And in another episode of Someone should install Doom on the Hubble, we find something slightly different. Last week saw the 2023 edition of the ISSCC conference, and Syntiant took the opportunity to talk about its chip NDP200. This silicon has been specially designed for deep learning and video processing, consuming just one milliwatt. The best part? Syntiant trained one of its chips to play a quick game of VizDoom…
One of the big challenges facing artificial intelligence is reducing its energy consumption and increasing efficiency. It takes a lot of firepower to train and run the most popular models, but it all depends on the type of application, and that’s how we get to Syntiant, which recently introduced its NDP200 chip, a "neural decision processor".
Doom, controlled by a one-milliwatt chip
The video (originally presented at the tinyML Summit 2022) is based on VizDoom, a special version with graphics similar to FreeDoom that is very popular for AI development. In relaxed terms, VizDoom allows the creation of bots that play using visual information from the screen buffer. The level is called "Defend the Circle" and it’s relatively simple: all the player has to do is spin on its own axis and shoot everything that approaches.
(Editor’s note: Skip to 10:34 in the second video for more details.)
At first, the NDP200 would make the mistake of emptying the magazine after killing its first monster, but it eventually "learned" that the best strategy is to make more precise shots and conserve ammunition. Still, we notice the occasional double shot, but that’s not bad for such a low-power chip. The rest will be left to the market: with a milliwatt of consumption and half a million parameters, the possibilities are quite broad.
Source: IEEE Spectrum
