Can It Run AI?

Nintendo 64 — a transformer LLM on 4 MB of RAM

93.75 MHz NEC VR4300 (MIPS III) · 4 MB RDRAM · libdragon SDK · no GPU, no network, no cloud

Legend of Elya runs an 819,000-parameter byte-level transformer directly on the N64's CPU, generating Zelda-style dungeon dialogue at roughly 60 tokens per second (61.8 tok/s in the captured run below). All inference math executes on the VR4300; sampling entropy comes from the CPU's own cycle counter.

Source + ROM on GitHub Citable release (DOI) Deep-dive page

As discussed on Hacker News.

Sun Cobalt Qube 3 — the same brain on an AMD K6-2

AMD K6-2/450 · stock 2001 Cobalt Linux · gcc 2.95, C89 · the little blue cube

The exact same 819K transformer was ported to strict C89 and compiled with the Qube's original gcc 2.95 toolchain on its unmodified 2001 Cobalt Linux image. It generates at ~12 tokens per second on the K6-2 (64 tokens in 5.30 s CPU), verified logic-exact against the N64 build — same weights, same outputs.

The Qube does double duty: it also runs rustchain-vintage-x86, a single-file C89 miner written specifically for this machine, using RDTSC clock-drift fingerprinting and honest hypervisor self-reporting to attest that it really is old silicon — RustChain being the verification layer that tells real silicon from a VM (how proof-of-antiquity works).

Qube 3 port (in the N64 repo) rustchain-vintage-x86 miner

PowerPC G4 / G5 — llama.cpp on Tiger and Leopard

AltiVec SIMD · Mac OS X 10.4/10.5 · big-endian · Scottcjn/llama-cpp-tigerleopard

A working port of llama.cpp to Mac OS X Tiger and Leopard on PowerPC — modern quantized-model inference on machines Apple last updated two decades ago, with AltiVec doing the SIMD lifting. The AltiVec vec_perm attention-path technique behind the lab's PowerPC work is published as a citable paper (DOI 10.5281/zenodo.21282030).

Third-party coverage: Hackster.io News wrote it up as “The PowerPC Has Still Got It” (Nick Bild) — currently the strongest external verification of the port.

Port on GitHub vec_perm paper (DOI) Hackster.io coverage

IBM POWER8 S824 — 147 tokens/sec, CPU only

2 sockets · up to 128 SMT8 threads · 512 GB RAM · VSX/AltiVec · no GPU involved

The lab's POWER8 S824 runs TinyLlama 1.1B at 147.5 t/s prompt processing (8.8× the stock llama.cpp build) and ~18.9 t/s generation — CPU only — using cache-resident dcbt prefetch, NUMA-aware weight banking (ram-coffers), and hardware-entropy sampling from the POWER8 timebase. Both techniques are published: PSE hardware-entropy framework and RAM Coffers.

Raw logs (fresh capture, July 20 2026)

A signed re-run was captured on the machine as it is configured today: a para-virtualized PowerVM partition exposing 56 of the 128 threads (“POWER8 (architected)”, pHyp). On that partition, the same build measured 47.9 t/s pp128 / 8.9 t/s tg32:

| model                  |       size | backend | threads |  test |          t/s |
| llama 1B Q4_K - Medium | 636.18 MiB | CPU     |      64 | pp128 | 47.90 ± 5.67 |
| llama 1B Q4_K - Medium | 636.18 MiB | CPU     |      64 |  tg32 |  8.90 ± 0.21 |

llama-bench-tinyllama.log (sha256 11a002a0…6f74b3c) · system-identity.log (sha256 2689cf71…feb12d01)

ram-coffers on GitHub PSE paper (DOI) RAM Coffers paper (DOI)

Commodore Amiga — native compute & a software ecosystem

Motorola 68k · classic AmigaOS / AROS · FS-UAE emulation, self-declared

This card does not claim AI inference. What runs on the Amiga is a native m68k software stack: rustchain-amiga ships a C miner, an SDK, developer tools, ported utilities, a micro-JVM, and a complete bootable AROS distribution for classic AmigaOS. It runs inside FS-UAE emulation and is honest about it — the miner self-detects UAE and reports it truthfully to the network rather than pretending to be real hardware.

rustchain-amiga on GitHub

Submit Your Vintage Machine

Got a PDP, a SPARCstation, a BeBox, a 486, an Acorn — anything with a CPU and a story? Make it do real work, capture the evidence under the rules above, and submit it. The museum is open to outside exhibits:

github.com/Scottcjn/can-it-run-ai — submission guidelines and the exhibit registry.