Netherite — Minecraft 1.11.2 rewritten in C/CUDA, bit-verified against Java, 7,200 parallel worlds on one GPU

TL;DR

Elliot Arledge announces Netherite, a from-scratch C + CUDA rewrite of Minecraft 1.11.2 designed to be a bit-verified RL simulator: 23/23 recorded input tapes replay with physics clean to 1e-9, CPU and CUDA kernels are bitwise-identical every tick, and an inventory-screen diff against the real Java game shows 442 differing pixels out of 112,796 (each ≤0.3%). Reported throughput on an RTX PRO 6000 + 9950X3D is 3.0M env-ticks/s at N=8192 batched worlds (~756k agent-steps/s), with 7,200 live worlds simulated in lockstep on one GPU; the exact 1080p renderer reaches 35.9 fps GPU / 4.5 fps CPU. Training mode drops the full rasterizer for a 64×36 block-ID camera with depth and edge planes so thousands of environments fit on one card. Positioning: an RL sim you can trust — train in Netherite, replay the same actions in Java Minecraft, get the same result.

Key claims

Netherite is a from-scratch C + CUDA rewrite of Minecraft 1.11.2 with no game engine and no game jar at runtime; users bring their own Minecraft copy and every Mojang-derived input regenerates locally byte-identical [tweet 8/9].
Bitwise-identical replay: 23/23 recorded tapes (up to defeating the ender dragon) replay with CPU and CUDA ticks byte-equal and physics clean to 1e-9 [tweet 4/9].
Pixel-level rendering fidelity vs Java Minecraft: 442 of 112,796 pixels differ on the inventory-screen diff, each by ≤0.3% (shown with 80× amplification) [tweet 3/9].
Reported throughput: 35.9 fps GPU / 4.5 fps CPU for exact 1080p rendering; 3.0M env-ticks/s at N=8192 batched sim; 756k agent-steps/s (RTX PRO 6000 + 9950X3D) [tweet 7/9].
Training-mode observation is a 64×36 block-ID camera with depth and edge planes rather than the full rasterizer, cited as the reason thousands of worlds fit on one GPU [tweet 6/9].
One demo shows a single trained agent producing a 2,058-action unbroken chain (spawn empty-handed → logs → planks → sticks → crafting table → wooden pickaxe → stone → coal) with nothing scripted [tweet 5/9].

Method

Netherite reimplements the Minecraft 1.11.2 game loop as a headless CUDA sim (for training) plus an exact 1080p GPU rasterizer (for replay/validation). Correctness is enforced by three gates shipped in the repo: input-tape replay against Java Minecraft, per-tick CPU-vs-CUDA byte equality, and pixel-diff against Mojang’s renderer. The RL-relevant path drops the rasterizer and exposes a compressed observation (64×36 block-IDs with depth + edge channels), which is what enables the N=8192 batched-tick throughput. The distributed licensing story is that users supply their own game copy; Mojang-derived assets are regenerated locally rather than shipped. No paper, formal benchmark, or third-party validation is linked; the announcement is a Twitter thread with a promised code + oracle-mod drop.

Results

Headline systems numbers (RTX PRO 6000 + 9950X3D):

Exact-render throughput: 35.9 fps GPU / 4.5 fps CPU at 1080p.
Batched sim: 3.0M env-ticks/s at N=8192; 756k agent-steps/s.
Bitwise CPU↔CUDA identity every tick; 23/23 tapes replay to 1e-9 physics precision including full ender-dragon runs.
Rendering fidelity: 442 / 112,796 inventory-screen pixels differ (0.39% of pixels, each ≤0.3% intensity).
One end-to-end agent demo: 2,058 actions, unbroken chain, unscripted, wooden-pickaxe-to-coal.

No RL training-recipe numbers, no comparison against MineRL / MineDojo / VPT baselines, no third-party throughput measurement.

Why it’s interesting

Netherite lands directly on the rollout cost axis that the RL Environment Platforms cluster has been circling. The recently-filed platforms in that cluster explicitly design around “verifier-cheap, action-expensive” — SETA: Scaling Environments for Terminal Agents (env repo) and Toolathlon-GYM: Large-Scale Long-Horizon Environments for Tool-Use Agents eat the action cost as Docker containers, and Anyscale’s Ray + Isaac Lab pattern (Scale Robot Policy Evaluation with Ray (Distributed Sim-Eval on Anyscale)) eats it as physics + rendering on a separate GPU pool. Netherite is a bet in the opposite direction for a single environment: collapse the action cost by rewriting the simulator itself against the training hardware, and get bit-exactness with the real target as a bonus. Swayam’s note (“this might make rollouts cheaper”) is the exact framing.

It also complements the sibling Pantograph introduces Pan-1 — Minecraft model with RL-based pretraining, arguing video games are the right testbed for robotics Pan-1 announcement Swayam originally shared in this thread. Pan-1 argues Minecraft-class games are the right RL testbed; Netherite is a candidate answer for how to run that testbed at RL-relevant throughput. If the bit-exact claim holds up under external scrutiny, Netherite would be to Minecraft-RL what Genesis World — Simulation platform for general-purpose robotics & embodied AI learning tried to be to robotics-RL: an in-house simulator engineered as a training substrate rather than as a viewer-facing game.

Two things to watch before betting on the numbers: (1) 3M env-ticks/s at N=8192 works out to ~366 ticks/s per world, which is plausible for a stripped-observation kernel but is not directly comparable to the fully-rendered path; (2) no external replay of the Java-parity claim exists yet, and the licensing model (users bring their own jar, assets regenerate locally) is unusual enough to want a legal-side sanity check before any team commits training runs to it.