Elliot Arledge Launches Netherite, A Minecraft Rewrite That Runs 7,200 Reinforcement Learning Worlds On One Card
Elliot Arledge has released Netherite, a C and CUDA rewrite of Minecraft version 1.11.2 engineered for reinforcement learning training. The framework strips the game's full rasterizer during policy execution, allowing a single graphics card to process 7,200 synchronized game worlds simultaneously.
The engine produces bitwise identical results to the official Java edition, with physics divergence capped at 1e-9 variance across recorded test sessions. Operating without a runtime game jar, it regenerates all internal inputs locally so trained policies transfer directly to the live client. Arledge has published the kernel source and oracle mod files for public development.
From the sources (4 posts)
@elliotarledgeIntroducing Netherite! Minecraft 1.11.2, rewritten from scratch in C and CUDA, bit-verified against the real game. This is one trained agent playing in it - then the renderer it actually trains through - then 7,200 live worlds stepping in
@elliotarledgeWhy exactness: it makes an RL sim you can trust. 23/23 recorded tapes replay with physics clean to 1e-9, CPU and CUDA bitwise-identical every tick (all the way to defeating the ender dragon). Train here, replay the same actions in the real
@elliotarledgeTraining mode throws the full rasterizer away. The policy sees a 64x36 block-id camera with depth and edge planes - right pane is the whole observation. That is why thousands of worlds fit on one card.
@elliotarledgeC, CUDA, no engine, no game jar at runtime. You bring your own copy of the game; every Mojang-derived input regenerates locally, byte-identical. I'm excited to see any and all constructive criticism and what the community can do with this.