[semi-extrinsic]

Poster above says the GPUs cost $6k total, and cases that can fit 4GPUs aren't that expensive nor elusive. So I guess the rest of the system is made from unobtainium, or they figure deep learning is the kind of red hot topic that attracts enough suckers with money that this will fly.

[jfb]

Or people for whom the money value of their time exceeds the margin over a custom build that nVidia is asking.

[semi-extrinsic]

I can spec out a system like that in less than 1h, but let's say it takes 2h, ending up with a total hardware cost of ~$9k plus $200 to have it assembled. Are you really valuing your time at more than $2000 an hour?

[jfb]

If someone's willing to pay, sure! But more realistically, a ML researcher doesn't want to know the differences between Haswell-E and Skylake, or which DIMMs are rated at which speed, or what's the optimal Linux driver situation, still less to know how to fix any of the myriad things that can go wrong. When Caffe starts screwing up, it's really nice to be able to call nVidia and make your overheating Titan their problem.

[semi-extrinsic]

I'm down with you about calling support when your hardware craps out. But an ML researcher who doesn't know the difference between Haswell-E and Skylake, etc., will likely make suboptimal choices when using the hardware. Using GPUs efficiently requires in-depth understanding of how they function and how they interact with the rest of the system.

[jfb]

This is true, and also, I think it's perfectly likely that nVidia is trying to soak people buying with other people's money.