NVIDIA used the SC26 supercomputing conference to announce a new member of its Rubin GPU family. Called Rubin HPC, the chip is aimed at scientific simulation and what NVIDIA calls agentic HPC workloads, in which AI agents run and steer large computing jobs. It is expected to arrive in 2028.
A record core count
NVIDIA says Rubin HPC will have “more than 50,000” CUDA cores, the most of any NVIDIA GPU so far. For comparison, Tom’s Hardware notes that the current Rubin GPU has 224 streaming multiprocessors and 28,672 CUDA cores, so the new chip would almost double that number.
How NVIDIA reaches that figure has not been confirmed. Tom’s Hardware suggests a four-die design similar to Rubin Ultra, but NVIDIA has not said so.
NVLink 7
The GPU will connect to other chips using NVLink 7, the next generation of NVIDIA’s interconnect. NVIDIA did not share bandwidth figures. Fast links between GPUs are critical in supercomputers, where thousands of chips work on a single problem.
What we do not know yet
Several important details are still missing: memory capacity and bandwidth, and whether NVIDIA has increased double-precision (FP64) performance, which matters for traditional scientific simulation. Recent NVIDIA designs have focused on lower-precision AI math, so the FP64 question is important for research centers.
Context: a slow year for new supercomputers
The announcement came as the November 2026 TOP500 list showed only 27 new systems, the lowest turnover in the list’s history according to The Next Platform. One explanation is chip roadmap timing: buyers may be waiting for the next generation of hardware, including chips like this one.
Why it matters
Rubin HPC is still two years away, and its specifications are partly unknown. But it shows that NVIDIA wants a GPU tuned for supercomputing and not only for AI data centers. Expect more details as 2028 gets closer.
Sources:
