V1.3.0 | Forge 1.20.1 | NeoForge & Fabric 1.21.1
TaCZ:Accel is a performance addon for TaCZ. It cuts down repeated rendering, animation, attachment, bullet and resource work that really does not need to happen every frame.
Same guns, same animations, same bullet behaviour. Just less CPU and RAM abuse, especially with bigger gunpacks :DDD
Basically TaCZ, except it stops bullying your PC so much xd.
Compatible guns can stay GPU resident instead of rebuilding transformed geometry every frame. Only changing bone data needs updating, while the CPU side reuses stable poses, keyframes, attachments and render lookups.
AUTO is enabled by default. It tests eligible CPU and GPU routes for each model and display context. It is not perfect, but it does not blindly force the GPU just because a gun is big xd.
Iris/Oculus shader packs keep their lighting, effects and framebuffer handling. Compatible passes can still use GPU skinning without replacing the pack's whole renderer.
Accelerated Rendering and TaCZ Mesh Loader keep their own renderers too. TaCZ:Accel optimises around them instead of trying to replace everything :L
Arcana underbarrel switching is handled properly as well, with stack specific model selection and live visibility callbacks preserved.
Complementary Reimagined r5.5.1 and Unbound r5.1.1 were tested on Forge. AUTO keeps unlisted shader packs on the native path unless explicitly opted in. The 1.21 ports still need wider shader and gunpack testing.
Collision checks are cheaper, not different. Travel time, drop, damage, penetration, headshots and player lag compensation still belong to TaCZ.
The replacement clipping path passed 300,000 randomized comparisons, with another 100,000 checks covering non-player bounds. A fresh collision clipping kernel ran at 18.19 ns normally and 13.92 ns with Accel, about 23.4% less CPU time in that microbenchmark.
| Test | Measured Result |
|---|---|
| M38 SPR, direct skinning vs packed CPU | 24.69% less render tick time, 20.22% less render-thread CPU/frame. |
| 512 M4A1 frame guns vs packed CPU | 24.497 ms → 12.089 ms render-thread CPU, 50.65% lower. |
| Reimagined ADS, forced GPU vs CPU compatibility | 63.16% less render tick time, 51.01% less render-thread CPU/frame. |
| Unbound scope cache vs native scope | 44.66% less elapsed time per scope call. |
1080p M38 idle, stock control against default AUTO. These are the middle results from two pairs, not a guaranteed FPS gain.
| Metric | Normal | Accel AUTO | Change |
|---|---|---|---|
| Median frame | 2.137 ms | 1.758 ms | 17.8% lower |
| Median FPS equivalent | 468 | 569 | 21.6% higher |
| P95 | 3.632 ms | 3.083 ms | 15.1% lower |
| P99 | 4.766 ms | 4.028 ms | 15.5% lower |
| Render CPU/frame | 2.252 ms | 1.844 ms | 18.1% lower |
Forge 1.20.1, RTX 3050 Laptop GPU. The two stock comparison pairs ranged from 5.6% worse to 51.7% better, and AUTO chose different routes. These values are measured central results, NOT a dependable 21.6% boost. The tables above also test different render paths, so do not add their percentages together xd.
These are from the older Forge 1.20.1 tests and the old description. Keeping the numbers here for reference, but they were not rerun or revalidated for this current OmniJar. The old detailed hand, shell and stage figures were removed from the current performance report, so do not treat them as current V1.3.0 measurements.
| Older Test | Normal | Accel | Old Result |
|---|---|---|---|
| Average frame time | 2.145 ms | 1.102 ms | 48.6% lower |
| Equivalent FPS | 466 | 907 | 94.6% higher |
| First person gun render | 760.73 µs | 412.59 µs | 45.8% lower |
| Gun model draw | 656.51 µs | 362.11 µs | 44.8% lower |
| Gun geometry | 83.95 µs | 47.00 µs | 44.0% lower |
| Buffer flush | 140.38 µs | 70.05 µs | 50.1% lower |
| Deferred left hand | 123.23 µs | 62.06 µs | 49.6% lower |
| Deferred right hand | 31.23 µs | 18.84 µs | 39.7% lower |
| Shell render | 6.10 µs | 2.19 µs | 64.1% lower |
| Scalar cube compiler | 0.55 µs | 0.26 µs | 52.7% lower |
| Bullet non-player hitbox check | 8.56 ns | 2.10 ns | 4.08x faster |
| Hitbox candidate allocation | 232 B | 64 B | 72.41% smaller |
With Accelerated Rendering's native GPU mesh path already running, the archived paired test reported 8.1% lower P95 frame time, 9.4% lower P99 frame time and a 63.2% shorter worst frame. Average FPS stayed within roughly 3% of the control. These are older test results, not new OmniJar gains.
Detached JSON/Lua source bytes can be kept in compressed 32 KiB pages instead of retaining their full raw size. Files are never rewritten, and live parsed objects do not get compressed underneath TaCZ xd.
In an older test using 1,184 detached JSON/Lua assets, the raw source bytes went from 30,758,670 B to 3,352,979 B after compression, or 89.1% smaller. That is source byte compression, NOT proof of an 89.1% reduction in the live game's RAM usage. This old test has not been rerun for the current OmniJar.
Quantified API handles bounded background work such as detached resource compression. Mutable Minecraft/render state stays on the correct thread because throwing OpenGL calls onto random workers and praying is NOT an optimisation :DDD
Requires TaCZ and Quantified API V2.1.0 or newer. Fabric also needs Fabric API and TaCZ Refabricated's normal dependencies. Mesh Loader is optional.
Use the same OmniJar on all three targets. Fabric 1.20.1 isn't included.
Forge: /tacztbench start runs a 30 second capture. Use /tacztbench gpu auto, off, on, instanced or direct to force routes.
For a stock control, launch with -Dtaczt.optimizations.enabled=false and keep the gun, world, camera and settings matched :]
Found a broken model, animation issue or crash? Send the log, loader version, TaCZ build and gunpack name so I can actually reproduce it :]
Admany | BlackRift Studios | 2026
