Team Ai
Apppublic

Agnuxo/chimera-benchmarks

sourceHugging Facemitupdated 5mo agoView on Hugging Face
0likes
App README

CHIMERA — GPU Compute Shader Benchmarks

Experimental GPU compute shader architecture for neural network acceleration via OpenGL. Public benchmark dashboard tracking research progress.

⚠️ Research prototype — not production-ready. Benchmarks are self-reported on RTX 3090 hardware. Independent reproduction welcome.

Benchmarks Tracked

  • —Matrix multiplication (2048x2048) — OpenGL compute vs CUDA baseline
  • —Self-attention simulation via cellular automata shaders
  • —Memory footprint comparison (OpenGL vs PyTorch/CUDA)
  • —Multi-GPU vendor testing (Intel, AMD, NVIDIA, Apple Silicon)

Architecture (experimental)

  • —Text encoded as GPU textures processed by fragment shaders
  • —Cellular automata evolution for recurrent computation
  • —Holographic memory encoding via single-pass GPU correlation
  • —No CPU/RAM in hot path — pure GPU pipeline (in theory)

Links

Reproducibility

Benchmarks run via demo_pure.py on RTX 3090. Results vary by GPU, driver version, and OpenGL implementation. Docker-based reproducible benchmark suite planned.