How to get the most Folding Points from your Hardware

[Lizard]

All credit for this thread should go to babychunder, as he set up the original thread on the old Custom PC forum - I'm simply recreating it here for continuity.

---------------------------------

I thought I'd jot down a few tips to pass on what I've learned, so that new / existing folders can squeeze those last few points out of their rig. Might not be 100% accurate so please feel free to improve the thread.

Best Clients

1. Windows GPU2 console client
2. Windows GPU2 systray client
3. SMP client (Native Linux/OSX)
4. SMP client (Windows with Linux under VMWare)
5. SMP client (Windows)
6. GPU no-nonsense console client (Windows)
7. GPU GUI client (Windows)
8. No-nonsense console client (Windows or Linux)


Best Operating System - all clients except GPU2

1. Linux/OSX
2. Windows with Linux under VMWare
3. Windows

Best Operating System - GPU2 client

1. Vista (uses less CPU)
2. Windows
3. Native Linux under WINE

Best GPU Hardware

1. 9800GX2 / 8800GX2
2. GTX280
3. GTX260
4. 9800GTX+ / 8800GTX+
5. 9800GTX / 8800GTX
6. 9800GT / 8800GT
7. 8800GS
8. 9600GSO
9. 9600GT / 8600GT

I have only listed nvidia hardware because ATI's isn't that good. But if you already have a top-end ATI card you will probably get more PPD than a low-end nvidia. Just don't buy ATI if you want to fold.

Best CPU Hardware

1. Quad Core 45nm Q9xxx
2. Quad Core 45nm Xeon E54xx
3. Quad Core 65nm Xeon E53xx
4. Quad Core 65nm Q6xxx
5. Dual Core 2 45nm E8xxx
6. Dual Core Woodcrest Xeons E51xx
7. Dual Core 2 65nm E6xxx
8. AMD Quad Core 9xxx
9. AMD Tri Core
10. Dual Core AMD Opteron
11. Core Duo Mobile
12. AMD 64 x2
13. Core Solo Mobile
14. Pentium M (Dothan) with 2mb cache
15. Pentium M (Banias) with 1mb cache
16. Single Core AMD 64
17. Older Pentium 4 Netburst Architecture, larger cache sizes are best
18. Older Pentium Mobile
19. Celerys and Simper-ons

Obviously the higher the clock speed the better. After that consider the FSB speed, higher the better. Some CPUs with the same clock but higher FSB are cheaper anyway. Last CPUs with larger cache will fold quicker. There will be some overlaps between processor classes, for example a high-end DC Opteron will smoke a low-end Core 2. Similarly, 2x Dual Core Woodcrests will outperform a single Core 2 Quad because of the extra cache and the higher memory bandwidth of the duallie motherboard. Pretty much all of these CPUs can be overclocked; however the Xeons are for experts only. If your mobo allows you to clock up an E2160 to faster than an E6750, then its worth considering despite the small CPU cache.

Best Memory Configuration

1. Four FB-DIMMs in Quad Channel mode
2. Two DDR2s in Dual Channel mode
3. Two DDRs in Dual Channel mode
4. Four DDR2s Dual Channel
5. Four DDRs Dual Channel
6. One or three DDR2s
7. One or three DDRs

In all cases you should have at least 1GB (2x 512) of memory. If running VMWare, you will need at least 2GB (2x1GB). With Quad Cores, shoot for 4GB if possible.

Latency matters, the lower the better, but it’s less important than clock speed. DDR2-800 with CAS4 is the minimum you should shoot for if it’s a new build. DDR2-1066 or better if you can afford it.

Best Client Mix

1. As many GPU2 clients as you can get GPUs in a box
2. 2x Quad Core Xeons – Four SMP clients
3. 1x Quad Core C2D – Two SMP clients
4. 2x Dual Core Xeons - Two SMP clients
5. High-end Dual Core machines – One SMP client
6. All other Dual Core machines – One SMP client
7. High-end Single Core machines – One Console client

The above assumes non-Windows SMP clients. You should only consider the Windows SMP client if you can’t run Linux or get VMWare working, or don’t have enough RAM to support VMWare.

