Hot AI News

Nvidia’s Rubin HPC packs 50,000+ CUDA cores, but won’t ship until 2028

Nvidia announced Rubin HPC at SC26, a Vera Rubin GPU for simulation with more than 50,000 CUDA cores. It is expected in 2028, and FP64, memory and power figures were not given.

Nvidia’s Rubin HPC packs 50,000+ CUDA cores, but won’t ship until 2028
Image: Nvidia

Nvidia used the SC26 supercomputing conference to announce Rubin HPC, a new GPU in its Vera Rubin platform built for simulation and data processing. The company says it has more than 50,000 CUDA cores on a single GPU and is expected in 2028.

A GPU tuned for feeding cores, not just adding them

According to Nvidia, Rubin HPC is meant for science, engineering and analytics workloads. Besides the core count, which Nvidia calls a record, it lists a sparse data acceleration capability and seventh-generation NVLink for faster links between GPUs.

Nvidia's Rubin HPC packs 50,000+ CUDA cores, but won't ship until 2028
Image: Nvidia

Nvidia describes the design as balanced: data movement and compute scale together, instead of compute alone. Its reasoning is that many scientific codes, from climate models to genomic analysis and financial risk, repeatedly pull scattered values from a huge dataset, compute a result and repeat. Cores that wait on memory are wasted, so the company says the sparse acceleration keeps them supplied. It names sparse matrix operations, stencil computations, sorting and reductions as patterns that benefit.

The pitch is also framed around AI agents. Nvidia argues that as agents propose more candidate materials, molecules and designs, more simulations are needed to test them. Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, said: “For agents tackling science and engineering, faster simulation means faster discovery.”

Nvidia's Rubin HPC packs 50,000+ CUDA cores, but won't ship until 2028
Image: Nvidia

Rubin HPC is expected to arrive in 2028, with more than 50,000 CUDA cores on one GPU.

Advertisements

What the announcement leaves out

The post gives no figures for double-precision (FP64) performance, memory capacity or type, or power draw. It makes no direct comparison with Blackwell or with other Vera Rubin parts, and it names no customers or supercomputer sites. The core-count claim and the claimed benefits are Nvidia’s own and have not been independently checked.

The timing matters too. A 2028 date means no one can buy or test Rubin HPC for roughly two years, so the announcement works more as a roadmap signal for centers planning future procurement than as a product launch.

Who gets something from the software today

For researchers, the nearer-term story is software. Nvidia published a companion post on CUDA-X libraries, which it positions as the tools agents call for simulation and analytics. According to that post, a library called Daviator provides batched GPU kernels for the VASP materials code and delivers quantum-mechanical calculations up to 22x faster at up to 75% lower cost. Another, cuPotential, is said to speed up existing molecular simulation software by up to 3x, and Nvidia says it is working with the Schrödinger, AMBER, NAMD, GROMACS and LAMMPS communities to bring it to their engines.

Those numbers come from Nvidia’s own post and describe current hardware, not Rubin HPC. They do show where the company expects the new GPU to pay off: existing scientific codes, ported through libraries, rather than new applications written from scratch. That dependence on a software layer is the same lever explored in our explainer on CUDA and ROCm, where the software stack, not the chip alone, decides which hardware researchers can realistically use.

For developers and lab teams, the practical advice is limited. Code that already uses CUDA-X libraries is the most likely to benefit when the hardware arrives. Anyone choosing a platform for the next two years will have to decide on current-generation GPUs, since pricing, system configurations and cloud availability for Rubin HPC have not been announced.

The launch also lands at an unsettled moment for the company, which recently saw its chief executive step down, as we reported in our coverage of Jensen Huang’s resignation and the interim leadership. Nvidia’s post does not address how, if at all, that affects its HPC roadmap.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like