Nvidia unveils Rubin HPC GPU with 50,000 CUDA cores for 2028
The new accelerator is designed to speed up complex scientific simulations and data analysis for researchers and engineers.
The short version
- Rubin HPC features more than 50,000 CUDA cores, nearly three times the amount found in the GeForce RTX 5090.
- The hardware is specifically designed for scientific simulations and data analysis rather than consumer gaming.
- It includes sparse data acceleration to improve calculations that retrieve limited info from large datasets.
- The chip will join the Vera Rubin platform and utilize seventh-generation NVLink technology for GPU communication.
- Nvidia expects the hardware to become available in 2028.
Why it matters
This hardware will allow scientists to run more complex simulations faster, potentially shortening the time it takes to discover new materials or climate solutions. It marks a shift toward specialized hardware that helps AI agents explore more possibilities before physical testing begins.
A new tool for science
Nvidia recently unveiled the Rubin HPC GPU, a specialized accelerator built for high-performance computing. Unlike the graphics cards used for video games, this hardware is intended for heavy-duty scientific simulations and data analysis. The company expects the chip to be available for researchers and industrial users in 2028.
As reported by Mezha.net, the Rubin HPC will be part of Nvidia’s Vera Rubin platform. It is designed to handle technical workloads that require massive amounts of processing power. By focusing on these specific tasks, Nvidia aims to provide a dedicated tool for complex engineering and scientific processes.
The hardware is intended to support the growing use of AI agents in research. These software agents can propose different candidates for materials or designs, and the Rubin HPC will help run the simulations needed to test those ideas. This allows teams to assess possibilities digitally before they commit resources to physical prototypes or laboratory experiments.
How the hardware works
The Rubin HPC is built with more than 50,000 CUDA cores in a single GPU. To put this in perspective, this is almost three times the number of cores found in the GeForce RTX 5090. These cores are the fundamental processing units that perform the math required for complex simulations.
Nvidia designed the chip to balance these numerous processing cores with faster data movement between components. This specific balance is meant to prevent the processing units from sitting idle while they wait for information during data-intensive calculations. Efficient data movement is critical when handling the massive datasets common in scientific research.
The chip also features sparse data acceleration to improve specific types of calculations. This technology helps the system retrieve limited information from much larger datasets more efficiently. It is particularly useful for tasks involving sparse matrices, stencil calculations, sorting, and reduction tasks, which are common in engineering software.
Connectivity and software
To work effectively in large computing systems, the Rubin HPC uses seventh-generation Nvidia NVLink technology. This system is designed to improve communication between multiple GPUs working together. Better communication allows for larger, more complex systems to function as a single cohesive unit for massive workloads.
Nvidia also connects this new hardware to its existing CUDA-X libraries. These libraries provide optimized software implementations of algorithms used in scientific research and data analysis. By combining the hardware with these software tools, Nvidia aims to accelerate calculations initiated by AI agents and other scientific applications.
While the hardware provides the raw power, the actual improvements in speed will depend on several factors. These include the specific software being used, the overall system configuration, and the measured performance of the specific workload. Nvidia has not yet published official throughput figures for this specific implementation.
Impact on research and discovery
The primary goal of the Rubin HPC is to help scientists and engineers explore more possibilities. By speeding up simulations, researchers can examine a wider variety of designs and materials in a shorter amount of time. This could lead to faster breakthroughs in fields like climate systems and industrial design.
Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA, noted that "For agents tackling science and engineering, faster simulation means faster discovery." He explained that the chip helps agents test ideas and guide their next steps, helping researchers bring breakthroughs closer to reality.
Because the hardware can handle complex processes like simulating molecules or materials, it serves as a bridge between theory and physical reality. It allows for a more iterative process where many digital failures can occur quickly, leaving only the most promising candidates for real-world testing. This reduces the time and cost associated with traditional laboratory experiments.
What to watch for next
Since the Rubin HPC is not expected to launch until 2028, the next few years will involve significant development. Researchers will need to adapt their software and workflows to take advantage of the new CUDA cores and sparse data acceleration features. The industry will be watching to see how these tools integrate with existing AI agent frameworks.
Independent testing will eventually be required to establish how much faster specific workloads will actually run on the new hardware. Until those tests are conducted, the practical performance claims remain theoretical. The tech community will look for these benchmarks as the 2028 launch date approaches.
Additionally, the rollout of the Vera Rubin platform will be a key indicator of how Nvidia plans to scale its high-performance computing offerings. The success of the Rubin HPC will likely determine how the company approaches the next generation of scientific and engineering accelerators.
What to watch
- Official throughput figures and benchmarks from Nvidia as 2028 approaches.
- How research teams integrate AI agents with high-performance computing hardware.
- The development of the Vera Rubin platform and its wider software ecosystem.
Try it
Look up the current number of CUDA cores in your own GPU to see how it compares to the 50,000+ cores in the upcoming Rubin HPC.
Sources and links
- ReportingMezha.net: the article we sourced
Read the original at Mezha.net
First reported by Mezha.net. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.