads produces 390k PPD
3900x@21Threads produces 300-350k PPD + 2thread slot of 16-17k + Radeon 7 with 1m-1.5m-2m PPD

That is very high, I'd be interested in your setup.

My stats:
OS: Win 10
CPU: Ryzen 9 3900X with Wraith Prism cooler
Other hardware: Mainboard MSI MPG X570, 32 GB RAM
Slot setup: 1x12, 1x8

With stock settings, I'd hit the temperature ceiling immediately, clock never exceeded 4GHz, and the rig sounded like a starting jet - for no more than 180k-200k PPD.
So I undervolted to 1.035V and limited to 3.8 GHz, saving 40% power with 180k-190k PPD, so not a significant loss. Happy to tinker again if I can get a better "bang for the buck".
Maybe it's just the number of CPUs per slot?

(Note: What I posted here for PPD went up slightly, it's probably down to the projects I got.)

uyaem wrote:

muziqaz wrote:3900x@24threads produces 390k PPD
3900x@21Threads produces 300-350k PPD + 2thread slot of 16-17k + Radeon 7 with 1m-1.5m-2m PPD

That is very high, I'd be interested in your setup.

My stats:
OS: Win 10
CPU: Ryzen 9 3900X with Wraith Prism cooler
Other hardware: Mainboard MSI MPG X570, 32 GB RAM
Slot setup: 1x12, 1x8

With stock settings, I'd hit the temperature ceiling immediately, clock never exceeded 4GHz, and the rig sounded like a starting jet - for no more than 180k-200k PPD.
So I undervolted to 1.035V and limited to 3.8 GHz, saving 40% power with 180k-190k PPD, so not a significant loss. Happy to tinker again if I can get a better "bang for the buck".
Maybe it's just the number of CPUs per slot?

(Note: What I posted here for PPD went up slightly, it's probably down to the projects I got.)

It's all project dependent. before COVID, 350k was norm for 21Threads, but now COVID projects has variety of Atom counts, and on small counts multithreaded CPUs like mine suffer.
You bottleneck is cooler 100%. I have Noctua D15 monster, and I do not limit anything, so I let it bounce at 80C all the time. Also I have 16GB of 3600Mhz CL14 RAM running. I really doubt your 32GB are clocked that high.
Also F@H PPD is driven by Quick Return Bonus (QRB), the quicker one returns the WU the greater the bonus is. Even 1 second of TPF can give a lot of extra points.
Another thing is Windows Scheduler. If you run multiple projects on a CPU, thread jumping is a thing which trashes cache, thus reducing performance. One project on one CPU gives access to 3900x's huge L2 cache, that seems to be the biggest advantage

muziqaz wrote:

bignerdguy wrote:OK good news is it was the RAM that was bottle-necking everything. The latewncy curve was awful! New memory has a MUCH better curve and a lot better latency, like 25ns less! Bad news is DRAM Calculator doesn't support my memory ... yet. The micron RAM SPD doesn't list what die it is and all the settings for the ones that Calc does list don't work with my memory. However i got it to work just fine by using the manufacturer suggested settings which the Mobo doesn't apparently do via SPD correctly. So now i have everything tested and running at DDR4-3200 and the CPU @ 4.1GHz with no issues. Now to get Amazon to ship me the second memory stick they forgot to send with the first one...

Thanks for the suggestions, i am adding the software to my always growing bag of Tech-tricks.

Download Thaiphoon 1.6. Read SPD info of one of the chips, and look at Die Density and Count section on right hand side. That will tell you what die chips are I just bought Team Group Vulkan 3600Mhz ram with some Hynix C-Die. DRAM Calculator gave me quick and stable timings to get me started
P.S. DRAM calculator is being actively developed. Author just announced he will be releasing new version of it with more ram support

I did download it and run it. i have v1.6.3 and it gives the following:

DRAM COMPONENTS
Manufacturer
Micron Technology
Part Number
MT40A1G8??-093E:?
Package
Standard Monolithic 78-ball FBGA
Die Density / Count
8 Gb / 1 die
Composition
1024Mb x8 (64Mb x8 x 16 banks)
Input Clock Frequency
1067 MHz (0.938 ns)
Minimum Timing Delays
15-15-15-36-51
Read Latencies Supported
16T, 15T, 14T
Supply Voltage
1.20 V
XMP Certified
1600 MHz / 16-18-18-36-55 / 1.35 V
XMP Extreme
Not programmed
SPD Revision
1.1 / September 2015
XMP Revision
2.0 / December 2013


That doesn't tell me what version of the chip it is. However i used the XMP settings at the bottom and it works just fine at 4.1GHz.

muziqaz wrote:It's all project dependent. before COVID, 350k was norm for 21Threads, but now COVID projects has variety of Atom counts, and on small counts multithreaded CPUs like mine suffer.
You bottleneck is cooler 100%. I have Noctua D15 monster, and I do not limit anything, so I let it bounce at 80C all the time. Also I have 16GB of 3600Mhz CL14 RAM running. I really doubt your 32GB are clocked that high.
Also F@H PPD is driven by Quick Return Bonus (QRB), the quicker one returns the WU the greater the bonus is. Even 1 second of TPF can give a lot of extra points.
Another thing is Windows Scheduler. If you run multiple projects on a CPU, thread jumping is a thing which trashes cache, thus reducing performance. One project on one CPU gives access to 3900x's huge L2 cache, that seems to be the biggest advantage

80°C, I wish... 91/92 was normal until I undervolted.

Indeed, the 32 GB RAM is 3200 C16 - seems I skimped at the wrong end.
EDIT: Just realized XMP does not seem to be enabled, so I may have something to gain here still.

I'll try your 21/2 setup and see what I get once I'm through with the current WUs.
Thank you so much for your input.

@bignerdguy

Look like micron E-die.
If it works with XMP, then awesome stuff

uyaem wrote:

muziqaz wrote:It's all project dependent. before COVID, 350k was norm for 21Threads, but now COVID projects has variety of Atom counts, and on small counts multithreaded CPUs like mine suffer.
You bottleneck is cooler 100%. I have Noctua D15 monster, and I do not limit anything, so I let it bounce at 80C all the time. Also I have 16GB of 3600Mhz CL14 RAM running. I really doubt your 32GB are clocked that high.
Also F@H PPD is driven by Quick Return Bonus (QRB), the quicker one returns the WU the greater the bonus is. Even 1 second of TPF can give a lot of extra points.
Another thing is Windows Scheduler. If you run multiple projects on a CPU, thread jumping is a thing which trashes cache, thus reducing performance. One project on one CPU gives access to 3900x's huge L2 cache, that seems to be the biggest advantage

80°C, I wish... 91/92 was normal until I undervolted.
Indeed, the 32 GB RAM is 3200 C16 - seems I skimped at the wrong end.
I'll try your 21/2 setup and see what I get once I'm through with the current WUs.

Thank you so much for your input.

Having 32GB of 3200Mhz RAM is not skimping Very hard and expensive to have 32GB clocking higher than that. I doubt if my 4000Mhz b-dies would do that if they were 32GB

91-92C is thermal limit of the chips, or near it, so you made a good call undervolting