VQFieldBenchmarks
Every published measurement, and how it was taken
A number is only useful if you can tell what was measured. These historical results name the data, the denominator or timing phase, and the build context. Test your own caches before choosing a setting.
Source · 4.0 GBQ50 · 341 MB- Whole sequence
- 341 MB file bytes
- Ratio
- 11.8 : 1 source ÷ encoded file
- Average per frame
- ≈ 2.05 MB rounded size ÷ 166
- Frame shown
- 1279 × 1193 original render pixels
Original PNG renders. Difference maps are image-space comparisons in sRGB, not voxel error metrics or quality scores.
Source VDBs, encoded volumes and Studio reports are not yet published.
Measurement ledger
All results, with their scope
Ratios and timings that describe different things stay in different rows. These are measurements of the listed historical workloads, not current-build or whole-render performance promises.
| Figure | Value | Data | Denominator or phase | Host · build · commit | Date | Kind |
|---|---|---|---|---|---|---|
| Size ratio at Q70 | 3.9 : 1 | Smoke plume, 166 frames · real | source ÷ encoded file bytes | 9950X3D · MSVC Release / AVX2 · revision not recorded in published table | 15 Aug 2026 | Measured |
| Size ratio at Q50 | 11.8 : 1 | Smoke plume, 166 frames · real | source ÷ encoded file bytes | 9950X3D · MSVC Release / AVX2 · revision not recorded in published table | 15 Aug 2026 | Measured |
| Size ratio at Q30 | 109.4 : 1 | Smoke plume, 166 frames · real | source ÷ encoded file bytes | 9950X3D · MSVC Release / AVX2 · revision not recorded in published table | 15 Aug 2026 | Measured |
| Size ratio at Q20 | 447.3 : 1 | Smoke plume, 166 frames · real | source ÷ encoded file bytes | 9950X3D · MSVC Release / AVX2 · revision not recorded in published table | 15 Aug 2026 | Measured |
| Size ratio at Q50 | 14.81 : 1 | Production file, 4 grids · private | source ÷ encoded file bytes | 9950X3D · MSVC Release / AVX2 · 6f072c6 | 15 Aug 2026 | Measured |
| Size ratio at Q50 | 56.42 : 1 | 256³ scalar field, 4,915 active blocks · synthetic | active FP32 bytes ÷ encoded bytes | 9950X3D · MSVC Release / AVX2 · 6f072c6 | 15 Aug 2026 | Measured |
| Prepare a sampler | 42.5 ms | One-grid density file · private | sampler creation, once per file | 9950X3D · MSVC Release / AVX2 · 6f072c6 | 15 Aug 2026 | Measured |
| First sample, uncached block | 66 ns | One-grid density file · private | prepared sampler · 125 uncached blocks | 9950X3D · MSVC Release / AVX2 · 6f072c6 | 15 Aug 2026 | Measured |
| Sample from cache | ~9 ns | One-grid density file · private | warm access · 8 MiB cache · approximate | 9950X3D · MSVC Release / AVX2 · 6f072c6 | 15 Aug 2026 | Measured |
| Q50 transfer at 1 Gb/s | ~2.7 s | Smoke plume, 166 frames · real | 341 MB × 8 ÷ 1 Gb/s; no protocol overhead | Arithmetic from rounded published size | 15 Aug 2026 | Derived |
| Decode + VDB write, Q20 | 0.63 s | Smoke plume · real | includes file read and VDB output; historical decoder | 9950X3D · MSVC Release / AVX2 | 15 Aug 2026 | Historical |
The original sequence table did not record a source revision. Commit 6f072c6 identifies the separate block codec benchmarks only. The historical decode-and-write timing includes I/O and cannot be compared directly with prepared sampling latency.
Methodology
Read the figures on their own terms
Sizes are rounded for display. Ratios come from the published measurements, so dividing rounded sizes may give a slightly different result.
Denominators
File bytes: source .vdb bytes divided by encoded .vqfield bytes, as stored on disk.
Active FP32 bytes: four bytes per active voxel and channel. This denominator ignores the source file’s compression and metadata. The two ratios are not interchangeable.
Timing phases
Prepare creates the sampler once. First access reads a block that is not cached. Cached access reads previously decoded values.
These phases stay separate. An individual lookup benchmark does not measure a complete renderer or transfer pipeline.
Quality settings
Q is an encoder setting, not a percentage of the original. The same setting gives different sizes and errors on different fields.
Check rendered sequences, field errors and the downstream workflow together. A ratio alone is not an acceptance test.
Comparison images
These supplied renders are aligned views of the same frame. The 416 × 416 crops are from the area of greatest image difference at Q20.
Difference maps show per-pixel differences in sRGB. Orange starts at the selected mean RGB difference threshold, which defaults to 13/255. The viewer reports the share of plume pixels above it. Render noise can contribute to this difference.
A comparison with OpenVDB + Blosc or half-float storage requires the same source data and a separate measured baseline. No such comparison is claimed here.