What If We Stopped Using GPUs?

The host traces how deep learning hardware shifted from optimizing FLOPs to prioritizing memory capacity and bandwidth for transformers. He argues that gains in FLOPs per joule have slowed and points to the brain’s roughly 20 watt power use as motivation to explore other approaches. He also outlines a proposal called SOMA: train small, specialized LSTM experts on clustered data and combine them at inference.

Ilker explains how light can carry information with low loss and high bandwidth, and how passive optical systems can perform parallel linear operations. He describes an application specific optical device for diffusion based image generation, using repeated light propagation to denoise images; tests used small datasets such as MNIST and Fashion MNIST. The reported energy advantage depended on fixed, passive weights, while data conversion, calibration, nonlinear operations, and optical memory remain challenges.

Aloque describes neuromorphic computing as borrowing principles such as intertwined memory and computation, event driven signals, mixed analog and digital behavior, and adaptation through learning. He says the field remains in R&D, with work spanning digital memory near compute, spiking systems for real time edge tasks, and devices such as memristors and coupled oscillators. He identifies memory and compute integration as a nearer term direction, while more brain like approaches remain early.

Sean describes treating cultured brain cells as a changing input output system: an encoder converts Doom observations into electrical stimulation, and a decoder maps the cells’ spikes to game actions. He says the encoding and decoding must be learned for a complex game, and warns that an oversized decoder can appear to perform well while doing the work itself. Feedback uses synchronous or asynchronous stimulation, scaled by surprise; Sean says how to scale biological computing for broad service remains an open question.