When using mixed clients always use "affinity changer" (choice of three tools) to do SMP in pairs and have a dedicated core per GPU. Google for details. Alternatively just right click on the folding processes in Task Manager to set affinity.

Never mind that Stanford recommend four cores per CPU client, it runs fine on two. Two clients gives better PPD. Your low end machine (Celery, Simperon etc) still has value and is best used as a mule for a GPU2 client.

Look out for electricity cost. Any of the Mobile chips on a Desktop mobo will save you money. After that, the Core architecture is best, followed by the AMD, with Intel P4/Xeon Netburst bringing up the rear. The best Points Per Day per Watt comes from Xeons and the Quads; the worst is from the P4 Netburst and Celeri. GPU2 client PPD/W is much better than SMP; a quad will pull at least 300W for about 2500PPD while a GPU about 100W for about 5000PPD (assuming an 8800GT).

Last if you are overclocking, don't push it too far. An unstable OC will cost you points and waste Stanfords time. You may think you have a stable overclock using Prime, Orthos etc but folding is much more demanding and will always find any weakness or instability.

[Lizard]

All credit for this post should go to IwantAbetterPC, as he set up the original thread on the old Custom PC forum - I'm simply recreating it here for continuity.

-----------------------------------

Hi all,

I managed to get both cores of a 4870 X2 folding under XP, so thought it might be worth posting how I managed to do it.

I initially installed the Catalyst Control Centre (CCC) and drivers off the CD with the card, then updated to 9.2 from the web, but it wouldn't fold on either core, with constant "Unstable Machine" errors.

So, I went onto the ATI Radeon site and downloaded just the 9.2 driver package to the desktop - not the CCC stuff. I then un-installed all of the CCC software and the drivers, and re-installed just the 9.2 drivers.

GPU console clients were set up with two shortcuts pointing to two different work folders, with the following flags added to each client

-gpu 0 -forcegpu ati_r700 -verbosity 9 -local
-gpu 1 -forcegpu ati_r700 -verbosity 9 -local

and this time both started cleanly and are clocking away on both cores at 95-100%

The only downside is that without the CCC software, you can't configure or overclock the card, but GPU-Z is showing it as CrossfireX enabled still, so unless you want to mess about with it, this will at least get it folding on both cores for about 3,800ppd at stock clocks. My PC is pretty old, and has PCI-E 1, so this may be constraining the performance somewhat, plus the CPU chip isn't great, and as the ATI GPU core still needs a whole CPU core to itself to perform at all (unlike the nvidia cards), this is another bottleneck for me.

Hope this helps!

Craig...

[leexgx]

hmm 8800gx2 :P

missing {GTX295 (xp only) }as that should be at the top (but have to use XP or second half of the card cant be used)

[SwiftDestiny101]

Originally Posted by leexgx

(but have to use XP or second half of the card cant be used)

Just put a dummy plug into the second card and tell it to extend the display to that monitor and you can fold on the second card under Vista.

More here https://www.evga.com/forums/tm.asp?m...5&mpage=1&key=

[leexgx]

id use XP 32 if i was(is) folding (well useing XP/2003 x64 but it locks up some times with Thread stuck in driver and i cant update the drivers as the new 180 drivers do not work in my 2 rigs only 178.24)

[coolamasta]

Does anyone know what a GTX 275 / 285 produce in PPD?

[Tim S]

Originally Posted by coolamasta

Does anyone know what a GTX 275 / 285 produce in PPD?

Right here

[coolamasta]

Originally Posted by Tim S

Right here

Awsome, nice one mate

GTX 260 - 6500 ppd-ish
GTX 275 - 7000 ppd-ish
GTX 285 - 7500 ppd-ish

[cgcox1]

I think XP is now definatly better than Vista for the GPU2 clients. With the latest drivers since this thread was originally written.

