Selected work
GPU · SystemsC++ · CUDA · TensorRT

Silky

60fps in, 240fps out — never leaving the GPU

A zero-copy video frame-interpolation pipeline: hardware decode, a TensorRT-optimized RIFE network, and hardware encode — with frames that never leave VRAM.

Role — Solo GPU / systems engineer
  • C++
  • CUDA
  • TensorRT
  • NVDEC / NVENC
  • ONNX

Silky decodes with NVDEC, runs RIFE optical-flow interpolation through TensorRT, and re-encodes with NVENC, keeping every frame on the GPU end to end. I wrote the pipeline orchestration, the NVRTC-compiled colour-space kernels, and the interpolator interface around the third-party model and NVIDIA libraries.

It targets 120, 144, or 240fps from 60fps source, in FP16 or FP32, encoding to H.264 or HEVC.

What's notable
  • Zero host round-trips — NVDEC → TensorRT → NVENC entirely in VRAM
  • NVRTC-compiled CUDA colour-space kernels
  • FP16 / FP32 precision, H.264 / HEVC output
  • 60 → 120 / 144 / 240fps targets