The Russian team from the Nizhni Novgorod National Research University exuded a quiet confidence as they prepared for the competition on Sunday. Despite the challenges of traveling great distances, the language and cultural differences, and last-minute hardware tweaking, they are composed and ready for whatever the SCC will throw them.
The Russian team and Stony Brook are the only teams using GPU accelerators. The Russians are using way more of them – a total of 10 GPUs, two on each compute node – vs. Stony Brook’s use of a single two-socket GPU board.
We all know that GPUs are pure processing speed: fast as a greased weasel on high-quality amphetamines. But there’s a trade-off here. The teams are limited to 26 amps total, and any GPUs they add will limit the number of CPUs they can deploy. Once they lock in a configuration, that’s the configuration they have to run – no swapping out boards during the competition.
The key question for the teams utilizing GPUs is how well they’ve adapted their code to take advantage of GPU awesomeness. If they’ve done a good job, the GPUs can outrun the fastest CPUs by a factor of 10x or more (sometimes way way more). But if their code doesn’t work, or doesn’t work on enough applications, then they’re stuck with inert hardware that’s sucking up watts without spitting out results.
Other teams looked at using GPUs in their builds but backed away due to the complexity – and doubts about the potential payoff. The Russians have rolled the dice on NVIDIA, Microsoft, and their personal prowess with CUDA. Now we’ll see if it’s going to pay off. Could be GPU-riffic (yeah, I’m still trying to coin that phrase).