Beyond GPUs: Optical, Neuromorphic and Biological Computing

YC Paper Club brings together François Chaubard, Ilker Oguz, Alok Vasudev and Sean Cole to discuss alternatives to conventional GPU-based AI compute. Their topics include optimization, optics, neuromorphic chips and biological neurons.

Chaubard describes how modern AI grew around a combination of model architectures, backpropagation and hardware. He asks whether power, memory and communication constraints could encourage different combinations in the future.

He sees memory capacity and bandwidth as major concerns for transformer workloads. Hardware and model design have influenced each other, so an alternative processor may also need algorithms designed around its strengths.

Chaubard predicts that alternatives to backpropagation could become more important. That is a research forecast; the panel does not establish that backpropagation or GPUs will disappear on a particular schedule.

Optical and neuromorphic devices can have different constraints from digital accelerators, including difficulty storing state, implementing nonlinearities or differentiating through physical operations. Those constraints influence which learning algorithms are practical.

Chaubard describes testing zero-order optimization methods. Instead of directly computing derivatives through every operation, one can perturb parameters, compare forward-pass losses and use the differences to estimate an update direction.

His example is simultaneous perturbation stochastic approximation, or SPSA. The estimator uses randomized perturbations; its practical cost and accuracy depend on the number of dimensions, samples and the problem being solved.

He questions whether architectures designed for gradients and GPUs are ideal for a forward-only optimizer. An alternative learning rule may call for a different architecture instead of a direct substitution into an existing transformer.

His proposed SOMA approach clusters data and independently trains small recurrent experts, then routes inputs among them. The [research preprint](https://arxiv.org/abs/2609.37899) describes the sharded zero-order setting and its evaluation.

The motivation is that gradient-estimate noise grows with the dimension being perturbed. Sharding keeps each expert smaller, but it introduces its own design and routing costs; the reported approach is research, not proof of a general replacement for backpropagation.

Chaubard suggests forward-only learning might benefit if physical hardware made forward passes much cheaper. The complete system must still pay for input, output, memory and repeated measurements, so that premise needs system-level testing.

Ilker Oguz introduces optical computing. Data centers already use optical links for communication; the separate question is whether light can also perform useful operations inside a computer.

Light can support parallel propagation and low-loss transmission under appropriate conditions. Oguz discusses how optical systems might perform matrix operations, but advantages depend on the hardware, precision and the full workload.

In digital electronics, moving and switching data consumes energy. Optical computation can exploit physical propagation, but that does not make the surrounding lasers, detectors, conversions and memory free.

Oguz explains why optical computing has not displaced electronic processors: the whole energy budget includes converting information between electronic and optical representations and getting data into and out of the device.

Another difficulty is nonlinearity. Passive optical components can implement large linear operations, but neural networks also use nonlinear activations, which add hardware and design complexity.

Oguz presents work on diffusion-based image generation. Diffusion models progressively transform noise toward an image using repeated denoising steps, so the same learned operation is applied many times.

The [optical diffusion paper](https://arxiv.org/abs/2407.10897) programs light propagation through transparent layers for a denoising task. Those passive layers are reused across denoising steps, while the broader system still includes electronic components.

Oguz discusses a design in which several devices can be reused rather than physically building a separate layer for every denoising step. The experiment explores a specific architecture and should not be read as a production replacement for a GPU.

He reports energy comparisons under assumptions including passive, fixed weights. Those assumptions matter: programmability, input-output conversion and system integration can change the overall result.

In questions, participants ask where optics loses its advantage. Oguz highlights workloads that must transfer large amounts of data relative to the useful optical computation; interfacing can overwhelm the savings.

The discussion covers fabrication, repeated passes and ways to encode multiple channels. Wavelength multiplexing may increase parallelism, but nonlinear layers, device size and conversion remain practical constraints.

Alok Vasudev introduces neuromorphic computing as an effort to borrow organizing principles from biology when designing circuits and chips. He does not claim that today’s chips reproduce a biological brain.

He points to the brain’s combination of sensing, action, memory and adaptation as motivation. Comparisons with biological energy use are suggestive, but the tasks, measurements and operating conditions differ from those of data centers.

Among the principles he highlights are memory and computation being intertwined, event-driven communication and changes in connection strength. Researchers translate selected ideas into circuits rather than copying every feature of neurons.

Biological neurons combine discrete-looking spikes with continuous dynamics. Vasudev describes attempts to build electronic counterparts and emphasizes that the analogy is imperfect.

Current deep-learning systems use architectural and training techniques that do not map directly onto biology. He cautions against assuming that a model optimized for one physical substrate will automatically work well on another.

Examples include bringing computation closer to memory, spiking neural networks and experimental semiconductor devices. Some ideas are more mature than others, and it is necessary to compare complete systems on concrete tasks.

During questions, Vasudev distinguishes early research from established products and considers how architecture and hardware should be co-designed. He also discusses whether noise can sometimes be useful rather than treated only as an error.

Sean Cole describes a hybrid experiment connecting living neurons on an electrode array to the game Doom. Software translates game state into stimulation and decodes electrical activity into game actions.

This builds on earlier work connecting neuronal cultures to Pong, while Doom introduces a more complicated action and observation space. The neurons are one component of a system that also relies on ordinary digital computation.

Cole describes a reinforcement-learning setup based on proximal policy optimization. A conventional neural network processes observations and provides a feedback signal, while the electrode array supplies biological activity used in the action path.

He explains how the decoder maps spikes to actions. If a powerful decoder learns the entire task on its own, the apparent success would not show that the cells learned; he discusses restricting the decoder to address that concern.

The team uses different patterns of stimulation as feedback tied to the game’s outcomes and prediction error. Cole describes the mechanism as a hypothesis and engineering approach, not a complete account of how biological learning works.

Questions probe what “positive” and “negative” mean, how feedback affects the cells and whether noise supports exploration. Cole stresses that the labels refer to game outcomes and discusses uncertainty in interpreting the biology.

Asked whether the biological approach can scale, Cole says the team is trying to find out. A compelling demonstration does not establish cost, reliability, learning mechanisms or competitiveness at commercial scale.