New dual Fermi card(s)!

Accelerated Computing CUDA CUDA Programming and Performance
1

There are strong rumors that a dual GF110 card (the GTX590) will be announced tomorrow at PAX 2011. We’ll probably hear the buzz and announcements. Some unofficial pre-buzz here.

But today, a dual GF104 card was announced by EVGA, the GeForce GTX460 2WIN.
Gotta love those marketing names!

I’ve been a huge fan of the GTX295… both my main tools scale perfectly with 3 of these dual cards in one box, and it’s hard to beat the compute density!
I’ll be the first to get the GTX 590 (assuming the rumors are true…)

2

Putting three dual GF110 in one box would be a real challenge though. I mean the numbers on that swedish site look reasonable, but 375Watt per GPU would make it hard to fit the box even into a 1500W power envelope - not to mention heat issues if you put three of those back to back. But I guess two of those is not unreasonable, the only question is if that is better than 3 GTX580 in a box.

Ceearem

3

CUDA apps typically use much less power than graphics apps, typically only 2/3 the wattage. I have had 3 GTX295’s running 24/7 for weeks at a time, with no special cooling.

We will see of course how the new dual Fermi cards perform. There are both power and cooling concerns of course.

With the new 460 2WIN, it will be interesting if their big custom side blowing fans would work well when multiple cards are sandwiched. I suspect they won’t work too efficiently… a top and back intake works for the GTX295 (even if much much smaller than the open framework of the 2WIN)

4

And a report onGalaxy’s version.
This report hints that the dual GF110 GTX590 will be announced on the 22nd. It also mentions the possibility of a dual GF114 card (GTX 560 2WIN?)

5

Interesting how they omit the memory specs of the card (in terms of memory SIZE available to each GPU). Apparently that number must not have been too impressive.

I want a dual GPU with 3 GB per GPU. I’d buy it for raytracing.

Christian

6

I bet it’s 1GB per chip. GF104 has a 256 bit memory interface, and 512MB is too small, so 1GB would be the sweet spot. Similarly, the GTX590 will likely have 1536 MB per chip.

If NVIDIA makes a dual-chip Tesla (unlikely) you could use GPUDirect2 to let each GPU use the other GPU’s memory. The sibling memory queries would be slow, but wouldn’t even leave the card.

I do this now for my app using host zero-copy, so it’s possible to have a 2GB+ scene database even though the GTX295 has only 768 MB of RAM.