NVIDIA GeForce RTX 4090

The RTX 4090's strongest measured trait is ray tracing retention at 71% of raster, which puts it in the 98th percentile — that metric describes how much of its raster frame rate survives with ray tracing on. Its weakest is frames per $100 at 9.1 fps/$100 (14th percentile). That is a wide spread for a single card: it is specialised rather than balanced, and whether it suits you depends on which of the two you actually load.

The RTX 4090 turns a 450W board power budget into 146.1 fps at 1440p, or 0.32 fps/W against a pool median of 0.32 fps/W. Power, not transistor count, is the binding constraint on a modern GPU — Dennard scaling stopped holding around 2005, and Esmaeilzadeh and colleagues showed at ISCA 2011 that the fraction of a chip that can be powered at full speed shrinks with every node (“dark silicon”). Frames per watt is the shortest honest summary of what an architecture is doing with that budget.

For scale: the closest NVIDIA card from an earlier generation is the NVIDIA GeForce RTX 3090 (2020), at 106 fps and 350W. The RTX 4090 is 38% faster and draws 100W more.

Every rendered pixel has to be fed from memory. At 1440p the RTX 4090 has 1,872 B/px of memory bandwidth available per pixel it draws, against a median of 1,808 B/px — roughly in balance for its performance class. This is the roofline model applied to a game frame: Williams, Waterman and Patterson (Communications of the ACM, 2009) showed that any workload runs into either a compute ceiling or a memory-bandwidth ceiling, and that the ratio of the two is what decides which one you hit.

On memory capacity, the RTX 4090 carries 24GB against a median of 16GB across the 3 other cards within 10% of its 1440p frame rate. It carries more memory than its performance class typically ships with, which mostly matters for high-resolution textures, local AI models, and viewport-heavy creative work rather than frame rate.

Method: percentiles are computed over the 77 non-legacy cards with a 1440p figure. Frame rates come from Tom's Hardware GPU hierarchy; specifications from manufacturer datasheets. Theoretical FP32 TFLOPS is deliberately not used as a cross-vendor comparison — RDNA 3 and 4 dual-issue FP32 doubles the paper figure without doubling game performance, and AMD itself declined to count it as extra shaders. The measured Steel Nomad score is used as the compute proxy instead. GPURanker does not run its own benchmarks.