I spent a month with three production path tracers and a rack of very different hardware. The headline numbers you see in marketing decks are real, and also close to useless for planning a facility.
The metric that actually matters
For a studio, the question is not "how fast is one frame on the fastest card." It is "how many artist-hours and watt-hours does a sequence cost from first light to final." Those are different questions, and they favor different renderers.
The renderer that lost most single-frame benchmarks won my facility test outright, because it degraded gracefully out-of-core instead of crashing, and its denoiser needed fewer samples to hold up under motion.
A renderer that never crashes at 2 a.m. is faster than a renderer that is faster.
Nomish is right about the metric and too kind about the vendors. Graceful degradation is not a feature they ship — it is a feature we discover they lack, usually on a Friday.
Where the hardware is heading
- Memory bandwidth, not raw FLOPs, is the wall most production scenes hit first.
- Ray-tracing cores help, but only if your BVH build keeps up — otherwise you are just waiting differently.
- The next real gain is in learned importance sampling, and it is coming from the renderer teams, not the silicon.
If you take one thing from this: benchmark your show, not their demo scene.