cuda & gpu computing

getting a 2016 GTX 1080 to do things it was never sold to do. custom kernels, parallel-processing tricks, and squeezing every last FLOP out of Pascal-era hardware (sm_61).

what i've built

• emulation kernels — 256 Game Boy instances resident in VRAM (PokeAI)
• GPU-resident PPO — full RL training loop with tensors that never leave the card
• custom PPU renderer — scanline rendering on-device
• parallel optimization — batching, memory coalescing, occupancy tuning
• single-unit compilation — inlined device functions to kill duplicate-symbol errors

stack

compute: cuda (sm_61), cudnn
bridge: pybind11, python
hardware: EVGA GTX 1080 8GB, ~180W under load

the point

HBM and modern accelerators are priced out of reach for most people. i'd rather prove what a decade-old gaming card can still do than pretend compute has to cost thousands.