Not everyone has a bottle of The Famous Grouse Scotch whisky as a liquid-coolant tank in their PC, but it seemed entirely appropriate in the Alpha Data demo at SC15, even if the coolant looks a lot more like Green Absinthe than Scotch. The demo shows an Auviz Deep Neural Network (DNN) demo classifying 1000 images/sec using 28W while running on a Xilinx Virtex-7 690T FPGA. The same demo running on an Intel-based PC with a Core i7 processor classifies 200 images/sec while consuming 90W. That’s a 16x performance/W improvement. But the kicker is the cooling system. The PC needs a liquid-cooling system, hence the bottle of Scotch whisky. The FPGA on the Alpha Data board, well that need a “little heat sink.”
Here’s a 1.5-minute video of the demo presented by Alpha data’s Managing Director, David Miller:
Note: If you’re wondering about the SC15 demo’s connection to the Scotch whisky coolant tank, as I did, Alpha Data is headquartered in Edinburgh, Scotland.
For more information about the Auviz DNN application that’s running on the Xilinx Virtex-7 690T FPGA, see “Machine Learning in the Cloud: Deep Neural Networks on FPGAs” and “A deep look at accelerating Convolutional Neural Network performance, from Auviz Systems.”