
Nvidia Teases Rubin HPC GPU: 50,000+ CUDA Cores, Ships 2028 - AI Weekly
Nvidia has unveiled its upcoming Rubin HPC GPU featuring more than 50,000 CUDA cores, with the chip expected to ship in 2028.
Nvidia used Supercomputing 2026 to announce the Rubin HPC GPU, which will feature 'more than 50,000' CUDA cores. The company describes it as a 'balanced design' that pairs increased compute with faster node-to-node data movement using NVLink 7, maintaining the 3.6 TB/s per-GPU bandwidth of NVLink 6. The design likely uses a four-die layout similar to Rubin Ultra, connected by NVLink-C2C.
Missing Specifications
Notably absent from the announcement are critical specifications that HPC buyers scrutinize most, including FP64 performance, memory bandwidth, and HBM capacity. These metrics are essential for procurement contracts, and competitors are promoting exactly these details at the same conference. This pattern of headline core counts arriving before procurement-relevant specs has been observed across 288 Nvidia stories tracked over the past 90 days.
Faster simulation means faster discovery.
Performance Context
The announcement comes with context about current supercomputing performance. Top machines achieve only 3% of their theoretical peak on sparse HPCG workloads, which are most relevant to Rubin HPC's target applications. Meanwhile, the Vera CPU rack, featuring 192 CPUs, approximately 300TB of memory, and 1PB of storage, is scheduled to arrive in early 2027—a full year before Rubin HPC ships.