[SazBard]

can anyone give some advice on how to setup the config files of the cpu and gpu for the best folding?

What extra command line arguments should I be using?

Thanks.

[cgcox1]

Try -gpu 0 -local -forceasm -advmethods -forcegpu nvidia_g80 -verbosity 9

for a GPU2 Nvidia graphics card.

-local -forceasm -advmethods -smp -verbosity 9 for an smp client

and

-local -forceasm -advmethods -verbosity 9 for a 'standard' client.

the -forceasm may now be out of date, i'm not sure; but has done me any harm having it in there.

As for the setup in the config files, well let me know the question and i'll tell you what i think is best.

[SazBard]

cool. thanks, although -configonly moans because the line is too long and cuts verbosity in half =/ but i got rid of advmethods because it asked me that already.

Did i explain that properly?

[uncle_fungus]

-advmethods is redundant as that can be specified as a client option anyway.
-local is deprecated and does nothing at all in the v6 clients. It was only kept to prevent drop-in replacements from dying with unsupported parameter errors.
-forceasm does nothing for GPU2 clients as they don't have and assembly optimisations to force on.

[SazBard]

Originally Posted by uncle_fungus

-advmethods is redundant as that can be specified as a client option anyway.
-local is deprecated and does nothing at all in the v6 clients. It was only kept to prevent drop-in replacements from dying with unsupported parameter errors.
-forceasm does nothing for GPU2 clients as they don't have and assembly optimisations to force on.

OK so I should use: -gpu 0 -forcegpu nvidia_g80 -verbosity 9?

Also gpuz says my gfx card is g92, should I replace g80 with g92?

[uncle_fungus]

Originally Posted by SazBard

OK so I should use: -gpu 0 -forcegpu nvidia_g80 -verbosity 9?

Yes

Originally Posted by SazBard

Also gpuz says my gfx card is g92, should I replace g80 with g92?

No, there is no g92 option, only g80.

[SazBard]

Regarding VMWare and Notfreds diskless folding client, is there a way to set cpu usage percentage?

Thanks.

EDIT: Mods, why does my folding sig keep disappearing?? I have to go into usercp and save my signature over and over again for it to reappear.

[Deadpunkdave]

Since it fits nicely in this thread, I'll ask here:

Could someone clarify for me the ideal graphics card clock settings? I know that OCing the shader clock is the primary concern but should memory and core clocks be left alone or underclocked?

Cheers

[Christopher N. Lew]

Originally Posted by Deadpunkdave

I know that OCing the shader clock is the primary concern but should memory and core clocks be left alone or underclocked?

You could try underclocking core and memory so that they produce less heat, which is often the limiting factor in overclocking. But increasing the shader clock often does not change the heat profile, so there will be little benefit from the underclocks. It'll vary from individual card to card (which you have no control over), and how well your particular system is cooled.

IIRC (can't find the link) people have reported more success with reducing core and memory, and increasing shaders, with factory overclcocked cards, than with standard cards. Quality control at manufacturers?

[Deadpunkdave]

Originally Posted by Christopher N. Lew

You could try underclocking core and memory so that they produce less heat, which is often the limiting factor in overclocking. But increasing the shader clock often does not change the heat profile, so there will be little benefit from the underclocks. It'll vary from individual card to card (which you have no control over), and how well your particular system is cooled.

IIRC (can't find the link) people have reported more success with reducing core and memory, and increasing shaders, with factory overclcocked cards, than with standard cards. Quality control at manufacturers?

Thanks Christopher, top man, next time I get paid I'll be getting 4 Zalman VF1000 2 Orbs, fingers crossed I should be able to drag another card's worth of ppd out of my rig.

[saspro]

VMWARE SMP Guide is here