Four GPU Motherboard

didde writes "The people over at Tom's Hardware are running a story on Gigabytes experiments with quadruple GPU's on one motherboard. Perhaps we'll need something cooler than liquid metal to keep this beast from running hot?" From the article: "About half a year ago, we learned that Gigabyte was working on a graphics card that integrates two GeForce 6600GT graphics chips. While we were impressed with the out-of-the-box approach from Gigabyte, there was of course the question, whether two of those cards could be combined for a total of four graphics chips."

Comments

Limitations

[taskforce]

One of the major limitations of the GB Dual GPU cards is that they only worked on their propreitary motherboards, which is useless for people who use other brands of motherboards; this was supposedly because it was using the SLi in some strange way. (2 SLi links accross the GPUs as opposed to 2 boards)

I would hope that they would be able to get these to run on all SLi boards, I've always thought one of the main strengths of building your own PC was the compatibility between differnet brands of components.

Re:Quad Cards?

[taskforce]

This is 2 cards with 2 GPUs on them each, not 4 cards. Last year Gigabyte launched their dual GPU cards, but they couldn't run in SLi. At the time one of the main comments from reviewers and fans who were shocked by the power was "Whoa, wouldn't it be cool to run 2 of those in SLi and have 4 GPUs!"

Re:So...

[matth1jd]

Other than Opteron server boards with HT slots, where is a motherboard that could hook in two grpahic cards?
Well here's the list from NewEgg.
SLI Equipped Motherboards.

Re:Quad Cards?

[Thinko]

Actually, it's both.. It discusses running in 4- 8x PCI-Express configuration for 4 Single Chip cards, and 2- 16x PCI-Express configuration for 2 Dual-Chip cards.

Re:Could someone please explain how this works?

[kebes]

The 4 GPUs are on two dual-core cards. You could use this in an SLI setup to run a single monitor with ridiculous amounts of graphics power, or two monitors with still amazing graphics rendering, or more monitors if you wanted to, I suppose.

SLI is Scalable Link Interface. It's a way to have two video cards running a single display. If, for instance, you have a video game with really high graphics requirements, but you don't want your frames-per-second (fps) to drop, then you could use the two graphics cards to render alternating frames. That way, you have high frame rate combined with the best graphics. In theory you can double the graphics complexity of whatever you are trying to render. In practice, of course, it can be hard to get it running, and for many games/applications won't make any difference whatsoever. It's still a very much "power gamer" setup, only for people who (1) have the money, (2) like tinkering, (3) enjoy being "bleeding edge" just for the heck of it, (4) really like their games to look slick... at any cost!

Despite the fact that SLI is currently seen to be sorta frivolous by many, it's quite possible that SLI (or multi-GPU cards) will become common in the future, and will in fact be required to play modern games.

Re:My God!

[psetzer]

In GPUs, they don't have multiple cores, per se, but they do have multiple rendering pipelines. Modern graphics cards can have anywhere from 1-16 pipelines, with each pipeline doing one pixel at a time, as long as RAM is sufficient. Each pipeline is like a core, but not exactly; a core can run its own instructions on its own data, while a pipeline runs the same instructions as all the other pipelines do. In computer architecture, a plain old CPU is single-instruction-single-data (SISD), a multicore chip is multi-instruction-multi-data (MIMD), and a GPU is single-instruction-multi-data (SIMD). Mixing things up is the fact that after the Pentium MMX, all Pentiums are capable of doing SIMD on integer arrays, but can't do it on floating point numbers. (In fact, IIRC they can't even do regular floating-point math at the same time they are using the SIMD instructions)

Multiple pipelines at a time allows you to increase the rendering speed almost linearly, as long as you accept the trivial restriction that you must get the same image as an output no matter what order you render the pixels in. It's the opposite of the CPU business. In CPUs, they started adding multiple chips first, and in the GPU side of things, they added multiple cores (or their equivalent) first. This is partly because it's easier to only decode one instruction at a time and send the decoded signal to every pipeline than it is to decode multiple instructions and send them to the correct cores. This isn't to say that it's impossible, but a couple million more transistors are enough to make you think twice about whether you really want two cores.

Re:My God!

[eofpi]

Mixing things up is the fact that after the Pentium MMX, all Pentiums are capable of doing SIMD on integer arrays, but can't do it on floating point numbers.

Everything else you said is right, but CPUs have had SIMD floating point instructions for quite a while now (in the forms of 3DNow, SSE, and SSE2). SSE (and possibly 3DNow as well; I can't find any explicit statement either way) is single-precision only though.