The simplest reason is blackwell was engineered for improvements in AI and not gaming. That brought a bunch of unintended effects and lack of or even regressive performance. To help combat that nvidia increased the cores on all the gpus which increased heat and watts. This unit was printed on the same nm so no improvments in transistor count. Ill list a few key points >Sacrificed High-Precision Math: To maximize AI math throughput (FP8/FP4), Blackwell sacrifices some traditional 64-bit (FP64) and even some FP32 floating-point compute resources used in traditional gaming rasterization, leading to smaller performance gains in traditional rendering. >Minimal IPC Gain: In rasterized rendering, the raw Instruction Per Clock (IPC) advantage of Blackwell over Ada Lovelace is reported to be quite low, sometimes averaging around only 1% improvement when normalized for clock speed. >Similar Shader Pipelines: The underlying shader architecture did not receive a revolutionary change in how it handles traditional vertex or pixel shading, focusing instead on Tensor Core advancements for AI-driven features like DLSS. >Cache Contention: Blackwell uses a monolithic 65 MB L2 cache, which is larger, but studies indicate it can increase contention and access latency compared to the fragmented L2 cache architecture of its predecessor, penalizing some gaming workloads. >AI Management Processor (AMP) Issues: There are reports suggesting a new AI management logic unit in Blackwell may be causing performance inconsistencies. If not properly utilized or supported by game engines, this unit can lead to "broken" performance, where high-end cards lose to previous-gen models.