Dual-track benchmark design

NeuroBench organizes evaluations in two complementary tracks:

  • Algorithm track: hardware-independent benchmarking for correctness and complexity.
  • System track: deployment-aware timing and efficiency benchmarking.

Algorithm track

  • Keyword FSCIL
  • Event Camera Object Detection
  • NHP Motor Prediction
  • Chaotic Function Prediction
  • GSC
  • DVS Gesture
  • NeHAR

Algorithm benchmarks: github.com/NeuroBench/neurobench

System track

  • Acoustic Scene Classification
  • Quadratic Unconstrained Binary Optimization (QUBO)

System benchmarks: github.com/NeuroBench/system_benchmarks

Metrics coverage

Each benchmark run can report metrics across correctness and efficiency dimensions.

  • Static metrics: Model Footprint, Parameter Count, Connection Sparsity, Model Execution Rate
  • Workload metrics: Classification Accuracy, COCO mAP, R2, sMAPE, MSE, Activation Sparsity, Synaptic Operations, Membrane Updates, Neuron Operations

See full metric documentation in the harness docs: Static metrics and Workload metrics.

Learn more