What struck me is that although I have at times been a huge consumer of computing power I am spectacularly unqualified to answer that question. Perhaps precisely because I have had access to the largest non-classified computing resources in the world. When you deal with massive computing systems, the nature of your interaction with them must necessarily be somewhat abstract.
The last time I fiddled with production hardware directly must have been some time around 1995-1996 or so. Back then I worked part-time for the local university to support myself while starting a company with a few friends. That was the last time I took a machine out of its box and put it into a rack.
After that hardware has largely been about numbers for me.
Quite early on in the game while I was at FAST, we moved our servers to the US. So from mid 1999 on I have rarely been on the same continent as the production resources I've been using. Back then I worked on software that ran on tens of machines. A few years later, at Yahoo, I worked on software that ran on thousands of machines. And finally, at Google, on "an undisclosed number of machines" (I said non-classified -- not non-secret ;-)).
Nowadays, at Comoyo, we run on a very modest (well, relatively speaking) number of EC2 instances from Amazon Web Services, but there are a few things that have not changed:
- I never physically see the machines. They are just numbers to me; in order of importance: RAM, CPU, bandwidth and storage.
- I am not on the same continent as the machines.
- I can provision whatever I need within minutes and ditch it when it is no longer needed
- The machines are extremely basic (no RAID, no expensive complicating bullshit)
Actually, the precursor of Comoyo did purchase a few servers about 3 years ago. It was an eye-opening experience. I was only peripherally involved in setting them up, but we burned an amazing amount of man-hours on dicking around with those servers. I think all of us had forgotten how much work it is to own physical hardware.
HP Moonshot looks like the right thing to do in many ways: packing lots of computing power into a small space and caring about things like energy-efficiency etc. But I actually think the bean-counters will be more qualified to say whether the machine is good or not.
That being said, I certainly would not mind having one for use at home or at work. I know precisely what I'd use it for (an OpenStack-based private cloud for experimenting).