AI & ComputeLOT 261122 · 3 min read

SNES ChatGPT Runs on a 21 MHz Super FX 2 Coprocessor From 1995

A 2.2M-parameter GPT-OSS-120b distillation runs on a 35-year-old SNES, with a 21.47 MHz Super FX 2 chip accelerating replies ahead of the 2.68 MHz main CPU.

By Sophie Lindqvist560 words

Key takeaways

  1. A 2.2M-parameter distillation of OpenAI's GPT-OSS-120b runs on a 35-year-old SNES.
  2. The Super FX 2 coprocessor runs at 21.47 MHz versus the 2.68 MHz Ricoh 5A22 main CPU.
  3. A 19 MB build embeds opening sentences of 150,000 Simple English Wikipedia articles via MSU-1.
  4. The Super FX 2 debuted with Yoshi's Island and also shipped in the Doom cartridge.
  5. Real-hardware use requires a flash cart with Super FX support, or MSU-1 for the larger build.
35-year-old SNES runs ChatGPT, accelerated with a Super FX 2 chip — 21 MHz of RISC muscle speeds 2.2M parameter chatbot
35-year-old SNES runs ChatGPT, accelerated with a Super FX 2 chip — 21 MHz of RISC muscle speeds 2.2M parameter chatbot — Carlos Irineu da Costa / unsplash

A 2.2-million-parameter language model now runs on a 35-year-old Nintendo SNES, with a 21.47 MHz Super FX 2 RISC chip handling the inference work instead of the console's 2.68 MHz main CPU. Vintage Computing blogger Benj Edwards built the project, and the gap between the two chips is the entire story: the Super FX 2 coprocessor runs roughly eight times faster than the 16-bit Ricoh 5A22 that hosts it.

Edwards called the chip "the Super NES's best GPU" and used it to accelerate chatbot responsiveness on hardware never designed for the task. The Super FX 2 debuted with Yoshi's Island and later shipped inside the Doom cartridge. Nintendo had to factory-install the chip on any cartridge that required it. The second version kept the polygon and advanced 2D capabilities of the original but ran at full internal speed rather than half speed and supported larger ROM sizes.

What model actually fits on a cartridge?

The software is a distillation of OpenAI's GPT-OSS-120b, compressed until it fits vintage silicon. Two builds exist. The 4 MB version is the smaller option. The 19 MB build adds the opening sentences of 150,000 Simple English Wikipedia articles, stored in the cartridge's MSU-1 data file, which lets the model act as a crude calculator and recall a few facts.

Edwards does not oversell the result. "The model is millions of times smaller than the models behind ChatGPT," he admits on the project's interactive page. His prediction for user experience is equally frank: "You'll get some fun and weird stuff."

How do you run it?

The project ships as a ROM with three deployment paths:

  • A browser page where Edwards embedded the SNES9x emulator, no hardware required
  • An emulator on your own machine
  • Real hardware, using a flash cart with Super FX support, plus MSU-1 support for the Wikipedia build

On real hardware, the practical constraints begin. You need a vintage SNES, a flash cartridge that supports Super FX, and a separate MSU-1-capable cart if you want the larger model. That is a niche within a niche.

Input is straightforward. The app prompts you to press Start, then accepts plain keyboard typing, with no joypad text entry required. The interface adds SNES-appropriate sound effects and a chip-tune soundtrack. None of this changes what the model can do.

Does it work beyond novelty?

Testing found the usefulness "very limited" beyond curiosity value, and Edwards concedes the same. That verdict matters more than the technical achievement. A 2.2M-parameter model cannot hold a coherent conversation, and the Wikipedia knowledge base covers only opening sentences of each article.

The engineering is still genuine. Squeezing a transformer distillation through a 21.47 MHz coprocessor from 1995, alongside a 2.68 MHz main CPU and cartridge-based storage, demonstrates how far model compression and emulation tooling have come. The project proves the pipeline works end to end on constrained silicon.

Whether it points anywhere commercial is doubtful. Nobody will run chatbots on 16-bit consoles by choice. But as a stress test of small-model inference on minimal hardware, the project sets a floor that modern edge-AI claims must clear, and future ROM updates could enlarge the knowledge base if Edwards continues the work.

Tom's Hardware (Source)

Filed under

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

News editor covering money, funding and market shifts around Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,238 articles

Read also

« Previous article