Thread

If you’ve been following neuromorphic computing, you know the frustration. Photonic chips promise the speed of light, but they have been tethered to an n2 data bottleneck. We’ve been burning massive energy just to convert digital memory into something the optical "brain" can understand. Basically, the "thinking" was fast, but the "remembering" was expensive. I caught a paper today that changes the math. Using Dynamic Electro-Optic Analog Memory (DEOAM), researchers have integrated memory directly into the light-paths. By using simple, elegant capacitors to hold synaptic weights, they have cut power consumption by 26x. In short, the real magic is how the hardware is designed to handle "leakage" just like a real brain. This paves the way for on-chip, online training.. Basically no more relying solely on GPUs for the heavy computing. We’re finally seeing hardware that doesn't just calculate, but actually functions like the architecture it’s trying to mimic.. if you are curious, you can read more about it here: www.nature.com/articles/s41...
Neuromorphic photonic computing with an electro-optic analog memory - Nature Communications
Neuromorphic photonic systems can incur significant energy for moving and converting data between digital and analog domains. This work shows that integrating analog memory into these processors can s...
www.nature.com