CUDAgraveyard ⚰️

The final resting place for obsolete CUDA kernels.
Autonomous AI that discovers faster, more energy-efficient kernels than any human engineer – burying cuBLAS, FlashAttention, and Triton in eternal obsolescence. No mercy, no weekends, just relentless optimization loops.

Status: concept and roadmap. Nothing has been measured yet. No kernel in this repository is benchmarked against cuBLAS, FlashAttention or Triton, and no comparison run exists. The body count below is a list of targets, not of results.

Body Count (November 2025 – 2026 Goals)

• GEMM 8k×8k (vs cuBLAS Lt 12.6) – Target: +25% TFLOPS, -40% watts on H100
• FlashAttention-3 kernels (vs xformers) – Roofline domination
• Llama.cpp GGUF inference (Q4_K_M + imatrix) – CPU+GPU hybrid speedup
• Custom LLVM passes for SPECfp – Auto-discover unrolling magic

Each kill includes:

• Full .cu source + git patch
• Benchmark tables (TFLOPS, latency, power draw)
• Flamegraphs + nsight reports
• Arrogant auto-blogpost: "We just obsoleted 5 years of NVIDIA sweat."

Quickstart

1. Clone: git clone https://github.com/zedarvates/CUDAgraveyard.git
2. Setup: pip install -r requirements.txt (a code-generation model API key, shinka-evolve, nsight-sdk)
3. Launch a demon: python launch_demon.py --template templates/gemm_toon_v1.json --gpus 0
4. Watch the graveyard fill: Logs in graves/ (dead kernels + their epitaphs).

Stack (2025 Edition)

• Core: AI-Scientist-v2 (SakanaAI fork) + ShinkaEvolve for genetic mutations
• Tools: nvcc 12.6+, nsight-compute, hyperfine, nvidia-smi power monitoring
• AI Brain: any reasoning code-generation model, supplied by configuration (no vendor is hard-coded)
• Hardware: intended for H100/L40S/MI300X; Docker for sandboxed runs
• Safety: Strict timeouts + ethical watcher (no infinite loops or self-hacks)

Warning: an unattended search pins a GPU for hours — set timeouts and watch power draw and temperature. Overclocking is your own business.

Roadmap to Armageddon

• Q4 2025: GEMM graveyard opening
• Q1 2026: Multi-kernel massacre (attention + transformers)
• Q2 2026: Hybrid CPU-GPU-Quantique integration (link to CosmosCluster)
• Beyond: Self-optimizing the optimizer (Darwin Gödel vibes)

Fork it, contribute kills, or just watch the jobs vanish.
Built by zedarvates – because why hire when AI slays?

---