At Hot Chips 2026, Nvidia opened the hood of Vera, the Arm processor that will accompany the Rubin GPUs: 88 Olympus cores, eight LPDDR5X controllers and an assumed bias for the IPC.
During the processor session at Hot Chips 2026, Nvidia introduced Vera, successor to Grace lineage, who was equipping Grace Hopper generations and then Grace Blackwell. This is an in-house Arm server processor, built on a new heart microarchitecture called Olympus, designed to train with Rubin GPUs future Vera Rubin systems. The chip features 88 cores, a choice that translates a priority to core performance rather than running to the number of units, and surrounds itself with eight 128-bit LPDDR5X memory controllers offering up to 1.2TB/s bandwidth.
The NVLink-C2C interface is used to match Vera with the Rubin GPUs, but it remains open: any NVLink-C2C implementing component can connect to it, including a second Vera to build a purely CPU dual-processor machine. Nvidia also put forward a spatial multithreading approach, where resources are statically partitioned between threads instead of being shared: the displayed objective is to achieve the performance level of the classical SMT while reducing the interferences between threads that dispute the same execution units.
Nvidia's argument is based on the nature of the agentic loads, presented as a sequence of tasks involving several tools and several types of calculation, and not as a single request that could be optimized in isolation. The presentation followed the publication of the White Paper on Vera and Olympus the previous month.
Read the full article on the original source
Go to the source →