Engineers teach their 1-milliwatt neural chip to play Doom, say this is serious work yo, everyone nods

pcgamer.com:

The NDP200 is an ultralow power chip that was not designed for your gaming rig. This thing is designed to essentially monitor video via neural networks while using barely any power and, when it spots something of interest like a person walking past, wake up beefier systems to do the 'real' monitoring. AI boffins being AI boffins, of course, they decided to show off its capabilities by teaching it to play Doom.

During the 2023 IEEE International Solid State Circuits Conference in San Francisco last week Syntiant(opens in new tab), the chipmaker behind the NDP200, showed how it had been trained to play VizDoom (thanks, IEEESpectrum(opens in new tab)). The latter is a stripped-back and lightweight version of Doom that's often used in AI research and reinforcement learning.

The level shown has the straightforward title Defend the Circle and features a round room that constantly spawns enemies which, when killed, are immediately de-spawned (so no death animations). The NDP200 was trained on a neural network consisting of several layers in order that the chip's 640 kilobytes of onboard memory could identify the demons of hell, aim at them, and fire.

IEEE fellow and Syntiant's recently departed chief architect David Garrett says the whole reason for training a chip like this to play Doom is to show «you can do meaningful detection and actions at this scale.» The NDP200 also had to learn over time that spamming fire is not a good idea. «After its first kill, it unloads the clip, but then it figures out that’s not a good strategy,» said Garrett.

I don't think I can quite wrap my head around 1 milliwatt of power being used in this fashion, but assume this was a conference highlight. Training an ultralow power chip to play Doom is