RTX 5070 = gtx 950 in the product stack if comparing cuda cores, says GN. GPU shrinkflation.
i was just about to post the video.
GN is using cuda unit count normalized to largest/flagship gpu relative to price adjusted to inflation. while it is an ok measure, die size should probably be added.
any asic maker has a choice on die size vs yield per wafer. any of the wafer/die calculators can demonstrate the math. after fabbing, the dies get binned for performance/temp/power. then marketing creates model numbers/segments to set prices. but regardless of what name marketing slaps on the product it is only ever going to be one of some number of die sizes that you can usually count on one hand. adding a lower model for the dies without 100% of the compute units/clusters for a lower price is part of the process.
when it was consumer flagship parts getting harvested to go to professional/research skus the die size scheme was consistent enough. when ai/render/compute server products took over the largest die size the xx50/xx60/xx70/xx80 segmentation started to fall apart.
the true generational performance gains come out of process node size shrinks. the non shrink product cycles are all minor tweaks and software features. because nv had no competition from amd and crypto/covid/ai scarcity drove up pricing, people have not been rationally evaluating the parts and branding. the 2080/2070/2060 were the same die, with the gtx 1660 taking the place of what a historical 2060 segment/die size would have been. the 3070 was the next die size down from the 3080 rather than being a cut down 3080. but these model numbering shuffles were used to price them at the next tier up.
die size is the the biggest tell for a product cycle. you are either getting the biggest die/biggest die with flaw, 2nd biggest die/die with flaw, 3rd biggest, etc.
nv has used ti/super/whatever to obscure the fact you were paying more for what would have been called the next segment down historically. but no one cared when we pointed this out. now it is another non-shrink product cycle and nv hubris pricing is when people sort of noticed. if it werent for lack of supply, plenty of consumers would have happily bought the 5000 series at release.
moore's law is slowing and new node prices are going up, but not enough to justify the 5000 series pricing given it is on the same node as the 4000 series.