Benchmarks

Catalyst vs the Competition

Standard neuromorphic benchmark suite. Fully reproducible. Every result deployable to Catalyst FPGA hardware.

Catalyst N3 Latest

All N3 models use adaptive LIF neurons with surrogate gradient BPTT and cosine LR scheduling.

Benchmark Classes Architecture Neuron Float Acc Params
SHD 20 700→1536→20 (rec) adLIF 91.0% 3.47M
SSC 35 700→1024→512→35 (rec) adLIF 76.4% 2.31M
N-MNIST 10 Conv2D+LIF→10 LIF 99.1% 691K
GSC-12 12 40→512→12 (rec, S2S) adLIF 88.0% 291K
DVS Gesture 11 Deep conv+rec adLIF 89.0% ~1.2M

Catalyst N2

All N2 models deploy to Catalyst N2 FPGA hardware via int16 weight quantization.

Benchmark Classes Architecture Neuron Float Acc Params
SHD 20 700→512→20 (rec) adLIF 84.5% 759K
SSC 35 700→1024→512→35 (rec) adLIF 72.1% 2.31M
N-MNIST 10 Conv2D+LIF→10 adLIF 97.8% 466K
GSC KWS 12 40→512→12 (rec, S2S) adLIF 88.0% 291K
MIT-BIH ECG 5 187→128→5 (rec) adLIF 90.9% ~35K

Catalyst N1

N1 uses basic LIF neurons only (no adaptation). Demonstrates competitive performance through model capacity alone.

Benchmark Classes Architecture Neuron Float Acc Params
SHD 20 700→1024→20 (rec) LIF 90.6% 1.79M
N-MNIST 10 Conv2D+LIF→10 LIF 99.2% 466K
DVS Gesture 11 Deep conv+rec LIF 69.7% ~1.2M
GSC-12 12 40→512→12 (rec, S2S) LIF 86.4% 291K

Competitive Context

Benchmark Catalyst N3 Catalyst N2 Catalyst N1 Loihi 2
SHD 91.0% 84.5% 90.6% 90.9%
SSC 76.4% 72.1% 69.8%
N-MNIST 99.1% 97.8% 99.2%
GSC-12 88.0% 88.0%

Loihi 2 results from Mészáros et al. 2025 (Table I, best per-dataset).

FPGA Hardware Characterisation

Kria K26 Edge (xczu5ev, 2-core variants, 100 MHz target)

Processor LUTs LUT% FFs BRAM DSP Fmax Power
N1 20,109 17.2% 30,847 52.5 (36.5%) 14 (1.1%) 100 MHz 0.642W
N2 26,431 22.6% 38,666 52.5 (36.5%) 16 (1.3%) ~97 MHz 0.688W
N3 53,420 45.6% 80,395 24 (16.7%) 20 (1.6%) ~58.5 MHz 0.867W

AWS F2 Cloud FPGA (Xilinx VU47P)

Processor Tests Pass Rate Throughput Frequency
N1 PASS 62.5 MHz
N2 28/28 100% 8,690 ts/sec 62.5 MHz
N3 19/19 100% 14,512 ts/sec 62.5 MHz

Methodology

Training

Surrogate gradient BPTT with fast-sigmoid surrogate. AdamW optimizer, cosine annealing LR, gradient clipping at 1.0. Event-drop augmentation for regularization.

Quantization

Float32 → int16 weight quantization. Membrane decay mapped to 12-bit fixed-point. Typical accuracy loss: <1%. Scale: w_hw = round(w_float * threshold_hw / threshold_float).

Deployment

Quantized weights loaded into Neurocore SDK Network. Deployed to Catalyst FPGA via USB/UART or AWS F2 cloud instance. Same test set evaluated on hardware.

Reproduce Everything

# Clone the benchmark repo
git clone https://github.com/catalyst-neuromorphic/catalyst-benchmarks.git
cd catalyst-benchmarks
pip install -e .

# Train any benchmark (auto-downloads data)
python shd/train.py --neuron adlif --hidden 1536 --epochs 200 --device cuda:0 --amp
python ssc/train.py --hidden1 1024 --hidden2 768 --recurrent2 --epochs 70 --device cuda:0 --amp

# Evaluate quantized accuracy
python shd/deploy.py --checkpoint shd_model.pt

# Full hyperparameter sweep
python sweep.py --benchmark shd --device cuda:0