Rubidium ( Salut_46 )

# Rubidium By Salut_46

Rubidium skips certain operations when the vanilla contribution to a structure’s burial is already known to be exactly zero. It keeps the original formula for all other cases and does not replace noises, seeds, or biome algorithms.

A standard installation preserves Minecraft’s rendering, graphics queues, and generation scheduling. No menu or configuration file is required. The mod automatically disables its computation integration when C2ME is present to avoid overlap; full mod combinations still need to be tested separately.

This first experimental release is available for Fabric 1.21.1, 1.21.11, 26.1, and 26.2. It is provided with its source code and test results. The observed generation improvement is modest and scenario‑dependent. A general FPS increase and the absence of stuttering are not demonstrated.

The name comes from rubidium, an alkali metal with atomic number 37. Technical identifier: rubidium_flow. Independent project, unrelated to Rubidium for Forge.

# Measurements and decisions — Rubidium 0.1.0-alpha.1

## Selected result

The release ships only the exact-computation shortcut. It contains no generation regulator or active render budget in a normal installation. The graphical gains of the prototypes below do not describe the delivered product.

## Reference and method

"Reference" means Minecraft with Fabric Loader, Fabric API and the same benchmark setup, with Rubidium optimizations disabled. This is not Minecraft without Fabric. Machine: Ryzen 5 PRO 7535U, integrated Radeon, 12 logical processors, maximum Java memory 4 GiB, Java 21. Client: 1280 × 720, view distance 8 chunks, simulation 5, VSync off, limit 260 corresponding to uncapped mode. 1.21.1 indicates Fancy; 1.21.11 indicates CUSTOM, with its default settings and the parameters above. Do not directly compare FPS between versions.

Each graphics run measures 60 seconds, after a 20-second warmup, with four stops in fresh terrain spaced 15 seconds apart. Revision 2 fixes the warmup camera to (60.5,110,-40.5). Worlds are fresh copies of seed 123456789. No benchmark or build runs concurrently. Kept runs indicate zero frames without focus and zero frames with a menu screen.

## Dedicated generation 1.21.1

144 chunks at FULL stage, request concurrency 8, 16 warmup chunks elsewhere, tickets preserved and five seconds of quiet. Order A/B/B/A, two runs per variant. The selected computation code was active in B, the Pulse regulator disabled. A dedicated server uses no graphical budget.

| Measurement | Reference | Rubidium selected shortcut |
|---|---:|---:|
| Average time for 144 chunks | 11.541 s | 11.391 s |
| Run 1 | 11.580 s | 11.371 s |
| Run 2 | 11.501 s | 11.411 s |
| Ticks measured > 50 ms | 0 | 0 |

About 1.3% less time, for this scenario only. Two repetitions do not demonstrate statistical significance nor a universal win. Sources: `final-perf-a1`, `final-perf-b1`, `final-perf-b2`, `final-perf-a2`. The historical field `gcPauseCollectionMs` measures collection time reported by Java, not exact pause durations. Tick durations are not FPS.

## Graphical prototype discarded — 1.21.1

Initial trials, revision 1: reference 285.87 FPS, 16 frames >50 ms, P99 14.064 ms; graphical budget without Pulse 305.74 FPS, 12 frames >50 ms, P99 14.086 ms. P99 did not improve. With Pulse, only 17 visible sections remained versus 74 for the reference: this prototype was rejected.

New trials, revision 2, in a Fabric development environment:

| Run | Avg FPS | P99 (ms) | Frames > 50 ms | Maximum (ms) | Sections at last sample |
|---|---:|---:|---:|---:|---:|
| client-final-a1 | 298.73 | 15.895 | 22 | 326.88 | 70 |
| client-final-b1 | 379.30 | 11.272 | 10 | 100.31 | 76 |
| client-final-b2 | 389.76 | 11.299 | 16 | 228.24 | 84 |
| client-final-a2 | 739.19 | 7.035 | 11 | 130.92 | 163 |
| client-final-b3 | 532.38 | 8.894 | 22 | 256.42 | 77 |

A denotes the reference. B denotes exact computation and a graphical budget of 2 ms/64 tasks, without Pulse. A1/B1/B2 were performed in the morning; an interruption of about five hours precedes A2/B3. Do not aggregate these series into a single percentage gain. The last B3 is lower than A2 and shows less terrain at the end. Changes in load, temperature, or power were not measured: the precise cause of variation between sessions remains unknown. This uncertainty and the unfavorable result motivate abandoning the graphical budget in the release.

The number of sections in the last sample is not an exact measure of the time for all chunks to arrive. Captures were also inspected. The tests do not permit any promise of a general FPS increase.

## Delivered JARs: runtime checks with normal Fabric

These trials use remapped JARs, a production Minecraft client and a separate bench. They verify loading, rendering, the run, and saving on close. The graphical budget remains disabled. The 1.21.11 reference uses the prior candidate with all its optimizations disabled; the active run uses the final JAR.

| Version | Load & run | Graphical budget active | Pulse threads | Sections at last sample |
|---|---|---|---:|---:|
| 1.21.11 | Success | No | 0 | 266 |
| 1.21.1 | Success | No | 0 | 84 |

The 1.21.1 "smoke" run checks the normal installation under a distinct bench identifier: zero Pulse threads and no redirection of the render queue. It does not form a performance pair with the older development trials. The 1.21.11 trials did not explicitly neutralize automatic FPS limiting when idle. These operational checks are classified performanceValid=false and are not used to claim an FPS gain. Raw frame times are preserved for traceability.

## Computation checks and limits

48 unit tests pass, including utilities for the prototypes kept in source. Each Minecraft version executed 375,956 comparisons of the real method with the vanilla formula in a Fabric server, then saved dimensions before closing. Grid, boundary values and pseudo-random doubles: bit-identical result, with NaN normalization in the audit.

On 1.21.1, block and biome fingerprints at NOISE stage are identical on 64/64 Overworld chunks, 16/16 Nether, 16/16 End. Two Overworld references also agree. Integrity audits are separate from performance measurements.

FULL final-world comparisons diverged between two references themselves. They remain inconclusive: NOISE-stage fingerprints do not prove full equivalence of reloaded final saves. The 26.x ports have no graphical test nor FPS measurement. Combinations with other mods are not universally certified.

The JSONs, JAR fingerprints and measurement sources are provided. P99 uses the upper rank; average FPS are the inverse of mean frame time. Sources allow reactivating prototypes only in separate benches, to reproduce the research. The mod remains an experimental first release with limited scope.