Physically-based GPU rendering on the RTX 5080, proven one measurement at a time

A kernel we wrote is compiled to PTX at runtime by nvrtc and launched through the CUDA driver API — entirely in userspace. No kernel-mode driver, no firmware, VBIOS or NVRAM write exists anywhere in the path, so this cannot brick hardware.

The toolchain finding that made it possible: nvcc is unusable here (no MSVC host compiler), so the way onto the GPU is runtime compilation + a gcc-built host that dynamically loads nvcuda.dll.

Colour bleed is visible — the floor picks up red on the left, green on the right. But "visible" isn't proof, so the walls were neutralised to grey and the scene re-rendered. If the tint is really indirect light off the coloured walls, it must vanish.

Under uniform illumination, a surface must reflect exactly its albedo (0.90 here). More than that is energy creation — a bug. The test caught, then confirmed the fix of, a diffuse/specular coupling error, and isolated the remaining loss as the known single-scatter deficit.

deficit rough metal 0.435 = single-scatter GGX loss — carried to rung 04.

Volumetric random walk — the method Arnold and RenderMan use for skin, not a diffusion approximation. A purely-scattering sphere (no absorption) must be invisible in a furnace: every photon that enters has to exit. Measured over all 1,048,576 pixels, not a patch.

Unifying the rough-dielectric surface (rung 03) with the colored subsurface interior (rung 05) into a single material — which also eats the debt of keeping them separate. Per-channel, full-frame: a white version stays invisible; a reddish one must tint the whole frame red with no channel exceeding the background.

honest physically correct, but it reads as dark reddish wax — not flesh. No texture/pores, isotropic (not forward) scattering, optically thick, harsh light.

Real skin scatters light forward, not isotropically. A Henyey-Greenstein phase function passes the energy check — a forward-scattering white sphere stays invisible in the furnace, because the phase function is normalized. But I predicted it would transmit more backlight, and the full-frame measurement refuted that flat out.

honest forward scattering redistributes the glow directionally (visible) but does not raise net transmission at this optical thickness — the optimistic claim died on the number, which is the point of measuring.

Perturbing the surface normal with high-frequency noise adds micro-geometry — the specular coat highlight shatters into a stippled, pore-like pattern, and energy stays conserved (a bump changes direction, not light). But measuring it was a cautionary tale worth publishing.

honest the 512-spp "detail" was ~80% Monte-Carlo noise — a per-pixel gradient metric is invalid until the noise is suppressed. And it reads as leathery/orange-peel, not human pores.

The render kernels had grown four copies of the same vector / RNG / GGX / noise code. Extracted into one shared device header, handed to the runtime compiler as an include; the skin kernel dropped to 67 lines. A refactor must change nothing — and the furnace proves it, to the byte.

honest a false regression (isolated by holding a variable fixed) and a false improvement (measured) — both caught before shipping. The other three kernels still carry their copies; that's next.

Value noise gives broad dimples, but skin is cells. Driving the surface normal from a Worley cell-distance field instead reads — by eye — as a fine pore-cell stipple, nearer to skin than the earlier orange-peel. The honest catch is in the number.

honest an edge-count metric measures detail quantity, not pattern type — it can't tell "pore-like" from "bumpy." That blind spot is the case for a perceptual vision-model judge, which this stack still lacks.