graphics card idea

[yoda291]

Ok, I read a post in this forum by CountZero on the difference between GPUs and CPUs and it got me thinking. According to things like Moore's law and whatnot, circuits are getting smaller and smaller so that, in time, we'll be able to snap in huge amounts of graphics processing power into our AGP slots in a few years. Here's my idea though. How feasible is it to, instead of a full graphics card, drop in an interface card, run something like a fiber optic cable to a seperate box with the actual GPU and memory. You're no longer constrained by space and can put like a gig of vid memory and multiple GPUs on a board. Granted, this wouldn't be cost efficient for the home user, but what about for a server-client model of some sort for professional graphics? Possibly linking multiple copies of these graphics boxes to make a scalable graphics architecture like what the Beowulf cluster does for raw CPU power. I'm pretty sure that optic cable bandwidth wouldn't be up to the task, but I suspect that splitting a signal between several cables would be possible. Has anyone heard of something like this? Thoughts? Comments?

 

[janlievens]

minor note:
fiber optics has more potential bandwidth than current electronic devices can handle.
agp 4x = 66MHz * 32bit * 4 = 1056 MBytes/s = 8.858 GBit/s
fiber = estimated limits around 1 TBit/s

 

[MustPost]

someday computers may use fiber on the inside.

This sounds just like a PBS thing I saw in like 97 or something.
The guy thought within 5 or 10 years everyone would be using dumb terminals.
everything except the I/O would be stored elsewhere.
when you wanted software you would just get it from legitement websits.

 

[yoda291]

I read about a company that tried something like that, they were selling PCs with a CPU, 1gb hard drive and a modem. IT would load up their OS and connect to the internet where all the data would be stored. I think it flopped tho.

 

[interchange]

You still have the nasty factor of speed-of-light delay in any electrical connection.

This is actually material dependent; in air or a vaccum or in fiber it is the commonly known c = 3x10^8 m/s...
The actual formula can be simplified to v|p = c/sqrt(epsilon|r), which explains the difference in different materials.

Additionally, there's a period in time where your signal bounces until it becomes stable, which can be eliminated by matching resistors which dissipate a lot of power.

But, assuming a completely ideal case, and you wanting to put say 1m of cable connecting your video card and your motherboard...
And also that your GPU is so fast it can do things in 0 time...
Your max clock rate is c (in seconds) = 300Mhz. Which isn't really good for next generation PCs.

The idea of dumb terminals is a better one, which requires all of your critically timed hardware to be in one place. From that, you can get a max refresh rate of 300Mhz .