digg

Facebook Connect Join Digg About
See the original image at extremetech.com

Not just for Games Anymore, The Present and Future of GP-GPU

extremetech.com — It wasn't so long ago that 3D graphics cards were only expected to deliver higher frames-per-second in your favorite 3D games. But now we are about to enter the age of "GP-GPU," general-purpose computing on a GPU, and it's about ready for the mainstream. Here's some of what you can look forward to.

Who dugg this? Made popular Jul 2, 2008

Digg DudeWhat is Digg?

Digg is a place for people to discover and share content from anywhere on the web. Digg surfaces the best stuff as submitted and voted on by the community. Take the Tour to learn more.

49 Comments

show profanity settings
expand all
  • MrIso, on 07/02/2008, -2/+11OpenCL: http://en.wikipedia.org/wiki/OpenCL
  • Kamael, on 07/02/2008, -3/+10Too bad there are again multiple standards working against the consumers. I've read somewhere that Nvidia was in talks with AMD to bring CUDA to the ATI cards, but I don't have much hope for that.

    We need an OPEN standard for GPGPU and another for Physics on the GPU. And no, DirectX doesn't count.
  • macbookhair, on 07/02/2008, -0/+5Imagine what my old 56k modem could do!!!!
  • mrBitch, on 07/02/2008, -1/+6Also, see OpenCL support in Snow Leopard.
  • Ramble, on 07/02/2008, -0/+5Psst. No it's not.
  • mrBitch, on 07/02/2008, -1/+6OpenCL (not OpenGL).
  • inactive, on 07/02/2008, -0/+3Thanks, "DickMasterson"
  • inactive, on 07/02/2008, -3/+6"In other words, the average GPU client is four times faster than a PS3. That includes relatively older and low-end graphics cards, too. Newer cards are likely six to eight times faster."

    Woah, dang. Consoles are really starting to show their age now, eh?
  • e2superman, on 07/02/2008, -0/+3Ok so I code with CUDA and am doing IRAD work on it for one of the big 4 Defense companies in the US. While CUDA (NVidia's API to execute code on their GPU) is great it has many fallbacks. GPU's are extremely parallel which is where they get their speed from. Unless you have code that can benefit from parallel operations then CUDA is not generally a benefit. Also there is the overhead involved in using GPU's. Specifically you have to copy your data from your main memory over the PCI-E bus to the card, process it, and move it back. VERY slow. The PCI-E bus is in the low gigabyte per second speed category versus the speed from main memory to the CPU. For example if I run a ton of FFT's on my graphics card (batch) there is something like a 50% overhead (wasted time) in just moving them to the card.

    Anyhow I could go on but needless to say this is not yet the holy grail. THere are certainly things to be excited about but you need to understand there are definate restrictions. Namely the PCI-e bus bottleneck. That said we are hoping to replace million dollar equipment with 30k worth of equipment using CUDA so we believe there may be some merit if your code base and ussage can leverage it.
  • macbookhair, on 07/02/2008, -0/+3As a gamer.. I have always spent about 2x more on my graphics card.
  • sg7791, on 07/02/2008, -0/+2There we go, measuring with consoles again.
  • tomz17, on 07/02/2008, -0/+2That's a TON of overhead from my experience (especially for larger array sizes). Are you sure your CUDA code is properly optimized? Are you transferring intermediate results?

    I have an algorithm that is pretty much entirely FFT's, where i've seen speedups on the order of hundreds over just running it on the CPU?

  • yabos, on 07/03/2008, -0/+2NO, it's Open Compute Library(OpenCL) for doing GPU calculations in regular programs.
  • kolyana, on 07/02/2008, -1/+3no, i don't think he did.
  • Anonymous3, on 07/02/2008, -0/+2Even the ones that play Dragon Ball Z: Burst Limit?
  • Ratteler, on 07/02/2008, -0/+2/ me thinks back to Amiga's custom chips.
  • supermanred, on 07/03/2008, -0/+2Can't wait. I can say my macbook utilizes both cpus a lot better in OS X than it does with Vista, it also uses my 4gb of ram better... I can only imagine how fast the bitch would run if it was using the cpu power of the graphics card as well....

    can't wait for snow leopard.
  • e2superman, on 07/03/2008, -0/+2Try this. Send some 1Million length complex FFT's (or whatever power of two or prime near that size) and batch them say to send 10-20 at a time. See how long it takes. Then go back and just copy the data to the card (and do not do the FFT on the card) and then copy back. The delta is the actual processing time (more or less) and that also tells you your overhead in piping the data. Post results here if you can. Thanks!
  • Voodrake, on 07/02/2008, -0/+2What the Frak Dick? :)
  • freshgrease, on 07/02/2008, -0/+2Holy buttmunch, Batman! Just think what you could do with a tri-SLI Nvidia 280 GTXs, dual 3.2 Quad-core Xeons, and loads of RAM! I personally would use it for my renderfarm to replace a couple of machines, but definitely throwing Folding@Home on it for its off time.
  • blackjack75, on 07/02/2008, -0/+2So my Radeon X1600 is going to give em PS3-like graphics! Wonderful. I believe it. Kind of.
  • justaboutdead, on 07/02/2008, -0/+2All in all... this was a well written article that brought up some very valid points. What is the market going to look like over the next couple of years? I don't know, but I can only hope that it means top-notch performance, cross compatible, and at a decent price.

    dugg
  • supermanred, on 07/03/2008, -0/+1Yeah, not gonna happen any time soon. Reboot into Windows XP and play your games then reboot back into OS X for stable computing.

