Promo At this point, every server vendor+dog has some sort of integrated CPU-GPU system that they’re offering to their high performance computing (HPC) and other number-crunching customers.
The recently introduced Tetra from Appro has upped the ante by offering more GPUs than any other manufacturer in a small 1U package.
This solution is super-dense with up to 40 server units per rack, each with dual CPUs and quad NVIDIA Tesla M2050 GPUs. Appro offers this server in both Intel and AMD flavors – giving customers a choice between the 4 or 6 core Intel Xeon or the 8 or 12 core AMD Magny Cours processors.
That’s one of the topics John Lee and I discussed in the video that accompanies this blog. My question to John was how customers decided between AMD and Intel. It comes down to the characteristics of the workload. The AMD procs can be configured with more memory and, of course, have more cores to address it.
The Intel CPUs have fewer cores but greater memory bandwidth and also great bandwidth to the GPUs. We also talked about the ratio of CPU to GPUs - typically a ratio of one to one. But this is changing rapidly as the model develops. According to Appro, its two GPU to one CPU ratio is the new standard but the company is starting to see customers looking for one CPU to three or four GPU configurations.
However, this can begin to strain the capacity of the PCI Gen 2 bus. Although this will be addressed next year with PCI Gen 3 (which will double current Gen 2 performance), the move to Gen 3 might highlight shortcomings in ability of operating systems to address large numbers of devices.
Our wide ranging conversation progresses to where we’re talking about the PCI bus becoming in essence a true system interconnect and speculating about what the future holds. The future will be speedy for sure...take a look at my Appro video (no registration) to hear more.