Folding Forum

I have an i7-6900K with 64 GB RAM running Win10 off an SSD array, and one EVGA GTX 970.

I would like to further optimize my system to run f@h and I think my only two moves would be upgrading to more/better GPU and/or migrating to Linux.

But GPUs are so expensive now that anything more powerful than my current card is out of my budget. (Unless someone has a suggestion?)

Is the only move I have left beside a GPU upgrade migrating from Win10 to Linux? Thanks.

Optimising your system is a lot more than just upgrading hardware or swapping operating systems.

Here's some of the things I've learned:-

Storage
FAH is not storage intensive.
It does not need to be big or fast, just big enough and fast enough.
My best system is currently running on a 500GB OCZ Vertex-4, even that is very likely overkill but it's what I have.

You don't need a storage array, you're only wasting electricity.
If this is a dedicated folding rig then disconnect any other drives in the system. Again, they're drawing power for no use.

Memory
Make sure XMP is enabled.

How is your 64GB made up? 4 sticks or 8?
If it's 8 take half of them out. The 6900K only needs 4 sticks to enable quad channel memory, any more is a waste of electricity.

FAH needs only a few GB to run even the biggest work units.

Graphics Card
Download and install MSI afterburner.
This is both a monitoring and an adjustment tool.
I use it to reduce the power limit of my graphics cards so they run cooler and quieter.
I think best efficiency is somewhere abut the 50-70% mark. Your mileage may vary.

EDIT: Make sure your drivers are up to date. viewtopic.php?f=24&t=37391

CPU
Disable turbo, anytime your processor turbos it uses disproportionally more electricity than the the performance gain.
I actually go deep down into the Windows power management settings to limit the performance of the processor, 99% and below disables the turbo so there's no need to enter the bios to disable it.

FAH Client
By default the FAHClient reserves a single thread for each graphics card slot and assigns the rest to the CPU slot, the setting for this is '-1'.
The i7-6900K is a 8 core 16 thread processor. Change the CPU slot assignment to 13 or 14.
This is to give the operating system an "airgap" in which it can operate and it also gives the graphics card a bit more room to breath.
I've found that by allowing the client to assign all spare threads to the CPU slot actually causes problems, the computer is less responsive and it actively degrades the performance of the graphics card.
Anecdotally, the graphics card performance seemed to be degraded to a greater extent than the CPU was capable of producing.


Try playing around with some of what I have suggested.
It's relatively easy to do in Windows so you can identify the changes you need to make should you decide to take the leap and switch to Linux.

aetch wrote:FAH needs only a few GB to run even the biggest work units.

... currently.

aetch wrote: You don't need a storage array, you're only wasting electricity.
If this is a dedicated folding rig then disconnect any other drives in the system. Again, they're drawing power for no use.

Dual SSD?

aetch wrote: Make sure XMP is enabled.

Sticks are rated 2,400 MHz for use with XMP. I haven't tried clocking them higher tho.

aetch wrote: How is your 64GB made up? 4 sticks or 8?
If it's 8 take half of them out. The 6900K only needs 4 sticks to enable quad channel memory, any more is a waste of electricity.

4x 16

aetch wrote: FAH needs only a few GB to run even the biggest work units.

Correct, on my system at least. Win10 + F@H on 7 CPU cores and the GPU uses 5.1/64 GB. I guess it's "future proof" in case F@H ever needs serious memory.

aetch wrote: Download and install MSI afterburner.
This is both a monitoring and an adjustment tool.
I use it to reduce the power limit of my graphics cards so they run cooler and quieter.
I think best efficiency is somewhere abut the 50-70% mark. Your mileage may vary.

Will check this out.

aetch wrote: EDIT: Make sure your drivers are up to date. viewtopic.php?f=24&t=37391

Yes.

aetch wrote: Disable turbo, anytime your processor turbos it uses disproportionally more electricity than the the performance gain.
I actually go deep down into the Windows power management settings to limit the performance of the processor, 99% and below disables the turbo so there's no need to enter the bios to disable it.

Will check this out in both BIOS and Windows. Running -1 for the CPU, meaning 7 cores crunching and 1 feeding the GPU, all 100% usage, 3.7 GHz. Which is to be expected. Not sure when F@H would trigger turbo, but I don't use this system other than for F@H.

aetch wrote: By default the FAHClient reserves a single thread for each graphics card slot and assigns the rest to the CPU slot, the setting for this is '-1'.
The i7-6900K is a 8 core 16 thread processor. Change the CPU slot assignment to 13 or 14.

