The Nvidia Blackwell architecture mostly rehashes Ada, using the same process node. Only the RTX 5090 stands out as a major (-ly expensive) upgrade.
Welcome to the modern era of Nvidia graphics cards, courtesy of the Blackwell architecture. Except, if we're being honest here — unlike Nvidia — not a whole helluvalot has changed architecturally with Blackwell relative to Ada Lovelace. We can sum up the major upgrades quite quickly.
First, Blackwell has native FP4 support on the tensor cores, which as of yet has only been used in a handful of applications, like a special version of AI image generation using Flux built into a special version of UL's Procyon benchmark. Blackwell also offers native FP6 support, which is sort of a hybrid between FP4 and FP8 that can potentially reduce memory requirements, but our understanding is that it's not really any faster than just using native FP8 operations.
Blackwell does offer some new features for ray tracing applications, but as with all other ray tracing tools, developer uptake can often be quite slow, particularly in regards to new games using the features. An enhanced triangle cluster intersection engine allows Mega Geometry (a new buzzword from Nvidia!) to better render massively complex scenes without bogging down. It feels a bit to me like the old over-tessellation approach where Nvidia got some games (like Crysis 2 or 3, IIRC) to utilize massive amounts of tesselation... on flat surfaces! Why? Because AMD's GPUs struggled mightily with the workload so it made Nvidia's GPUs look better.
