Kevin D Johnson
But can it play DOOM? Never mind your technical papers! I want to know if it can play DOOM! Immensely practical, I know.
After hearing about Cortical Labs training DOOM with human brain cells on their own chip, I knew we had to give the Akida/Symphony Neuromorphic Hive Mind a shot as well. So, let's ask the question. Can the Hive Mind play DOOM?
The answer is yes, fast and furious.
Cortical Labs put 200,000 living human brain cells on their CL1 chip and taught them to play DOOM. Their system keeps neurons alive in a nutrient bath while 59 electrodes translate game frames into electrical stimulation and read back spike responses as game actions, an amazing scientific achievement that demonstrates adaptive goal-directed learning from biological tissue on silicon.
The
BrainChip Akida/
IBM Neuromorphic Hive Mind is a distributed neuromorphic system. Ten BrainChip AKD1000 spiking neural network processors, each on its own single-board computer, all orchestrated by an IBM Spectrum Symphony cluster. Each chip contains 80 neuromorphic processing units running spiking neural networks at roughly one watt. Every game frame goes to all ten chips simultaneously. They each make a specialized decision and the convergence engine fuses their votes into a single action, 35 times per second.
The difference is not just silicon versus biology. The difference is architecture. Our system is a distributed compute platform. It scales. You can add chips, assign them new roles, retrain individual nodes without taking down the hive, and let Symphony handle fault recovery if a chip goes offline. The same Symphony infrastructure that runs HPC batch jobs, financial risk calculations, and quantum workloads now coordinates a ten-chip spiking neural network playing a first-person shooter.
Each AKD1000 completes inference in about 4 milliseconds. All ten run in parallel, so the full hive mind produces a coordinated decision in under 5 milliseconds against a frame budget of 28 milliseconds. When you watch the video below, the agent moves so fast it is hard to follow. The speed is not manipulated here, it's real-time neuromorphic inference on ten spiking chips drawing about ten watts total cranking through the game at light speed.
Getting models onto the hardware meant working through various architectures to find one that fit the AKD1000 well. We trained a DQN reinforcement learning agent on VizDoom to generate gameplay data, then used behavioral cloning to transfer that knowledge into a convolutional model designed for the chip.
Both projects prove the same thing from opposite directions. Biological neurons on an electrode array and silicon spiking networks on a distributed cluster can both learn to play DOOM. One is a breakthrough in biological computing. The other is a scalable neuromorphic compute platform that runs on edge hardware, standard orchestration software, and ten watts of power.
500 million+ members | Manage your professional identity. Build and engage with your professional network. Access knowledge, insights and opportunities.
www.linkedin.com
on 12 March:
thestockexchange.com.au
