Blackwell
NVIDIA announced a generation of accelerators in which two dies of 104 billion transistors each work as one chip, joined by a 10-terabyte-per-second link.
Why it matters
Manufacturing came up against the physical limit on die size, and the answer was to glue dies together rather than make them larger.
A single die is limited by the field of the lithography scanner, and Hopper at 80 billion transistors was already close to it. Blackwell goes around the limit: two dies in one package, coupled tightly enough that software sees a single device. That made scaling compute a question of packaging and cooling rather than lithography alone. One accelerator cost tens of thousands of dollars, and the queue for them decided who could train frontier models at all.