A CPU that runs entirely on GPU

Original Article ↗ Hacker News Discussion ↗

A GitHub project implements a full ARM64 CPU using neural networks running on a GPU, effectively emulating every ALU operation with learned models instead of using the GPU’s native arithmetic units. Commenters see it as an impressive but wildly inefficient hack—fun for running things like a Doom raycaster or exploring differentiable execution, but nowhere near practical compared to real CPUs or more direct GPGPU approaches. The conversation broadens into whether GPUs could ever replace CPUs, with most arguing instead for heterogeneous systems where CPUs, GPUs and NPUs coexist and share memory.

Project concept

  • The project implements an AArch64 CPU simulator that runs entirely on a GPU using neural networks for ALU operations, including add, mul, sqrt.
  • Commenters frame it as a “because we can” hack, akin to CPUs in Game of Life or Minecraft, more about exploration than practicality.
  • Some note it’s closer to “a CPU on an NPU that happens to be a GPU” than a conventional CUDA-based GPU CPU-emulator.

Performance and practicality

  • One estimate: ~625,000× slower than a 2.5 GHz CPU for addition/subtraction.
  • People question real-world utility, but others argue it doesn’t need a practical purpose.
  • There is curiosity about how many such CPUs could run in parallel on one GPU and whether massive parallelism could offset slowness, but skepticism remains.
  • Comparisons are made to more efficient GPU CPU-emulation approaches (qemu-style dynamic translation to shaders), which could be orders of magnitude faster.

Neural arithmetic and exactness

  • Discussion around whether an LLM (or neural system) should perform exact arithmetic without external tools, versus just calling out to conventional hardware.
  • Some question why one would train networks for operations like sqrt when the GPU already has fast, precise hardware instructions.
  • The inversion where multiplication is much faster than addition is highlighted; explained by lookup-like parallelism for mul vs. carry chains for add.
  • Idea raised that a fully neural CPU makes execution differentiable, enabling backpropagation through programs for program synthesis, though not useful for normal OS work.

GPU vs CPU roles and future

  • Extended debate on whether GPUs could replace CPUs.
  • Consensus trend: CPUs and GPUs solve different problems (latency-sensitive, branchy vs massively parallel), and full replacement is unlikely.
  • Many expect continued convergence into heterogeneous systems (APUs, unified memory, mixed units like GPU, CPU, NPUs, FPGAs) rather than dominance of one.

Uses, OS-on-GPU, and culture

  • Project author’s stated long-term dream: an OS running purely on GPU or on “learned systems.”
  • Some link to prior work on parallel operating systems and “compute in memory” concepts.
  • Thread contains the usual “can it run Doom?” jokes, references to Doom-on-GPU, and playful renamings of “GPU,” capturing both amusement and admiration for the hack.