The plummeting cost of deep learning silicon

[strangecosmos]

Brett Winton on Twitter:

2016: nVidia introduces the drive PX 2, a 16 tera-op liquid-cooled supercomputer for the car to enable autonomous capability at the low-low price of ~$15,000

2018: Buy 3 iPhones for ~$3,000 and shove them into the glove compartment; you get roughly the same amount of performance

Bodes well for the 2019 deep learning hardware that Tesla is going to put in Hardware 2 cars.

 

[strangecosmos]

Fascinating and exciting that Tesla is taking a bet on big data sets (requiring production fleet learning) and big neural networks (requiring ~100 tera-ops of processing power). If you want to increase the capabilities of deep learning, bigger data sets and bigger neural networks are exactly what you want. No one else in the world is trying to train neural networks on billions of miles of driving data. Waymo only has ~10 million miles of data. Tesla is taking this two orders of magnitude higher over the next year, then three orders of magnitude higher in the following years.

This approach is not guaranteed to work. Deep learning could yet fail us. But the preliminary results are really encouraging. On certain narrow tasks, deep learning is already superhuman.

I’m excited for Autopilot v9. Could that be the inflection point where Tesla’s strategy is publicly vindicated?