GPU folding box?

edited January 2009 in Folding@Home
okay, so i'm getting a bit addicted to this folding business. :) would it make long term sense to buy a twin 8800 box with a dual core right now? 9k ppd am i correct?

will the 8800 still be able to get 4k ppd a year from now? and are ATI cards going to start to see the kind of performance people get from nvidias anytime soon?

Comments

  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited December 2008
    Wow, I just love this kind of project!

    With a slight shader overclock, G92-based video cards, 8800 and 9800 series (NOT original 8800GTX, which is G80) can get 6,000PPD. I've had to GPU clients run slightly over 6,000PPD just today. For a while, the 57XX series work units were outputting around 3500 (depending on shader clock). The newer series of projects are much more productive.
    ATI cards going to start to see the kind of performance people get from nvidias anytime soon?
    I really doubt it. That would require a nearly complete architectural revision at the GPU nano scale. Just recently ATI GPUs have seen a significant Folding performance increase due to driver revisions. There is still headroom left in coding, but nothing that I know of that will put ATI in the same arena with Nvidia with respect to Folding for probably at least six months to a year, probably the latter.

    Please, ATI fanboys, don't take it as a slam. ATI was the first to support GPU Folding@Home and were pioneers in this field of processing. For now though, they have been leapfrogged by Nvidia.
    will the 8800 still be able to get 4k ppd a year from now
    I don't think anyone can predict drivers performance and Pande Team Folding projects that far into the future.
    so i'm getting a bit addicted to this folding business
    It's a good addiction! I got hooked nearly 8 years ago.
  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited December 2008
    With a Q6600 quad class CPU or better running Windows SMP Folding plus two G92s processing GPU clients, 14,000PPD+ is attainable for a single machine. The best I've seen is a couple really dedicated guys with rigs running three or four 9800GX2s - that's six or eight GPU clients per computer!

    The attachment shows two GPUs, each one core of 9800GX2 card, performing at better 6K PPD.
  • edited December 2008
    I'm switching all my dedicated rigs to Ubuntu 8.04.1 because of the added SMP performance. GPU2 is tricky to set up, but can be done, and the boost to SMP is quite impressive (literally double what you would achieve under Windows). For example, my E2200 @ 3GHz running alongside a GPU2 client went from 1000PPD under Windows Vista to 2100PPD under Ubuntu and an overclocked C2Q can hit >5500PPD.

    Just throwing that out there. :D
  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited December 2008
    That's amazing.
  • SnarkasmSnarkasm Madison, WI
    edited December 2008
    OSX also sees a nice bump. Everybody seems to forget that the Windows clients are essentially ports. Everything runs better on the *nix side.

    My old laptop was too slow to run SMP in Windows, so I put 8.04 on it and just let it churn. It only got 700ppd or so, but it's way better than nothing.
  • edited December 2008
    hmm. got some questions. is the 9600GSO usable for the next 6 months or so? i heard its having trouble coping with the new WUs. also, should i go dual core or quad core? amd or intel?
  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited December 2008
    I've not been reading about the 9600GSO having problems. I'm a little bored right now, so I'll take a look.

    Do you have any links concerning such problems? I'm not challenging you, I just want help steer you in the right direction.
  • _k_k P-Town, Texas
    edited December 2008
    Right now a 9600GSO is hitting for 3300ppd with 600/1576/800 clock on it. They will scale up higher but my old 939 board has a sucky choke buzz when I run certain clocks; overall 2 9600GSOs and a 939 x2 4200+(stock) turn out about 7k+ when they are running. The 4500ppd GSO is kind of a myth unless you can run some crazy OC like 675/1800+/975, in my experience and opinion.

    The 9600GSO is alright for just farming out if you can find a good deal on them. When in contrast 9600GSO xfx card(don't like their ref boards) is $78 while a new 9800GTX is $150. It is double the price for double the production. Then if you are looking at that you can find 9800GX2s for 300 or less and that is a 10-12k ppd card by itself. This all really depends on if you have the power supply for it and how much money you want to sink into this.

    The 9600GSO out performs all ati cards excluding a x2 and cheap for the ppd produced. They are not doing so well on the newer WUs and since the newest set has hit the cards my 8800 GTs' are pushing up to 5700; they can be had shipped for roughly 120ish. Over the years the 88s have held up. After buying two 9600GSOs for 180 shipped, freaking had to pay for 3-saturday delivery was a sweet afternoon project though, I kind of wish I had just bought a 9800GTX so I could save a few dollars now and have another spare video slot to put in another GTX later.

