Computer Hardware3 min
Nvidia quietly lowers the memory speed of its Rubin AI chip as shipments begin
Nvidia has posted its final shipping specifications for the Rubin data-centre GPU, ServeTheHome reports. Memory bandwidth drops from the 22 TB/s first promised to 19.2 TB/s, about 11% less. The NVLink chip-to-chip link drops from 3.6 TB/s to 3.0 TB/s, about 17% less. Networking to other servers rises from 0.4 TB/s to 0.45 TB/s. Compute stays at 50 PFLOPS for FP4 inference, and memory stays at 288GB of HBM4. The Vera Rubin NVL72 rack changes by the same amounts. Nvidia has not said why.
Why it matters: Memory bandwidth often limits how fast AI models run, so cloud customers renting Rubin systems will get somewhat less than first advertised. Nvidia had said its early figures could change.