Energy-proportional spiking transformers
Benchmark attention-style models mapped to spiking substrates and measure joules per inference against conventional edge accelerators.
[CORE_03 · Facility]
Neuromorphic Computing Research Testbed
The testbed is the experimental platform behind the centre's Neural Architecture programme. It supports neuromorphic computing research on biologically inspired, energy-efficient hardware — from spiking-transformer models to analog inference at the edge — and has been available to external collaborators since April 2026.
Facility overview
- Architecture
- Analog spiking neural substrate with mixed-signal synapse arrays
- Scale
- Multi-die configuration supporting spiking-transformer workloads
- Instrumentation
- Per-core energy measurement for energy-proportional inference studies
- Access
- Open to external collaborators since April 2026
Platform capabilities
- Workload mapping
- Spiking conversions of attention-style and convolutional models, plus natively spiking architectures written against the centre's mapping toolchain.
- Measurement
- Per-core energy and latency capture on every run, reported as joules per inference alongside accuracy so results stay comparable across sessions.
- Operating regimes
- Runs can be swept across supply, sparsity, and timestep budgets to characterise behaviour under realistic edge power limits rather than a single tuned point.
- Baselines
- Matching runs on conventional edge accelerator hardware, so a testbed number always ships with a like-for-like comparison point.
Exact device generations, die counts, and instrumentation models are confirmed during scoping, since the configuration is set per campaign rather than fixed.
What researchers use it for
Edge AI hardware characterisation
Profile latency, sparsity, and noise tolerance for always-on perception workloads under realistic power budgets.
Learning rules on analog hardware
Evaluate local plasticity and surrogate-gradient training when device variability and drift are part of the system.
Access terms
- Who can apply
- Academic groups, doctoral researchers, and industry R&D teams. Applicants outside the centre are welcome; no prior relationship is required.
- Cost
- No charge for collaborative academic work. Industry engagements are scoped case by case before any run is scheduled.
- Review
- Proposals are reviewed on a rolling basis by the Neural Architecture group. Expect an initial response within two weeks of enquiry.
- Allocation
- Runs are scheduled in supervised sessions with a member of the group rather than granted as unattended remote access.
- Outputs and credit
- Collaborators keep authorship of their own work; testbed staff are credited where they contribute intellectually, and configuration details are shared so runs can be reproduced.
- Data and IP
- Measurement data belongs to the collaborating group. Anything commercially sensitive is agreed in writing before the first session.
How a collaboration runs
- 01
Enquiry
Send a short description of the workload and what you want to measure.
- 02
Scoping call
Thirty minutes to check the workload maps cleanly and to size the run.
- 03
Proposal
A one-to-two page plan covering models, sweeps, baselines, and outputs.
- 04
Scheduled runs
Supervised sessions on the testbed with measurement capture throughout.
- 05
Results
Raw traces and a summary handed back, with support for write-up if useful.
Enquire about testbed access
External access runs as a collaboration rather than a rental: a short proposal describing the workload, measurement plan, and expected outputs, followed by scheduled runs with a member of the Neural Architecture group. Results from testbed runs feed the centre's spiking-transformer programme and are presented at the annual Winter Seminar on neuromorphic computing at the edge.
Email the group with a few lines about your workload — the link below opens a message with the details we need already laid out. There is no application deadline.
Or write directly to csdiscoverylab@gmail.com