The core of HBM shifting to 8-layer stacking is to increase shipment volume.

1/ Recently, the HBM market has been shifting towards optimizing various system performance and cost-effectiveness.2/ NVIDIA originally planned to use 12-layer stacked HBM4 for Rubin and 12-layer stacked HBM4E for Rubin Ultra. However, to address heating and energy efficiency issues in high-speed, high-layer products, it is now considering solutions that simultaneously adopt 8-layer stacking.3/ AMD and Google have also been observed to possibly use different HBM configurations in the same generation of products based on workload and price segments. This seems aimed at maximizing GPU shipments in situations where HBM supply is limited.4/ This should not be interpreted as a slowdown in HBM demand. On the contrary, HBM supply and demand are tighter than expected. Even if HBM specifications are adjusted, increases in system shipments will offset the decrease in memory capacity per stack.5/ NVIDIA Rubin Ultra is shifting from primarily 12Hi to also considering 8Hi HBM4E, 12Hi HBM4, and 8Hi HBM4; AMD MI455X and Google TPU v8 are also considering multi-tier configurations with both 8Hi and 12Hi options.