I’m new to the GPU/Cuda form of accelerated computing. I am the IT guy and my engineers say they need a server with 2 K80’s on it. They say they don’t need much of a cpu and ram, but that’s not much for me to go on. They said they would be using Matlab and Python. I saw somewhere that 2 cpu’s will help with performance.
can anyone recommend cpu and ram for a server like this?
thanks
Adam
Use of dual K80s suggests that some very intense processing is envisioned, and it would help to know what that is. It would also help to know what your budget is.
A common mistake in configuring high-end systems for GPU-accelerate computing is underprovisioning the host system. Some generic thoughts:
(1) For the host system to communicate with both GPUs at full speed, you will need a CPU with ~40 PCIe lanes (2x16 for the GPUs, plus some for SSDs, etc)
(2) While dual K80s can make short work of the parallelizable portion of the planned workload(s), you do not want the serial portion of workloads to become a bottleneck, which is why I would suggest a CPU running at >= 3.4 GHz. The vendor buzzword is “frequency optimized” processor, e.g. Xeon E5-2637 v3, Xeon E5-2643 v3, Xeon E5-2637 v4, Xeon E5-2643 v4 etc.
(3) My general experience is that sizing system memory at 2x to 4x the total GPU memory size makes for a well-balanced system. However, the sizing will likely be strongly determined by the exact type of processing intended to run on the system.
(4) To keep the cost of running the system to a minimum (and cut down on waste heat), I would suggest use of an 80 PLUS Platinum power supply (some vendors even offer 80 PLUS Titanium PSUs, which are a tad more efficient still).
You should choose an OEM server that is qualified for K80 usage. Building your own is fraught with peril, and you can find many examples of the peril if you search around on this forum.