BENCHMARKS

Performance & Quality

Independent benchmarks measuring Knonik training speed, compression quality and dataloader throughput.

benchmarks

Every Number Here Links
To The Run That Made It.

We benchmarked Knonik against standard tooling on real robotics data, newest first. GR00T N1.7 trained end to end in a third of the time, streaming from S3. Loader throughput measured end-to-end against lance and LeRobot, on a harness you can run yourself. Our research paper trains 60 policies to test whether the storage format changes what they learn. Compression quality verified across three policy architectures.

TrainingOct 2026View details →

GR00T N1.7 Training Speed

How much faster does a real model train?

Throughput vs stock: 2.8× faster
Same loss: 3× sooner
Cost per run: 62-66% lower

NVIDIA GR00T N1.7 fine-tuned on BridgeV2 for 3,000 steps on 8× A100, with only the dataloader changed. Wall-clock time, convergence, GPU idle time and cost against GR00T's own loader.

ThroughputAug 2026View details →

Loader Throughput

How fast do decoded frames reach the GPU?

Throughput vs lance: up to 1.56× faster
Throughput vs lerobot: up to 12.3× faster
Storage vs lance: up to 18.6× smaller

knonik vs lance vs LeRobot on three robot datasets from 180×320 to 720p, each driven through its own public API. Frames/s and samples/s, the latency split, host memory, on-disk footprint, and a public repo to reproduce every number.

ResearchMay 2026View details →

Knonik Is All You Need for Scaffolding Robot Learning

Does the storage format change what a policy learns?

Success vs HDF5: +9.9 pp
Lossless storage: 3.51× smaller
Lossy storage: 18.23× smaller

60 BC-Transformer trainings on LIBERO-Object (500 demonstrations, 149,014 RGB frames) across 3 storage modes, 10 tasks and 2 seeds, with only the data format and dataloader changed. Task success, stability across seeds, storage size and fetch latency, plus the full paper as a PDF.

CompressionAug 2026View details →

Training Quality Benchmark

Does Knonik compression preserve learning signal?

Domain gap: < 0.004 MAE
Models tested: 3
Best MAE ratio: 0.774×

Diffusion Policy, ACT, and Flow Matching Policy trained on compressed vs uncompressed data. Full learning curves, domain transfer analysis, and per-model breakdowns.