India's Neuromorphic Push Tests Whether the Brain Can Teach AI to Run Cooler

AI's appetite for compute is outrunning what denser GPUs alone can deliver. A growing cohort of Indian researchers is betting the next architectural answer may not come from faster silicon at all, but from chips that compute more like neurons, firing only when an input crosses a threshold and idling the rest of the time, with the 20-watt human brain as the standing benchmark.

"Neuromorphic" describes chips built to mimic how biological neurons process information, not by running every operation every cycle, but by sending short electrical spikes only when inputs matter. The human brain runs on roughly 20 watts; a rack of AI accelerators can pull tens of kilowatts. That gap is the structural bet, in the framing of Alex P. James, director of India's IIITM-Kerala, in an EE Times interview.

James does not pitch neuromorphic as a GPU replacement. He frames it as a portfolio question. The category covers spiking neural networks, cellular neural networks, hierarchical temporal memory, and analog in-memory crossbars. India is not converging on one of those; it is working on several at once, at device, circuit, architecture, and system levels, per the EE Times feature on the field.

The most concrete result sits at IIT Bombay. Prof. Udayan Ganguly's group built an ultra-low-power spiking neuron, documented in the group's IEEE-published work, that repurposes the quantum-tunneling leakage current that analog designers usually try to suppress. Fabricated in a standard 45-nm CMOS process from GlobalFoundries, the design reportedly draws roughly 1,000 times less power than comparable sub-threshold neurons and trims process, voltage, and temperature variability by about five times, per the EE Times interview.

Ganguly's lab also built an end-to-end spiking speech-recognition pipeline, filters plus a liquid-state-machine reservoir plus a classifier, that reportedly hit about 99% accuracy on a standard keyword benchmark using only about 100 neurons, with under 2% error on two-class yes/no and start/stop tasks. The fabricated recurrent stage uses 36 neurons. Per-neuron specs, as reported: about 100 kHz firing rate, 40 µm² silicon area, 2.70 picowatts of standby power, and roughly 8 femtojoules per spike, a figure the EE Times piece describes as close to the human brain's per-spike budget.

A second Indian bet runs in parallel at the Indian Institute of Science (IISc) in Bangalore. Profs. Sreetosh Goswami and Navakanta Bhat reported a ruthenium-based molecular memristor that switches deterministically across four orders of magnitude with linear and symmetric conductance updates, achieving 14-bit analog resolution and 4.10 TOPS/W energy efficiency. The work was framed in a follow-up EE Times piece on the molecular memristor, and the device is being integrated into a mixed-signal AI inference accelerator in 22-nm TSMC, with analog crossbars for the heavy linear algebra and digital control for the rest, via a startup incubated at IISc's Centre for Nano Science and Engineering. The architecture framing tracks a survey of in-memory computing in Foundations and Trends in Electronic Design Automation.

The pattern is two research groups working at different layers of the same architectural question: one at the device-plus-circuit boundary using standard CMOS plus clever leakage, the other at the device-plus-architecture boundary using an exotic molecular switch. Both are being steered toward working systems, not just papers.

James flags the real gating factor: applications, not device physics. A biologically inspired algorithm does not automatically map to an efficient chip, and the field has not converged on a winning substrate. Ganguly sketches a longer horizon, a roughly 20-year vision of neuron-and-synapse silicon that could one day interface with damaged brain tissue, framed in the EE Times feature as a research aspiration rather than a product roadmap.

For now, the 20-watt brain stays a benchmark, not a target. The portfolio question is whether toolchains, training flows, and a real business proof-of-concept catch up to the device work. India's labs have not answered that, but they have made it a question with named, on-record contenders.