MLX-DLSS: Hacker runs NVIDIA's DLSS 5 neural renderer on a Mac, within 0.005 MAE of the original
WaveCut · reddit · 2026-09-05
Reddit user WaveCut open-sourced MLX-DLSS, extracting NVIDIA's two DLSS neural networks — the DLSS 5 renderer and the frame generator — and getting them to run on Apple Silicon via MLX/Metal and anywhere via PyTorch.
- Accuracy: the renderer is within 0.005 MAE of the original DLL; the frame generator matches the library at 59.9 dB and stays within 0.03 dB on full clips.
- Performance: on an M2 Max, frame generation takes 6 ms per 540p frame; the renderer does 11 fps at 512×448.
- The library runs in FP8 and half precision with approximate softmax and a hash-based noise generator, so a single ulp of difference shows up in the image; recovery required matching rounding against per-kernel captures.
- No proprietary code is included: the weight tool extracts tensors from the user's own nvngxdlssnr.dll / libnvidia-ngx-dlssg.so.
- DLSS Super Resolution was skipped: without engine motion vectors it loses to plain Lanczos.
More from Infra
- Anthropic IPO investors press for revenue per token and per gigawatt metrics — Hesamation · 2026-09-05
- Tauri-Based Wan2GP Desktop Launcher Brings One-Click Open-Source Video Gen to Windows — Extension_Affect_483 · 2026-09-05
- ESP32 voice assistant: AIMET AdaRound cuts wake-word model 8x without the 3.7% accuracy hit of naive 4-bit — carrycooldude · 2026-09-05
- Qwen3.8-27B on 2× RTX 5070 Ti: One vLLM Flag Kills Speed 11×, Full Benchmarks Inside — puthre · 2026-09-05
- Dev boosts GLM 5.2 TPS on a B300 and swaps it into Claude Code in place of Anthropic models — abhijithneil · 2026-09-05
- What forces LLM teams to optimize inference when going from MVP to production? — Ok_Philosophy_4031 · 2026-09-05