    Sorry, it's reality.
  • yabos, on 07/02/2008, -1/+2This is going to be huge in the next few years. Current high end GPUs do 500GFlops. Next up 1TFlop on a single GPU. OpenCL is going to kick ass if it gets adopted by everyone.
  • DickMasterson2, on 07/02/2008, -0/+1Now you are a fag for mentioning that!
  • Ramble, on 07/02/2008, -1/+2Nothing will happen until AMD, nVidia and Intel decide to unify across their cards. Preferably using OpenCL.
  • supermanred, on 07/03/2008, -0/+1I know the Mac isn't a gaming platform, but some high end Macs do have nice graphics cards in them and the next version of OS X (Snow leopard) will have this technology built in and allow you to utilize your graphics cpu for gen purpose computing... kind of neat. Hope it's built into the next version of Windows (you know, the one after XP) :)
  • inactive, on 07/02/2008, -3/+4thumb me down !
  • Schmich, on 07/02/2008, -0/+1You mean think of the price? I rather wait and spend money on the best bang for the buck card than going all out.
  • supermanred, on 07/03/2008, -0/+1Mine drove me insane.

    Actually, it was my 300 b modem that drove me nuts. The 56 kb modem finished the job off.

    ***** im such a nerd.
  • themastersb, on 07/02/2008, -0/+1Nvidia has made some large strides over the past couple years and are starting to almost become unrivaled with their graphics cards which is sort of a bad thing since they need more pressure from competitive companies to keep making high-quality and better products. (or so I assume)
  • supermanred, on 07/03/2008, -0/+1Of course they recommend splitting budget between cpu and graphics card. They want half the gamer money.

    ***** them all, I bought a 360 and wont upgrade until the next gen of consoles.

    Yeah Yeah I know... Just cant afford to be a pc gamer any more and I can still play games for only a few hundred bucks (console) and they look great on my hdtv.... pc gaming rocks, just sayin........

  • mrBitch, on 07/03/2008, -0/+1True, Snow Leopard will be the first OS upgrade that will :

    1. use LESS RAM
    2. be FASTER than the previous OS
    3. have LOWER min. requirements to run than prev OS

    I wonder if Windows 7 will be out by June 2009?
  • e2superman, on 07/02/2008, -0/+1You need roughly one CPU to GPU to manage it. Also the PCI-E bus goes to hell when you are using CUDA which can hurt any RAID cards you have attached on the PCI-E bus. Just sayig, there are still issues to overcome.
  • inactive, on 07/03/2008, -0/+1This may be one of the few times that I wish to bury myself for stupidity of not RTFA before commenting...
  • e2superman, on 07/02/2008, -0/+1Current High End GPU's do a TeraFlop however getting a "usable" TF is nill. We see far less "actual" performance due to latency and overhead.
  • inactive, on 07/02/2008, -1/+1When is GPU hardware acceleration coming to VM?
  • Depravo, on 07/02/2008, -0/+0No, you're thinking of mens bottoms.
  • inactive, on 07/02/2008, -0/+0it's sarcasm. the bad kind i guess
  • medfreak, on 07/02/2008, -3/+3The prospects are virtually infinite. It is going to be exciting watching how different pieces of hardware we took for granted on the PC are going to embrace new roles away from the traditional though in the coming years.
  • Baryn, on 07/02/2008, -1/+1I remember when the GeForce came out, the "NV10", and invented the term "GPU"...

    Never thought it would stick! *****!
  • DickMasterson2, on 07/02/2008, -4/+1No I meant X-bawks' is for fags, sorry to mix you guys up. The guys that play PS3 are alright!
  • inactive, on 07/02/2008, -5/+1oops - I believe you meant, "but can it run run Doom?"
  • tito13kfm, on 07/02/2008, -6/+1Badaboom is gonna rule
  • mjesales, on 07/02/2008, -7/+2nice concept. Lets see how long it takes to go. If you look at nvidia's site. They recomend splitting your budget between CPU and graphics card.
  • iSkylla, on 07/02/2008, -9/+3But can it run Crysis on Ultra High?
  • inactive, on 07/02/2008, -6/+1Psst, it's OpenGL.
  • inactive, on 07/02/2008, -7/+0OMG like intel is going DOWN

  • Add a Comment

    Login or register to comment!
    Or, sign-in super fast with your Facebook account ...


© Digg Inc. 2009 — Content created and posted by Digg users is dedicated to the public domain | Terms of Use Updated | Privacy Policy
DIGG, DIGG IT, DUGG, DIGG THIS, Digg graphics, logos, designs, page headers, button icons, scripts, and other service names are the trademarks of Digg Inc.