Skip to content

Benchmarks

This page presents reproducible performance benchmarks for DSG‑JIT, demonstrating the efficiency and scalability of our differentiable Scene Graph optimizer across SE(3) pose chains, voxel grids, and hybrid problems.
All benchmarks were run on macOS (Apple Silicon) using JAX CPU back‑ends.

🚀 Overview

DSG‑JIT achieves 31-1000× speedups over pure Python Gauss‑Newton solvers thanks to JAX vectorization, JIT compilation, and careful manifold‑aware optimization.

Benchmarks included:

  1. SE3 Gauss‑Newton Chain Benchmark
  2. Voxel Chain Gauss‑Newton Benchmark
  3. Hybrid SE3 + Voxel Gauss‑Newton Benchmark

1. SE3 Gauss‑Newton Benchmark

Solves a chain of num_poses SE(3) variables connected by odometry constraints.

200‑Pose Chain (JIT Enabled)

Elapsed time: 51.802 ms
pose0 (opt):   [7.94e-06 4.59e-06 0 ...]
poseN-1 (opt): [1.990e+02 5.88e-04 0 ...]

200‑Pose Chain (No JIT)

Elapsed time: 376,098 ms   (~6.27 minutes)

Result: JIT is ~ 7250Ă— faster.

2. Voxel Chain Gauss‑Newton Benchmark

Solves a smooth‑regularized chain of voxel centers.

500‑Voxel Chain (JIT Enabled)

Elapsed time: 96.192 ms
voxel0 (opt):   [0.1457 0.0425 0]
voxelN-1 (opt): [494.0813 0.018 0]

500‑Voxel Chain (No JIT)

Elapsed time: 3,044.991 ms

Result: JIT is ~ 31Ă— faster.

3. Hybrid SE3 + Voxel Benchmark

Solves both a 50‑pose SE(3) chain and a 500‑voxel chain simultaneously.

JIT Enabled

Elapsed time: 149.832 ms
pose0:     [-4.19e-19 4.00e-10 0 ...]
poseN-1:   [4.90e+01 1.28e-08 0 ...]
voxelM-1:  [4.989959e+02 -1.199e-03 0]

No JIT

Elapsed time: 97,500 ms

Result: JIT is ~ 650Ă— faster.

đź“Š Summary Table

Benchmark JIT Time Non‑JIT Time Speedup
SE3 Chain (200) 51.8 ms 376,098 ms ~7250Ă—
Voxel Chain (500) 96 ms 3,045 ms ~31Ă—
Hybrid SE3+Voxel 150 ms 97,500 ms ~650Ă—

📦 How to Run Benchmarks

From the project root:

python3 benchmarks/bench_gauss_newton_se3.py
python3 benchmarks/bench_voxel_chain.py
python3 benchmarks/bench_hybrid_se3_voxel.py

đź“ť Notes

  • The first JIT call includes compilation time (can take minutes for large graphs).
  • Timings reported above exclude compilation and measure cached execution.
  • Performance scales linearly with problem size once compiled.

🔬 Reproducibility

To ensure consistent benchmarking:

export JAX_PLATFORM_NAME=cpu
export XLA_FLAGS=--xla_cpu_enable_fast_math=true

âś… Conclusion

DSG‑JIT dramatically accelerates differentiable scene‑graph optimization, delivering up to 7000× speedups on real‑world sized problems.

These benchmarks confirm:

  • JIT + vectorized Gauss‑Newton is extremely efficient
  • Hybrid problems remain fast and stable
  • DSG‑JIT is ready for large‑scale SLAM and differentiable graphics tasks