# DLSS 5 Neural Rendering in ComfyUI on a Tesla V100 — no NGX, so the architecture gate isn't there — retained record

> Retained reddit record from reddit:comfyui.

Canonical: https://highsignal.app/data/records/e1d47f9ec2d4e785

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:comfyui)  published 2026-09-26 02:50 ingested 2026-09-27 02:40

[original source ↗](https://www.reddit.com/r/comfyui/comments/1wqfce3/dlss_5_neural_rendering_in_comfyui_on_a_tesla/)

## retained text



```
Every DLSS 5 NR node pack I came across runs NVIDIA's NGX runtime through D3D12, which cuts off at Turing. On a V100 (Volta, SM 7.0), \`CreateFeature(18)\` just hands back 0xBAD00001. No driver version fixes it. No Wine prefix either.



So I skipped NGX entirely. The network runs as plain PyTorch: no NGX, no D3D12, no ReShade, no Wine, just tensor code pointed at whatever \`torch\` device you've got. The architecture gate isn't bypassed. It's absent. Weight extraction is the only step that ever needed an NVIDIA GPU, and that part runs on CPU.



Here's what I measured on a Tesla V100-SXM2-16GB:

\- 1920x1080 — 0.700 s/frame, 1732 MB peak VRAM

\- 320x320 — 0.295 s/frame, 730 MB peak

\- roughly 0.83 GB per megapixel of output



I honestly expected FP8 to break everything. The network rounds activations to E4M3 at its publication points, and PyTorch only ships float8 kernels for Ada and newer. But here's the thing: it's a \*rounding\* step, not a matmul. The missing FP8 tensor cores don't actually matter. Works fine on SM 7.0.



Nodes available: DLSS5PyTorchModelLoader, DLSS5PyTorchEnhance (13 widgets), DLSS5PyTorchVideoEnhance, DLSS5PyTorchClearCache. They show up under \`DLSS 5/PyTorch (experimental)\`.



[https://github.com/lshan99q/comfyui-dlss5-pre-rtx](https://github.com/lshan99q/comfyui-dlss5-pre-rtx)



\*\*Being straight about what this is and isn't:\*\*



This is a reimplementation, not NVIDIA's binary. Quality is DLSS-5-\*like\*, not DLSS-5-\*equal\*. Parts of the weight layout are reconstructed from consumer evidence, and nobody's established numerical equivalence against NVIDIA's original output.



You supply your own \`nvngx\_dlssnr.dll\`. It's NVIDIA property, so it's not redistributed here. The installer takes it as a \`--dll\` argument.



Don't chain passes. The output converges rather than diverges, but what accumulates is a tone shift, not added detail. Tune \`intensity\` or \`local\_structure\_strength\` instead.



Credit where it's due: the ComfyUI node and PyTorch pipeline come from \`levzzz5154/ComfyUI-DLSS5-PyTorch\`, and the weight extractor plus recovered network definition from \`iamwavecut/MLX-DLSS\`. What I added is the pre-RTX packaging: Volta compatibility analysis and its verification, install/verify tooling, docs, and the performance numbers above.



Other Pascal/Maxwell cards and RTX 20-50 series should work but are untested on my end. And honestly, "pre-RTX" is kind of a misleading label. It's really a \*no-NGX\* path. Pre-RTX is just where it's your only option.
```



document fetched 2026-09-27 00:38 · canonical https://reddit.com/r/comfyui/comments/1wqfce3/dlss_5_neural_rendering_in_comfyui_on_a_tesla
