Research Bits: July 8

2D TFETS for neuromorphic computing

Researchers from the University of California Santa Barbara and Intel Labs used 2D transition metal dichalcogenide (TMD)-based tunnel-field-effect transistors (TFETs) in a neuromorphic computing platform, bringing the energy requirements to within two orders of magnitude (about 100 times) the amount used by the human brain.

The 2D TFETs have lower off-state currents as well as a low subthreshold swing (SS), which describes how effectively a transistor can switch from off to on. A lower SS means a lower operating voltage, and faster and more efficient switching, according to UC Santa Barbara electrical and computer engineering Professor Kaustav Banerjee.

“Since the power efficiency of these chips is constrained by the off-state leakage, our approach — using tunneling transistors with much lower off-state current — can greatly improve power efficiency,” Banerjee said in a release. [1]

Real-time biomimetic SNN

Researchers from the University of Bordeaux, the University of Tokyo, the University of Genova, and the Istituto Italiano di Tecnologia developed a low-cost, embedded, flexible, and real-time biomimetic spiking neural network (SNN), called BiœmuS.

This embedded system offers a user-friendly solution for closed-loop applications, especially when compared to server-based infrastructures or complex systems, which can be expensive and difficult to integrate into experimental setups. For example, even software alternatives accelerated by GPUs often struggle to achieve the low latencies needed for closed-loop applications.

In a release, University of Bordeaux’s Timothee Levi, the senior and corresponding author of the study, said: “We envision our system as a crucial step toward the development of neuromorphic-based neuroprostheses for bioelectrical therapeutics, enabling seamless communication with biological networks on a comparable timescale. The embedded real-time functionality of BiœmuS enhances practicality and accessibility, amplifying its potential for real-world applications in biohybrid experiments.” [2]

Readout IC for brain surface recordings

imec researchers proposed a compact neural readout IC for high-resolution, large-scale micro-electrocorticography (µECoG) arrays, which can capture brain-surface activity with single neuron precision.

“The IC is implemented in 22nm CMOS technology and records from 3,072 electrodes via 96 multiplexed channels, where one channel registers signals from 32 electrodes. Each multiplexed channel occupies only 0.0136mm2 and consumes 14.02μW,” according to a release.

The architecture keeps track of the electrode DC offset in each electrode and cancels it while keeping power consumption low. This is essential for high-density arrays because power-hungry electrodes generate heat.

References

[1] Pal, A., Chai, Z., Jiang, J. et al. An ultra energy-efficient hardware platform for neuromorphic computing enabled by 2D-TMD tunnel-FETs. Nat Commun 15, 3392 (2024). https://doi.org/10.1038/s41467-024-46397-3

[2] Beaubois, R., Cheslet, J., Duenki, T. et al. BiœmuS: A new tool for neurological disorders studies through real-time emulation and hybridization using biomimetic Spiking Neural Network. Nat Commun 15, 5142 (2024). https://doi.org/10.1038/s41467-024-48905-x

[3] X. Huang, X. Yang, A. Lodi, C. Van Hoof, G. Gielen, C. Mora Lopez (2024). A 3072-Channel Neural Readout IC with Multiplexed Two-Step Incremental-SAR Conversion and Bulk-DAC-Based EDO Compensation in 22nm FDSOI. In: 2024 IEEE symposium on VLSI technology & circuits, June 2024, Honolulu, Hawaii.