Who a GeForce RTX 4090 is actually for
This is flagship silicon, and flagship silicon has an awkward property: at 1080p it spends most of its time waiting. There is simply not enough pixel work at that resolution to keep 16,384 shading units busy, so the processor becomes the limit and a large part of what you paid for sits idle.
It earns its price at 4K, at high refresh rates, or with ray tracing turned up — workloads heavy enough to actually saturate it. If you play at 1080p on a 60 Hz panel, a card several tiers down will give you the same experience.
A narrow memory bus
Relative to its shader count, this card has less memory bandwidth than a balanced design would carry. Our model docks it roughly 9% for it. In practice that penalty grows with resolution: the card holds up at 1080p, where its cache absorbs most requests, and falls off faster than its specification suggests at 1440p and above.
Why it is not as fast as its shader count implies
With 16,384 shading units this is one of the widest dies in the range, and very wide dies do not scale linearly. The command processor and geometry front end run out of work to dispatch before the shader array runs out of capacity, so the last portion of that array contributes less than the first. Our model applies about a 15% correction for it — the same reason a card with 68% more shaders than its sibling is not 68% faster.
Is 24 GB enough?
At 1440p and high settings, 24 GB clears every title we model with room to spare. Memory is not the constraint on this card, and it is unlikely to become one within its useful life.
Where it sits
Ranked 2 of 126 desktop cards we model. Nearest faster: GeForce RTX 5090. Just below: GeForce RTX 5080, GeForce RTX 4080 Super. Full method on the methodology page.