Doom is a classic game, but it has recently been tested on a wide variety of unusual hardware by computer enthusiasts curious about the game’s compatibility. It’s become something of a novelty to get the game working on technology like a calculator, Microsoft’s Notepad, and more due to the game’s legendary popularity.
One such example is a recent story by IEEE Spectrum. Some experts were able to Doom on much inferior hardware but that’s just not it. The test were conducted on Syntiant’s NDP200 (Neural Decision Processor) and it surprisingly utilized 1/1000th of a Watt (or a Milliwatt). The 320W of the GeForce RTX 4080 is therefore more than 300,000 times more powerful than this. The video below demonstrates Doom running on the neural chip.
The chip is described as having “highly accurate inference capability at less than 1mW” inside its introduction. The fact that we can fire demons in Doom with such a minimal power demand is evidence of how far AI hardware has advanced.
VizDoom, a variation of Doom used for AI research and reinforcement learning from unprocessed visual data, was utilized by Syntiant to train the neural network of the NDP200. To train a neural network, it was necessary to decipher its output, which included, first and foremost, recognizing the opponent and, finally, determining a reaction. The “player” is entrusted with protecting a circle-shaped chamber against ongoing assaults.
In order to play Doom, the neural network needed to understand the game’s mechanics, including how to manage its ammo effectively. The neural network had over 600,000 parameters, all of which had to fit within the NDP200’s 640Kb of Memory and neural core, which operated at 9 GB/s.
This was a great illustration of the “bounding-box person detection” skills of Syntiant’s NDP200, which is designed for low-power activities. While this brief Doom presentation may be entertaining, it also serves as excellent marketing.