    Quad-intel. For all instances. With the current cuda drivers you do not load CPU cores in order to run GPU2 clients which allows you to run SMP on all cores and run dual SMPs if you want to. Intel is just faster and worth the price.
  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited December 2008
    You might want to consider looking for a used 8800GT or 9800GT. They both will outperform any 9600 and will cost about the same used as compared to a new 9600. I just sold a 9800GT three days ago for $85, shipped. The big advantage that 8800/9800 have over over the 9600 is stream processors count, which is 112 and 96 respectively. That is the most important GPU factor for Nvidia GPU Folding. Core frequency and GDDR play minor roles in comparison.
  • edited January 2009
    sorry leo, i didnt mean errors. i meant the ppd drop after the new nvidia WUs and yes, the reason mentioned was the same, the stream proc count.

    my 4850 can do 4.2k ppd. the gap between nvidia and ati seems to be closing somewhat. from the figures i see around the net, anyway.

    i am half considering waiting for the next generation of cheap nvidias to start this folding box project actually. if you ask me, taking a previous gen gpu and just changing the 8 to a 9 is a bloody con job.

    i was thinking of a pure GPU folding box, 2 9600gsos or 9800gt's because if i can just drop an amd athlon x2 in there id save a bunch of money. but that quad intel really seems tempting. :) how much ram do you think id need for this endeavour?
  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited January 2009
    i meant the ppd drop after the new nvidia WUs
    As a general rule, that applies to all video cards/GPUs. Newer, more demanding work units come out to take advantage of new, more powerful GPUs, points 'go down' for less productive hardware, then Pande Group co-develops new GPU and CPU cores to take advantage of newer hardware. It's a never ending cycle, no different than gaming, multimedia editing, or any other processing-intense software/hardware cycles.
    how much ram do you think id need for this endeavour?
    Ok, right now I'm looking at Task Manager for a system running a quad CPU with SMP and two GPU2 clients. Not much else is running on that system except for hardware monitors, antivirus, anti-malware, and an idle web browser. Current memory demands are 1.11GB. The system is a Windows XP SP3 platform. I don't know about Vista's or Linux' memory requirements.
    i am half considering waiting for the next generation of cheap nvidias
    That generation is here, now. It's 8800GT. OK, I understand I can't speak to your market, but here in North America prices are falling on G92 based GPUs. I'm very active in the pre-owned market and I'm consistently seeing 8800GTs for $70-85 shipped. That is tremendous performance per price. The next big bang for buck will be in about nine months, maybe six, when the Nvidia 2XX series are on the used market in quantity.

    ATI is a wild card...pardon the pun. The high end Nvidia and ATI offerings essentially are at parity in the gaming arena, just based on which games the consumer prefers. I don't see ATI catching up with Nvidia in Folding for a long time due to significant nano-scale architecture differences.
  • edited January 2009
    i am half considering waiting for the next generation of cheap nvidias to start this folding box project actually. if you ask me, taking a previous gen gpu and just changing the 8 to a 9 is a bloody con job.

    So, you'll be happy to know that Nvidia is going to be re-branding G92 again to the GTS 200 series. :D
  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited January 2009
    Oh, good grief! 'If you can't impress them with brilliance, dazzle them with bullsh!t.'

    To make a long story short, if the story turns out to be accurate, Nvidia will be doing with the "240" and "250" what the original intent was to be with the 9800 series. That is, the 9800 GPUs were to be the 55nm die shrink of the 8800 series 65nm die. Well, Nvidia fumbled that one, but went ahead anyway and renamed the GPUs. Maybe they actually will get it right this time.
  • edited January 2009
    Ive decided Nvidia are idiots, and also that GPU folding is just not a sound decision financially right now. decided to build a SMP box for under 400USD with a 6600 and a G31. Anyone know how this chipset does OCing? and also, has anyone had any experience running notfreds USB folding disk in dual SMP? how would i set affinity for this?
  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited January 2009
    notfreds USB folding disk
    Never heard of it. I believe I'll take a look. There is always more to learn!

    You don't need to do any manual settings changes for affinities. Just download and install Affinity Changer for dual client SMP Folding. Affinity Changer is load balancing tool that keeps the WinSMP requirement shifting to the different cores that are under utilized. (It's not necessary, but production will increase significantly.) If you are really adventurous, you could run two SMP clients under a virtual Linux layer within Windows - VMWare. You'd see a 30-40% better production that than directly running on Windows. On that topic, I know about it but don't know how to do it. But with multi-client SMP Folding, Linux and OSX are the kings.

    Sorry the GPU Folding didn't work out. One of the nice things about the FAH project is that there are clients for just about any hardware/software configuration around.
  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited January 2009
    Notfred's: Here

    That looks interesting. Maybe we were writing about the same thing. I believe I'll give that a try!

    Magic, I'm working on it (VMWAre + Notrfreds) now on another Windows platform. I'll report what I find in a new thread. Thanks for getting my interest perked. I had not realized that there was an easy route to VMWare virtual machine Folding.
  • edited January 2009
    i dont plan on installing windows in the folding box. i initially planned on running that notfred linux straight from a USB flash drive. but i am curious as to how i would install drivers for wireless network cards and such, seeing how its so stripped down. do you know how to do this?
  • LeonardoLeonardo Wake up and smell the glaciers Eagle River, Alaska
    edited January 2009
    No clue. I am completely new to Linux. My last couple hours experimenting with USB-VMware-Notfreds is my ONLY Linux experience. I've got Notfreds/SMP running now, but I haven't seen any completed frames.
Sign In or Register to comment.