Wouldn't a more apt comparison with the XX90 series be the Titan series? I've always seen the XX90 series as more like the Titan series than the XX80ti (especially when in the Ampere generation we had an XX80ti.)
If you do the comparison using Titan GPU's, it makes sense why this was an incremental improvement generation for the 5090.
| GPU | Process Node |
Microarchitecture |
Release Date (Rounded to nearest month) |
3dmark (Timespy) | Difference | Point Gain per day | % Gain |
| Titan X (Maxwell) | 22N | Maxwell 2.0 | 3/2015 | 5721 | |||
| Titan X (Pascal) | 14N | Pascal | 8/2016 | 9565 | 519 | 7.40655106 | 67.19% |
| Titan Xp | 14N | Pascal | 4/2017 | 10213 | 243 | 2.666666667 | 6.77% |
| Titan V | 14N | Volta | 12/2017 | 12507 | 244 | 9.401639344 | 22.46% |
| Titan RTX | 14N | Turing | 12/2018 | 14306 | 365 | 4.928767123 | 14.38% |
| 3090 | 10N | Ampere | 9/2020 | 18178 | 640 | 6.05 | 27.07% |
| 4090 | 5N | Ada | 10/2022 | 30478 | 760 | 16.18421053 | 67.66% |
| 5090 | 5N | Blackwell | 2/2025 | 854 | 9.63 (Projected) | Projected 27% |
Edited: To include Titan X (Maxwell)
RTX 3090 -> RTX 4090 was a huge leap because 10N -> 5N was a huge jump. But as you can see, the Pascal -> Volta -> Turing leaps were much more moderate. A 27% increase is roughly in line with the Titan RTX -> 3090.
We're going to see things slow down like this because Moore's law is slowing down. That's why Nvidia is chasing low-hanging fruit (given that they also run a GPGPU compute business) like neural-rendering and ray-tracing acceleration. It's a lot easier than trying to innovate a paradigm-shifting technology that creates a new S-Curve in terms of hardware acceleration.
I do agree that the pricing (below the xx90 series) is bad though, even adjusted for inflation/costs of production.
Last edited by sc94597 - 9 hours ago
Essay Pro