NeuroCUDA SpiNNaker Silicon Confirmed
Two independent SUCCESS runs on two different physical SpiNNaker-1 boards at the University of Manchester. This is neuromorphic silicon - not a simulator.
TL;DR - Physical silicon
NeuroCUDA SpiNNaker physical silicon is confirmed. Job #420148 (13 Jul 2026, board 10.11.242.169, 43 chips) and job #420186 (17 Jul 2026, board 10.11.196.49, 48 chips / 851 cores) both returned SUCCESS with neuron 0 = 2 spikes and neuron 1 = 2 spikes. Platform: EBRAINS Neuromorphic Computing Service, Manchester. Source: github.com/Krishnav1/neurocuda.
For months, neuromorphic tooling conversations blurred simulators and chips. NeuroCUDA already shipped GPU, CPU, and a Loihi 2 simulator. Today the SpiNNaker path is different: it executed on physical SpiNNaker-1 silicon through the EBRAINS queue - twice, on two boards, on two dates.
Confirmed SpiNNaker physical silicon runs
| Job | Date | Board IP | Hardware | Result |
|---|---|---|---|---|
| #420148 | 2026-07-13 | 10.11.242.169 | 43 chips, 766 cores | SUCCESS |
| #420186 | 2026-07-17 | 10.11.196.49 | 48 chips, 851 cores | SUCCESS |
Both runs produced identical spike counts: neuron 0: 2 spikes, neuron 1: 2 spikes, total 4 spikes. Same network, different boards, same outcome. That is the point of repeating the experiment - a single lucky queue slot is no longer a plausible explanation.
What ran on the silicon
Job #420186 used a minimal PyNN validation network designed to prove the hardware path:
- Input: 2 SpikeSourceArray neurons
- Output: 2 IF_curr_exp (LIF) neurons
- Synapse: OneToOne, weight 5.0, delay 1.0 ms
- Simulation: 10 ms at 1 ms timestep
- Firmware: SC&MP 4.0.0
- Stack: sPyNNaker 7.4.2, PyNN 0.12.4
A cosmetic log warning (cfg has no version) appeared and does not affect the neural computation. Status remained SUCCESS.
Why "not silicon" still appears for Loihi
Older NeuroCUDA pages said backends were not claimed on physical neuromorphic silicon. That sentence covered the general case and especially Loihi 2, which remains an IF-neuron simulator checked against Intel's published equations - not Intel silicon, not Lava. SpiNNaker is now the exception that is proven: real boards at Manchester.
Do not rewrite Loihi claims as silicon. Do not hide SpiNNaker silicon behind simulator language either. The honest matrix is:
| Backend | Status (July 2026) |
|---|---|
| GPU / CPU | Shipped - multi-seed accuracy published |
| SpiNNaker-1 | Physical silicon confirmed (2 boards, 2 jobs) |
| Loihi 2 | Simulator only (equation validation) |
| NIR export | Shipped - ResNet-18 residual bit-exact |
| FPGA | HLS C++ proof of concept |
What this does and does not prove
It proves
- NeuroCUDA's SpiNNaker / PyNN path can reach real SpiNNaker-1 boards via EBRAINS
- Spike delivery is reproducible across two independent machines
- The neuromorphic silicon claim for SpiNNaker is no longer theoretical
It does not prove
- ResNet-18 / CIFAR-10 accuracy on SpiNNaker chips (those 94.61% numbers are GPU/CPU)
- Energy or latency superiority vs GPU on silicon for large models
- Physical Loihi 2, Akida, or BrainScaleS deployment
Calling a 4-neuron SUCCESS run "ResNet silicon accuracy" would be false marketing. Calling it "SpiNNaker physical silicon confirmed" is accurate and citable.
Where to verify
- GitHub README (updated 2026-07-17): github.com/Krishnav1/neurocuda
- Job record:
rEPORT 1/job420186_result.txtin the NeuroCUDA repository - Product page: quantaracore.in/neurocuda
- EBRAINS Neuromorphic Computing Service (University of Manchester SpiNNaker)
Primary sources
- NeuroCUDA GitHub - SpiNNaker physical silicon section and job #420186 result file
- EBRAINS Neuromorphic Computing Service - SpiNNaker-1 job queue (jobs #420148, #420186)
- University of Manchester / SpiNNaker project - SC&MP firmware and sPyNNaker stack
- QuantaraCore NeuroCUDA product page - backend honesty labels
Frequently asked questions
Has NeuroCUDA run on physical SpiNNaker silicon?
Yes. Two SUCCESS jobs on two different SpiNNaker-1 boards (#420148 and #420186) via EBRAINS Manchester.
Is Loihi 2 also physical silicon now?
No. Loihi 2 in NeuroCUDA is still a simulator. Only SpiNNaker-1 is confirmed on physical boards in this announcement.
What network ran on SpiNNaker?
A 4-neuron PyNN LIF validation network (2 sources, 2 IF_curr_exp, OneToOne weight 5.0), 10 ms @ 1 ms.
Does this prove ResNet accuracy on SpiNNaker?
No. It confirms the silicon execution path. Full-model accuracy remains published for GPU/CPU backends.
Next: pip install neurocuda · Product page · GitHub