Vera CPU and Rubin GPUs designed to operate within 600kW-class rack power envelopes.
On stage at GTC on Tuesday, CEO Jensen Huang unveiled Nvidia's next-generation of high-end datacenter architectures: Vera, a CPU architecture; Rubin, a GPU architecture; and Rubin Ultra, a refreshed version of Rubin. Rubin Ultra is designed so that 576 GPU dies can be crammed into a single rack consuming 600 kW of power, with full deployment expected in late 2027. However, Vera CPU cores and Rubin GPUs will arrive first. Named after American astronomer Vera Rubin, known for her research into dark matter, Vera represents Nvidia's first Arm-compatible CPU architecture since Grace was announced in 2021.
The Vera CPU will feature 88 custom-designed Arm Cores — replacing Neoverse — with SMT pushing the thread count to 176 per socket upon its arrival late next year. Like Grace, the chip will feature integrated NVLink chip-to-chip connectivity for interfacing with Nvidia's upcoming Rubin GPUs. Rubin borrows heavily from Blackwell's design architecture, featuring two reticle-limited dies capable of up to 50 petaFLOPS at FP4 precision with 288 GB of HBM4 memory good for 13 TB/s of bandwidth. The parts will be packaged as a Superchip and deployed in Nvidia's rackscale NVL144 chassis. Unlike Blackwell and Blackwell Ultra, Nvidia is counting Rubin's twin dies as two separate GPUs on a package rather than a single logical chip.
The Vera-Rubin NVL144 will deliver 3.3x higher floating point performance compared to the GB300 NVL72, topping 3.6 exaFLOPS of dense FP4 for inference and 1.2 exaFLOPS of FP8 compute for training. The system will feature Nvidia's 6th-gen NVLink switch fabric providing an aggregate 260 TB/s (1.8 TB/s per die) of interconnect bandwidth and will utilize its upcoming 1.6 Tbps ConnectX-9 NICs. When Rubin Ultra arrives in late 2027, it will double the number of GPU dies and HBM modules to four and 16 respectively. Each Rubin Ultra package is expected to top 100 petaFLOPS of FP4 performance and cram 1 terabyte of faster HBM4e memory.
The rack-scale system will contain 144 Rubin Ultra packages along with an unspecified number of Vera CPUs, rated for 600 kW of power consumption and thermal output. In total, the system is expected to deliver 15 exaFLOPS of FP4 inference performance and 5 exaFLOPS of FP8 for training. Nvidia plans to transition to faster NVLink7 interconnects for chip-to-chip communications, but will stick with 1.6 Tbps ConnectX-9 NICs for scale-out communications. However, a significant challenge remains: few existing datacenter facilities will be able to support such a dense configuration. While HPC vendors like Cray offer EX4000-series cabinets capable of supporting up to 300 kW, those systems are not typical 19-inch racks and are usually designed around custom datacenter shells. Cooling 600 kW of compute capacity in what appears to be a 19 or 21-inch OCP open rack form factor will almost certainly require a custom build.
Nvidia's datacenter roadmap extends beyond Rubin Ultra, with the next GPU architecture named after American theoretical physicist Richard Feynman slated for 2028.