iiinova — where your digital self keeps evolving

Replies

Ant Stanley @iamstan.dev · 17/04/2026

This is a good post on the challenges implementing the GB200 platform. introl.com/blog/gb200-n... Take all of that, and double it. Then you have the new Vera Rubin powered NVL72 units.

introl.com

GB200 NVL72 Deployment: Managing 72 GPUs in Liquid-Cooled Configurations

Seventy-two GPUs operating as a single computational unit is now production reality. The GB200 NVL72 consumes 120 kilowatts and delivers 1.4 exaflops of AI compute in a single rack.¹ The architecture

120

Ant Stanley @iamstan.dev · 17/04/2026

The upside to the new platform having double the CPU/GPU density, cooling, power and physical infrastructure requirements, is that the performance is significantly more than doubled. Expected performance is somewhere around the 3.5x to 5x range if reports are to be believed.

152

Ant Stanley @iamstan.dev · 17/04/2026

Real world this means the performance per watt will improve significantly. One of these new Vera Rubin NVL72 racks will do the work 3 to 5 of the "older" GB200 NVL72 racks using 2 times the power. This means inference costs will drop, though don't expect that to be passed on to end users.

151