Dedicated rig High end GPU Low end CPU?

[PS3EdOlkkola]

In defense of Fury X, I'd say muziqaz has it right. I run 9 (nine) Fury X's on four separate rigs with CPU types including FX-8350, 4771, 4790 and 5930, and they've been incredibly reliable, quiet and they run fairly cool. I've added an extra Noctua 120mm fan on each one, running at about 50% of rated speed, and the temps dropped another 3 to 4 deg C, and it's still very quiet. When a Fury X rig is running, I can pretty much forget about needing to tweak it to run reliably - it just does. The average PPD of each Fury X slot is 387,056 right now. Personally, I like to see how AMD performs against Nvidia, but the mix is about 30% of my total folding slots use AMD systems (9 Fury X and 1 R9-295-X2).

[muziqaz]

In defense of Fury X, I'd say muziqaz has it right. I run 9 (nine) Fury X's on four separate rigs with CPU types including FX-8350, 4771, 4790 and 5930, and they've been incredibly reliable, quiet and they run fairly cool. I've added an extra Noctua 120mm fan on each one, running at about 50% of rated speed, and the temps dropped another 3 to 4 deg C, and it's still very quiet. When a Fury X rig is running, I can pretty much forget about needing to tweak it to run reliably - it just does. The average PPD of each Fury X slot is 387,056 right now. Personally, I like to see how AMD performs against Nvidia, but the mix is about 30% of my total folding slots use AMD systems (9 Fury X and 1 R9-295-X2).

when compared to nvidia cards, AMD does not look good in regards to folding. Heavily clocked 980ti would usually net you ~600k ppd. I have quite a lot of hope for Polaris though

[JonasTheMovie]

I don´t want a heavily clocked 980ti next to me while relaxing in the sofa....

[muziqaz]

I don´t want a heavily clocked 980ti next to me while relaxing in the sofa....

I was just saying for reference

[JonasTheMovie]

muziqaz, I know, I know.

PS3EdOlkkola, do you oc the FuryXs at all? How come you choose to cool them that low, any performance gain? Any idea why I average 330-350?

[muziqaz]

muziqaz, I know, I know.

PS3EdOlkkola, do you oc the FuryXs at all? How come you choose to cool them that low, any performance gain? Any idea why I average 330-350?

the amount of PPD depends on WUs
mine are averaging as well around 350k

[JonasTheMovie]

To have 9 slots averaging 10% higher is significant.

[PS3EdOlkkola]

@JonasTheMovie - I do not overclock Fury X. I've found that overclocking even a little seems to make them unstable, primarily in the phase where after one work unit is uploaded and the next downloaded (next-unit-percentage =100) the GPU will hang and not start folding the next work unit. Others may have a different experience overclocking, but for me it isn't worth the hassle on AMD GPUs.

I add the extra fan because it's a cheap solution to reducing heat, which prolongs the life of the GPU and the system overall by moving heat out of the unit. I like to get two+ years of use out of my GPUs folding 24 x 7. As it stands now with 36 GPU folding slots, I'm replacing an average of about one GPU every couple of months, so running them at stock clocks with a little extra help from an auxiliary fan lengthens the life, and serves as a backup if the stock fan goes toes up (which has happened on the R9-295X systems). The Noctua's are incredibly reliable (although not cheap), quiet and move a good amount of air. I'm not a fan-boy (pun intended) of Noctua, just my experience with them has been very good: In the 15 folding rigs I have running now, there's an average of 5 Noctua fans in each one. Not one has failed, including after recycling Noctua fans from my old BigAdv rig into GPU rigs that are folding today (oldest Noctua fans running are vintage early-2011). I use extra fans on all my closed-loop water cooled systems, including the 980ti and Titan X GPUs.

As for performance, I do find that the Fury X's in systems with 2011-v3 sockets (with 40 lanes) on PCIe 3.0 x16 slots perform better than the 1150 sockets (4771 and 4790) on PCIe 3.0 x8 slots, which in turn performs better than the FX-8350 on PCIe 2.0 x8 interfaces. I hesitate to make a blanket statement about speeds and PCIe lane width, because there are many other variables to consider; but anecdotally what I tend to see performance-wise is in the order I mentioned. Also, I don't use the CPU to fold any work units on these systems, which can also negatively impact GPU performance. I'm speculating, but I believe that reducing the latency from main memory to the GPU will positively impact performance, which is why if the CPU is engaged in folding, the GPU has to wait to get serviced, increasing latency and reducing PPD. Again, a broad-brush statement that I'd have to put some hard numbers against to prove definitively.

[JonasTheMovie]

Ok, interresting. I keep my CPU free, but its a FM2+ boards with a a4-6300 apu. Been thinking of upgrading it later, but only if it would benefit folding. What kind of memory speed do you use?

[PS3EdOlkkola]

FX-8350 is Dual Channel DDR3 @ 2000 MHz (11-11-11-28)
i7-4771 is Dual Channel DDR3 @1600 MHz (11-11-11-28)
i7-4790 is Dual Channel DDR3 @ 1600 MHz (9-9-9-24)
i7-5930 is Quad Channel DDR4 @ 3000 MHz (15-15-15-35)

[JonasTheMovie]

Since my main interest is to contribute I havent changed the next unit percentage. But I will change it to 100% to see what effect that gives on the ppd.

[bruce]

...which is why if the CPU is engaged in folding, the GPU has to wait to get serviced, increasing latency and reducing PPD. Again, a broad-brush statement that I'd have to put some hard numbers against to prove definitively.

When you do put the numbers together, be sure to consider RAM. If the total of the working sets of all active programs plus OS plus disk cache leaves enough pages to retain all the pages needed to manage the GPUs exceeds the free ram, then it will take time to bring the necessary code in from virtual memory (disk). If the pages can sit in RAM until they're accessed, the latency is extremely small -- provided the priority of the GPU FAHCores is higher than the priority of the CPU cores. Except if it has to wait for the pages to load, Interrupting a background task is really very prompt.

I'll bet that loading from virtual memory is producing most of the latency.

When you add CPU processing, how many CPUs are typically free? I only notice the latency if all the CPUs are kept busy.

It should also be noted that FAHCore_21 offloads some important parts of the analysis to the CPUs, and FPU congestion would certainly delay the GPU processing if the GPU has to wait for the CPU to finish whatever has been assigned to it.

[JonasTheMovie]

We are kind of on topic here. My CPU is never heavy loaded, and logs of memory use shows 3GB max.

[muziqaz]

bruce, intel for me was always less tolerating to multitasking. whatever the CPU I had. maybe it is the nature of HT, or QPI or something else in the cpus, but whenever I used to use AMD CPUs, be it athlon 64 x2, phenom 1 x4, or phenom 2 x6 or FX 8350, I could throw at it whatever I like, system didn't care. Play games and fold, no problem, no stuttering, render a video, while doing something else heavy, no problem. But with intel, gaming while folding is no no, since I start getting stuttering. Rendering videos in background has some effect to the responsiveness of the system. And I am talking about 6 core E series CPUs with 4 memory channels. Obviously what you are talking about is bit different of what I am speaking, but I thought I would share my experience

[JonasTheMovie]

Today I got myself a powermeter and the results are 200W at bootup, 500W when folding. Wish I had the meter before I upgraded the drivers.
It´s nice to see that the new drivers keeps the load at 100%, no fluctuation as I had before. Increase of PPD seems to be 15%+, really nice.