Hi,
There seems to be quite a few companies out there selling compute servers with 8 Tesla cards in a single 4U rackmountable box.
Does anyone know about any attempts to use 8 295 GTX cards in a single box?
/Lars
It will be fun to see how many Fermi it will take, to get the same performance that the Fastra II is currently getting…
Granted, I’m sure their Fastra II system will also be updated with Fermi.
Also note Fastra II uses a custom BIOS:
[url=“http://fastra2.ua.ac.be/?page_id=628”]http://fastra2.ua.ac.be/?page_id=628[/url]
I don’t think there are. The usual choices are 4 x GTX295, which gives 8 GPUs (not 8 cards), or a pair of Telsa S1070s (each contains 4 x1000 series Telsa in 1U), plus a 2U host with enough PCI-e slots to hold the interface cards. The FASTRA 2 seems to be the only attempt to put that many cards in a single host (and as noted, they need a custom 64 bit BIOS and operating system patches to map the memory apertures of all the cards correctly so they can be used). I am not aware of any commercial 8 card solutions.
There was a company in the recent GTC in San Jose that put 8 C1060 in one tower machine… maybe you can google
that. But I agree with avidday most likely 4 GTX295 are easier
eyal
I stand corrected on that one, and thanks for the information. Colfax’s previous Tesla based products were all either 1 or 2 C1060 in rack mount or 3 or 4 C1060 in a deskside tower. That new box looks to be using a dual Tylersburg IOH board (so I guess the Tyan “gpu blade” board), with custom PCI interface cards using PEX8647 PCI-e switched to share the PCI-e slots amongst the C1060s.