The Stanford chips are biomimetic, trying to copy how the brain is thought to work. Artificial Learning's are machine-learning theoretic, embodying proven ML algorithms in hardware.
Ah, thanks for the clarification.
Artificial Learning's largest designs are 180nm tech so some way from Moore's Law limits. Efficiency advantage over GPGPU implementation calculated as better than 10,000 fold in terms of learning updates per second per watt per kg.
Is there are real advantage to have this running on the device? Why not use lower profile hardware like electric imp and run everything in the cloud?
Not if you want to use it in autonomous drones...
How does this compare to CUDA implementations?
A better question would be how it will compare to CUDA implementations when it's released?
Moore's law killed all such initiatives in the past, however, perhaps this time it will be different?