NVIDIA announced DGX-1, their new “GPU supercomputer”. The spec is impressive. Performance, even more so (training AlexNet in 2 hours with 1 node).
Costs $129K. Running this would take around 3KW. That’s like keeping an oven going.
The cheapest (per hour) best config you can currently get from AWS is g2.8xlarge:
So for $129K you get around 9y worth of AWS GPU time, but that’s not going to give the same performance as 15B cores connected by a high-speed fabric. So, should you get the DGX-1? I would get my hands all over it if I had access, but I’m also reminded of the lean startup movement. A lot of the startups in the 90s dot com boom went bust because they sunk in a ton of money in hardware equipment.
The most important change brought by cloud computing, however, might be in spending. The dot-com bubble was fueled by a hardware and software buying explosion as companies built and grew computer networks, but that spending was cyclical, and after people and companies bought what they needed, it crashed: U.S. technology spending by companies grew at a compound annual growth rate of 13% from 1995 to 2000, before falling to -3.5% over the next three years.
There is a lot of interesting work that can be done with existing commodity GPUs. Be clever. Be resourceful. Above all, be frugal. 2016 is not a time to squander.