GeForce · Kepler → Blackwell · 2012–2026

Fourteen years of learning not to draw the pixels

Every generation shipped one headline idea, and most sat unused for years before a game needed them. Meanwhile the games people actually log hours in stopped caring about graphics cards somewhere around 2016. Both halves of that are on this page.

The readout

drag the year

What shipped, what it meant, and where the pixels came from. The two bars are the point: the top one is what an enthusiast chasing every feature could run, the bottom is a typical mainstream setup in the same year. They only start diverging once upscaling arrives.

2012
20122016202020232026
Architecture
Kepler
What the press said to buy
GTX 680
What the year's games needed
GTX 680
Where one frame's pixels come from
Enthusiast, 4K, every feature on
Typical setup, 1080p, quality preset
Rendered by the shaders Reconstructed by upscaling Invented between frames
What shipped
What people were playing
Full entry for this generation

What "23 of every 24 pixels" actually looks like

NVIDIA's own figure for DLSS 4.5. Each square is a pixel on your screen; the dark one is the only square the graphics card genuinely drew.

Left: 2016, everything rendered. Right: 2026 with upscaling and 6× frame generation. The middle is where most people actually sit.

The dormancy gap

shipped ≠ useful

Faint bar: the feature exists on the die and nothing uses it. Solid bar: games depend on it. The tick above each solid bar names the game that ended the wait.

Why one ray is expensive

A ray has to find what it hits. Testing it against every triangle in a scene is hopeless, so the scene is wrapped in nested boxes — a bounding volume hierarchy — and the ray walks down the tree, discarding whole branches. RT cores do that walk in fixed-function hardware instead of in shader code.

Roughly a dozen box tests instead of thousands of triangle tests, per ray, several rays per pixel, every frame.

What frame generation does to latency

The argument in one picture. Generated frames make the image smoother without making the game respond any sooner — the inserted frame is built from two frames the GPU already has, so it can only be shown after the second one is ready.

Double the frames on screen, and your input still waits for the next real one. Reflex claws some of that back, which is why NVIDIA requires it.

The 2026 price distortion

same silicon, new price

No new consumer generation shipped this year. These cards are physically what they were in 2025. Bars are shown as a multiple of launch price so the mid-range doesn't vanish next to the 5090 — the marker is where each card sat in June, two months before the August reading.

launch price1.5×2.5×
Launch MSRP Median listing, August 2026 Where it sat in June

Which card runs what

the part that answers the question

Not a timeline — a ladder, ordered by rasterization performance, starting below the PC entirely. A bar starts where a game becomes playable and goes solid where it's comfortable. What happens after that is the interesting part, and it's three different things.

How running out of VRAM actually fails

Not as a slower card. Average framerate can look fine while the 1% lows collapse, textures pop in late, and GPU utilisation drops — because the card is waiting on the bus rather than working. That last part is why people misdiagnose it as a slow GPU and buy the wrong upgrade.

Same card, same scene, one setting more texture detail than the frame buffer holds.

Resolution
Frame rate target

Where it stops scaling
Playable, settings down Comfortable CPU-bound — a faster processor helps, a faster GPU doesn't Capped — nothing helps

The two halves don't meet

Fourteen years of silicon went into ray tracing, tensor cores, mesh shaders and neural rendering. Those features are real, they took five years each to matter, and they now define a set of maybe thirty games.

The games with the most hours played on PC — the ones at the top of the ladder above — were finished with your graphics card around 2016. What limits them is a processor, a frame cap, a network, or an anti-cheat policy. Almost never the GPU.

Both facts are worth holding at once. The feature race produced genuinely better-looking games and genuinely better upscaling, which every card back to 2018 got for free. It also produced a market where the mid-range costs double what it did last year, sold on capabilities most people's libraries will never ask for.

Glossary

every term on this page

Any dotted-underlined word above opens its definition where you are standing. This is the same set, grouped by what kind of thing it is rather than alphabetically, because the category is usually what you actually want.