GPU Server - Exxact Tensor TS4-264546-DP2?... Is this an old April fools joke?

[RCS]

I stumbled upon this article on Exxact's blog that covers the advantages of using PLX chips and single root PCIE complexes for GPU deep learning etc.

Example 4 highlights their new "Tensor TXR414-1000R" system which can apparently take up to 20 GPU's, all while using 5 PLX switches to essentially eliminate any previous P2P bottleneck (besides the limits of PCIE obviously).

"It takes the concept of PCIe P2P communication to the limit supporting full P2P communication across up to 20 GPUs. This system can incorporate 10 standard Titan-X or M40 GPUs or, with modified single width high performance heatsinks, 20 single wide Titan-X or M40 GPUs in a single system image."

Other than this odd blog post, I can't actually find any info on this. The system that is linked in the article takes me to the "TS4-264546-DP2" system, which just going off specs, seems identical to Supermicro's 4028GR-TRT2.

Original blog post: Exploring the Complexities of PCIe Connectivity and Peer-to-Peer Communication

Is this real or is this an April fools joke from 2016?...

​

 

[EffrafaxOfWug]

PLX PCIe switches got super expensive following the broadcom buyout, so it's possible such a system didn't make economic sense after that. You're also still limited to a single 16x link between the CPU and all of the GPUs, effectively giving less than 1GB/s bandwidth per GPU when all 20 slots are populated. Chassis form factor might also be an issue.

I think it'd be an interesting workload that required 20 GPUs on a single host and wouldn't be bandwidth limited in such a situation, when five GPUs on four hosts would likely work out cheaper and more scalable (but then I'm not really involved with GPU-based computing).

 

[Patrick]

Perhaps this also will help https://www.servethehome.com/superm...-16x-or-20x-nvidia-tesla-t4-inferencing-gpus/

You tend to want to use bigger GPUs first, then more GPUs, then more systems for training in that order.

 

[RCS]

Thanks for the reply's!

Makes sense. Would have been one hell of a machine, but I'm sure also very very expensive.