Nvidia has unveiled Rubin HPC, a new graphics processor built for scientific simulation rather than AI training, with more than 50,000 CUDA cores on a single chip. The company announced it at the SC26 supercomputing conference in Chicago, with availability expected in 2028.
What Rubin HPC Is
Rubin HPC is a new member of Nvidia’s Vera Rubin platform, aimed at simulation and data processing for science, engineering and analytics. It supports precisions from FP64 to FP4, adds sparse-data acceleration and connects graphics processors over seventh-generation NVLink interconnect technology.
The headline figure is the core count: more than 50,000 CUDA cores, which Nvidia calls a record for one of its GPUs and about three times the count of its Hopper generation. The chip is designed so those cores spend their time computing, not waiting for data.
Rubin HPC arrived as part of a wider SC26 package. Nvidia also announced Nemotron for Science, a set of open models and tools for researchers, and expanded Vera CPU racks for high-performance computing, due from Dell, HPE, Lenovo and Supermicro in early 2027.
Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, said faster simulation means faster discovery for scientists working with AI agents. Agents test ideas through simulation, he said, and Rubin HPC shortens the loop between a hypothesis and a result.
What Nvidia Did Not Say
The announcement carries no memory specification, no bandwidth figure, no power draw and no price. Nvidia has also named no launch partners or customers, and the chip’s double-precision performance — the figure scientific buyers watch first — remains unpublished for now.
That matters because the workloads Rubin HPC targets — sparse matrix operations, stencil computations, sorting and reductions — wait on memory more than on arithmetic. On the HPCG benchmark, which models those patterns, even the fastest supercomputers reach only a small fraction of their peak speed.
Our Read
Rubin HPC is Nvidia stretching its Rubin franchise from AI training into scientific computing, a market where AMD and custom silicon are gaining ground. The pitch is not raw speed but balance: cores, memory movement and interconnect scaled together, with agents driving the simulations.
The 2028 date is the honest part of the story. Rubin HPC arrives a year or more after the parts that will define the 2027 exascale procurement round, and its specifications can still change. Until independent groups measure real kernels, 50,000 cores is a count, not a forecast of speed.
The competitive context is AMD, whose MI450X accelerator is positioned as a direct challenge to Vera Rubin. Nvidia’s answer is a purpose-built science chip; AMD’s is an open ecosystem. Buyers planning 2028 systems now have two distinct bets to weigh.
Why It Matters
If Rubin HPC delivers, it gives Nvidia a foothold in the simulation budgets of national labs, climate centres and engineering firms, not just AI training clusters. That diversifies a business that currently depends on a handful of hyperscalers buying training hardware.
For researchers, the promise is shorter paths from idea to discovery, with agents running more simulations and testing more candidates. For Nvidia, the risk is that a chip announced two years early, without a full specification sheet, is judged on its core count alone.
Background
Nvidia’s standard Rubin GPU, with 28,672 CUDA cores, is already in production, and the company reported record quarterly revenue of $115.8 billion this week. The same week brought turbulence: Nvidia delayed its 10-Q filing, and regulators are probing its China chip sales.
Source: NVIDIA’s announcement


0 Comments