[BOUNTY] N64 LLM Speedrun — 5 tok/s on Real N64 Hardware (300 RTC) #1876
Description
N64 LLM Speedrun — 300 RTC
The challenge: Make legend-of-elya-n64 hit 5+ tokens/sec on real Nintendo 64 hardware.
Hardware Constraints
- CPU: NEC VR4300 (MIPS III) @ 93.75 MHz
- RAM: 4 MB RDRAM (8 MB with Expansion Pak)
- RSP: Reality Signal Processor — 8x 16-bit SIMD lanes
Requirements
- Achieve 5+ tok/s sustained inference on the 819K-parameter transformer
- Must use RSP SIMD for matrix multiplication (not just the main CPU)
- Video proof on real N64 hardware (controller input visible, CRT/TV output)
- Submit PR to legend-of-elya-n64
- Include benchmark numbers in PR description
Bonus: +100 RTC
If inference runs from a physical N64 cartridge (flash cart or EPROM), not just via USB upload to dev kit. Show the cartridge in the video.
Why This Matters
The N64 VR4300 is a MIPS III processor from 1996. If a transformer can run at usable speed on 30-year-old console silicon, it proves that LLM inference belongs everywhere — not just in data centers. This is the most constrained inference target in the world.
Approaches to Explore
- RSP microcode for 8-wide SIMD matmul (the RSP is basically an embedded DSP)
- Aggressive quantization below Q4 (Q2? Binary weights?)
- Vocabulary pruning to reduce embedding table size
- Loop unrolling and MIPS III instruction scheduling
- DMA pipeline: overlap RSP compute with RDRAM fetches
Resources
- libdragon SDK — Modern N64 development
- N64 RSP documentation — Vector unit reference
- awesome-n64-development — Dev resources
Reward: 300 RTC (+ 100 RTC cartridge bonus)
Wallet: Provide your RTC wallet address in the PR
Part of the Elyan Labs retro compute initiative. See also: RustChain proof-of-antiquity mining with vintage hardware bonuses.