Neural Chip Plays Doom Using a Thousandth of a Watt
Doom is the game that became the benchmark. From humble beginnings on a 386 PC, it has been ported to run on everything, even the humble Raspberry Pi Pico. “It’s no big deal” you say This story is from IEEE Spectrum Stay away from running Doom on low spec hardware. Instead, make sure the ultra-low-power chip has learned how to play Doom with just 1 milliwatt of power.
Let’s quantify the power of 1 milliwatt. Even its low power consumption, albeit at a thousandth of a watt, is hard to fathom. Take Nvidia’s RTX 4090 for example, this card draws around 400-450 watts of power. That’s about 400,000 times more power than Syntiant’s NDP200 uses. Admittedly, the NDP200 doesn’t make the list of best GPUs because of its focus on using data to make decisions based on training. Killing Doom is just plain fun.
Syntiant’s NDP200 (Neural Decision Processor) is an ultra-low power chip for neural networks. It is primarily used to monitor video and audio to trigger events that other systems react to. The NDP200 runs up to 100 MHz and has 26 GPIO pins just like the original Raspberry Pi.
Syntiant trained a neural network for the NDP200 using: Bizdoom, a version of Doom used for AI research and reinforcement learning from raw visual information. Training involved understanding what the neural network was seeing. Mainly identifying enemies and ultimately defining reactions. In this case, see the demon and shoot the demon. The “player” is tasked with defending a circular room that is under constant attack. A neural network had to learn how to play Doom. This also meant learning how to conserve ammo. The neural network consisted of about 600,000 parameters, all of which were compressed into his 640Kb of RAM and neural core on his NDP200, running at 9 gigabytes per second.
The purpose of the demo is not to show how well Doom can be played, but to show how efficient the NDP200 is at “bounding box people detection” which usually requires a much more powerful processor. . Capable of scanning 6 frames of video and performing this task with just 1 milliwatt of power, the NDP200 can be easily integrated into vehicle and home security systems. Syntiant claims this is 1/100th the power of his Arm Cortex A53, the same Arm chip that powers the Raspberry Pi 3.
For now, NDP200’s AI limits carnage to demons. I can only hope that it doesn’t start communicating with Bing’s chatbot.