Meteorhead

Proper GPU workstation

Discussion created by Meteorhead on May 18, 2011
Latest reply on May 22, 2011 by galmok
any intention?

Dear All!

From time-to-time I look around the market to see if there are any pure AMD based workstation solutions or just standalone pieces of HW that can be put together for something capable.

I am somewhat puzzled why it is, that there are practically none AMD based workstations, and it's impossible to make one. Intel and NV has several good solutions, the nF200 chip is extremely useful for splitting PCI-E lanes (create two x16 slots from one, both working at x16, but they share the 16 lanes, so when both are running, they compete for bus bandwidth.) Naturally the nF200 is non-existent on AMD chpiset motherboards, but what is more surprising is that nothing of the like exists on AMD motherboards.

If I wish to create dense AMD cluster, I would need tonns of host machines, I could not have any hosts with 4-6-7 PCI-e x16 slots each of them running at at x16, even if they are split lanes competing with each other. FX990 chipset has 32 lanes, and always they are put as x16/x8/x8/x4, there are no x8/x8/x8/x8 motherboards, not to mention x16/x16/x16/x16 with an nF200-like solution.

Most likely there is no time anymore to develop something like this for the FX990 chipset, but next chipset will most likely feature PCI-E 3.0, less slots required for the same bandwidth... to put it short, are there any plans to create something, so that one could hook up 4-6 dualGPU cards to a single host with decent bandwidth?

ps: Anyone know of PCI-E extenders that have some lane-splitting capability? Or any that are capable of creating two slots from one?

http://www.cubix.com/product/gpu-xpander-rack-mount-16

This Cubix solution looks really good, but I do not know how they achieve one host card to create two slots in the extender rackmount. Is it really that special/custom-made that one cannot find it standalone in the market? It would be favorable for many multi-gpu applications to be able to have as many GPUs comunicate as possible without the use of LAN. Enterprise animation studios, research centers, etc. could benefit from such solutions. This isn't black magic though, one could build it at home, I just cannot find the means to split slots and lanes. Ideas?

Outcomes