Path Tracing
Path tracing is the Monte Carlo global-illumination algorithm (Kajiya, 1986): many rays per pixel, multiple light bounces, solving the full rendering equation -- vs. rasterization (zero rays, 120fps+ on mid-tier GPUs) and hybrid ray tracing (1-2 bounces, direct light/reflections only, ~1 ray/pixel). CD Projekt Red's Cyberpunk 2077 "RT Overdrive" (patch 1.62, Apr 2023) is the reference real-time build: RTX 4090 native 4K = 22.4fps, DLSS3 Quality+Frame Gen = 103.5fps. On NVIDIA's current flagship RTX 5090 ($1,999 MSRP, Jan 2025, Blackwell, 4th-gen RT cores): native 4K path tracing ~55-59fps, DLSS4 Quality ~70fps, DLSS4 Multi-Frame-Gen ~200-290fps (Quality-to-Performance mode; NVIDIA's own marketing figure is 290fps). Unreal Engine 5.8 (June 2026) ships Lumen Lite (~2x faster than Lumen High) plus a reference Path Tracer with NFOR denoising. Unity HDRP's path tracer accumulates samples per pixel, restarting on camera move. Current version/status: real-time viable on RTX 40/50-series w/ DLSS4 (as of mid-2026); algorithm dates to Kajiya, 1986. Maintained by No single owner -- rendering technique implemented by NVIDIA (RTX/DLSS), Epic (UE5 Path Tracer/Lumen), Unity (HDRP), and offline renderers (Blender Cycles, Arnold).
Pick path tracing for pre-baked cinematics, product viz, or archviz stills where minutes-per-frame is fine and physical accuracy sells the shot -- that's Blender Cycles or Arnold territory, not a game engine. Shipping a real-time title? Skip full path tracing unless your player owns an RTX 4080+ -- use Lumen (UE5) or baked GI instead. Even Cyberpunk 2077's own RT Overdrive needs DLSS4 Frame Gen to clear 100fps on a $1,999 RTX 5090; native 4K sits at 55-59fps on the newest flagship. NVIDIA's own docs admit the real trick: at ~1 ray/pixel the image is unusably noisy without ReSTIR + Ray Reconstruction AI denoising doing most of the "path tracing" work. Known weakness: At the ~1 ray/pixel budget real-time engines can afford, raw path-traced output is far too noisy to ship -- Cyberpunk 2077 alone stacks five separate AI denoisers per frame to hide it, meaning the "accuracy" is largely reconstructed, not sampled.