NVIDIA has discussed its plans to push supercomputer technology and enable much more intelligent artificial intelligence (AI), demonstrating its CUDA for Machine Leaning system. Discussing image detection, gesture and speech recognition, recommendation engines, and more, NVIDIA set out its stall for why its GPUs are the natural platform for AI.
In an on-stage demo during the opening keynote, NVIDIA showed an NYU image-recognition system that can not only identify dogs, but their specific breed. NVIDIA invited GTC 2014 attendees to tweet photos of their dogs, and then used AI to pick out the type of dogs shown.
The test system, developed by NYU, uses a 1.2m image training set processed over two weeks by seven GPUs. It took a total of 25 EXAFLOPS to fully train the system.
NVIDIA has been pushing its GPU supercomputers for some years now, last year contrasting its three-array system with Google’s far larger, far more expensive Google Brain image recognition system. Google Brain demands 600 kWatts and cost $5m; in contrast, three GPU-accelerated servers using twelve GPUs cost just $33,000 and use 4 kWatts.