You weren't kidding... Going from 7 cores to 5 (I have HT turned off) still has four cores at nearly 100% usage and four cores at 75% or so usage with nothing but F@H running.

plector wrote:

aetch wrote: Make sure XMP is enabled.

Sticks are rated 2,400 MHz for use with XMP. I haven't tried clocking them higher tho.

By default XMP is disabled and your memory runs at 2133MHz.
If you open up the task manager -> performance tab -> Memory readout
On the right about halfway down should be a readout of your current memory speed.
2400MHz will mean XMP is enabled.

aetch wrote:

plector wrote:

aetch wrote: Make sure XMP is enabled.

Sticks are rated 2,400 MHz for use with XMP. I haven't tried clocking them higher tho.

By default XMP is disabled and your memory runs at 2133MHz.
If you open up the task manager -> performance tab -> Memory readout
On the right about halfway down should be a readout of your current memory speed.
2400MHz will mean XMP is enabled.

Yes, that's right. They are running at 2,400 MHz.

Neil-B wrote:

aetch wrote:FAH needs only a few GB to run even the biggest work units.

... currently.

Is there reason to believe that WU will become larger or somehow else require much more memory? I think one of the two crunching tasks eats about 240-260 MB of RAM, the other much less so. EDIT: I have never seen more than 1/2 GB usage on the GPU during F@H either, 4 GB total, so the GPU RAM capacity is not (yet) an issue.

Mm comes in different forms and different challenges .. as models get larger and incorporate more factors it is easy to think of scenarios where very large amounts of ram might be needed to support the gpu .. whether that will happen is anyone's guess .. I was simply pointing out that things can and do change.

In the past when the compute power was more limited scenarios that might use large amounts of ram were fah (sorry couldn't resist the pun) less possible but with the power of the current crop of fast/wide gpus such as the rtx 3000 series the possibility of running such models and needing larger ram support and delivering novel scientific insights becomes more feasible .. of course this could/would only happen if there was there was enough kit available that could actually run the models.

Actually with a potential pool of gamer/enthusiasts with rtx 3000 series and 32, 64 or even 128GB rigs the opportunity for some researcher to push the modeling boundaries further might be here?

Neil-B wrote:Actually with a potential pool of gamer/enthusiasts with rtx 3000 series and 32, 64 or even 128GB rigs the opportunity for some researcher to push the modeling boundaries further might be here?

The fah dev team and the science people talk amongst themselves and perhaps with one another about this once in a while, right? It would be kind of cool if something could use the massive 64 GB RAM I have. As I mentioned above, Win10 Pro 21H2 + fah on a 6900K and a GTX 970 uses about 5.1 GB out of 64 total. I do not use it as my primary system and generally don't do video editing, CGI, gaming, etc.

My gut instinct (and I am not a scientist/researcher) is that use of lots of ram will probably be paired with the latest fastest/widest gpus if it were ever to happen ... the factors that might utilise this also would want loads of really fast compute cores ... my total guess would be that even rtx2000 series could possibly struggle with the loadings ... but this is speculation on my part ... if something like this were to develop (and I have no idea if it ever will) it would be the needs/desires of the researcher and the proposed modelling that would shape the minimum hardware requirement - there would pretty much always be a trade of between being able to find enough of the very best resource (say rtx3000 series or better with 128GB ram which might be very scarce) and doing the most extreme modelling or paring back the model a bit to allow "lesser" setups to be usable - who knows?

I am however treating myself to an upgrade over the festive season and pushing my best kit (last one in my sig block) to 128GB anyway for other reasons ... I have accepted that the 512GB in my cpu servers won't even get utilised for FaH !!

plector wrote:

aetch wrote: By default the FAHClient reserves a single thread for each graphics card slot and assigns the rest to the CPU slot, the setting for this is '-1'.
The i7-6900K is a 8 core 16 thread processor. Change the CPU slot assignment to 13 or 14.

You weren't kidding... Going from 7 cores to 5 (I have HT turned off) still has four cores at nearly 100% usage and four cores at 75% or so usage with nothing but F@H running.

With 4 CPU cores, 4 cores are between 90-100% and the other four are between 60-75% usage. CPU PPD went from 145k - 155k @ 7 cores to 105k @ 5 cores to 80k @ 4 cores. GPU went from 450k @ 7 CPU cores to 545k @ 4 and 5 CPU cores. May have hit a wall for maximum GPU performance with three free cores but I will continue to watch. Also waiting for these WU to finish before I look at some of your other recommendations. EDIT: So I guess that means each 6900K core can push 20-21k PPD each and that the GPU needs three free CPU cores for its max performance. So I traded 70k PPD CPU for 100K PPD GPU.