SIAgentNews.com
← All news
DevelopingChips

NVIDIA Rubin HPC GPU to feature over 50,000 CUDA cores by 2028

NVIDIA is expanding back into traditional high-performance computing with a new GPU architecture designed for complex scientific simulations.

· First reported by Wccftech

Blueprint drawing: picture for the story 'NVIDIA Rubin HPC GPU to feature over 50,000 CUDA cores by 2028'
AI-made blueprint drawing for this story. No real people or products are shown.
50,000CUDA cores per Rubin HPC GPU

The short version

  • The Rubin HPC GPU features over 50,000 CUDA cores on a single chip.
  • It utilizes 7th Gen NVLink technology for high-speed GPU-to-GPU connectivity.
  • The hardware is expected to be available for science and engineering workloads in 2028.
  • Server trays will eventually pack over 200,000 cores using multiple Rubin chips.

Why it matters

This signals NVIDIA's return to the traditional high-performance computing (HPC) market after focusing heavily on AI. It provides researchers with the specialized hardware needed to run massive simulations faster.

A return to traditional HPC

NVIDIA recently announced the Rubin HPC GPU, a new hardware solution designed specifically for science and engineering workloads. This move marks a notable shift for the company, which has spent a significant amount of time focusing primarily on the artificial intelligence market. By introducing Rubin HPC, NVIDIA is expanding its portfolio back into traditional data center use cases.

The new chip is built on the Rubin graphics architecture and is intended to accelerate various data processing and simulation workflows. While the company has not provided an exhaustive list of specifications, the primary focus is on providing massive compute power. This hardware aims to serve researchers who need to run complex calculations that go beyond standard AI training.

This expansion comes at a time when competitors like AMD are offering specialized variants for high-performance computing. By re-entering this space, NVIDIA is positioning itself to capture the needs of scientific discovery. The company plans to make these GPUs available for these specific workloads starting in 2028.

Core counts and architecture

The Rubin HPC chip is designed to pack over 50,000 CUDA cores on a single GPU. These cores are dedicated to accelerating simulations and analytics workflows that require high-intensity processing. To ensure these cores remain productive, NVIDIA is incorporating sparse data acceleration capabilities. This technology helps keep the cores fed with data so they spend more time computing rather than waiting for information to arrive.

The hardware architecture features a large GPU die positioned in the middle of the chip, surrounded by eight HBM sites. HBM, or High Bandwidth Memory, is a type of memory designed to provide high-speed data access to the processors. The physical layout suggests that two Rubin GPUs can be joined into a single unit using high-speed interconnects.

NVIDIA is also utilizing 7th Gen NVLink interconnect technology to facilitate GPU-to-GPU connectivity. This balance between core count and data supply is what optimizes the chip for scientific computing. It allows for the efficient handling of recurring patterns found in every simulation workload, such as sorting and reductions.

Accelerating scientific discovery

The primary goal of the Rubin HPC is to speed up the simulations used by scientists and engineers to test new ideas. Faster simulations allow for quicker discovery and more rapid exploration of possibilities. By reducing the time it takes to run a simulation, researchers can move through the trial-and-error process of engineering much faster.

The hardware is specifically tuned to handle sparse matrix operations and stencil computations. These are common mathematical patterns found in many scientific models, including those used in physics and chemistry. By accelerating these specific patterns, the Rubin HPC helps researchers bring breakthroughs closer to reality. It provides the raw power needed to handle the most demanding data processing tasks.

Ian Buck, a vice president at NVIDIA, noted that "For agents tackling science and engineering, faster simulation means faster discovery." This quote highlights the company's intent to provide a tool that guides the next steps of scientific exploration. The hardware is meant to be a foundational piece of infrastructure for the next generation of engineering breakthroughs.

Scaling to server trays

While the individual GPU features over 50,000 cores, NVIDIA plans to scale this technology significantly within server environments. Each Superchip will include two Rubin HPC chips and a Vera CPU. This configuration will pack over 100,000 cores into a single Superchip unit for high-density computing.

For even larger requirements, NVIDIA will offer trays that include the latest NVLink switches. These trays are designed to pack over 200,000 cores, providing a massive amount of compute power in a single rack. This scaling allows for the processing of even larger datasets and more complex simulations than a single GPU could handle alone.

The architecture of these servers is designed to maximize the efficiency of the interconnects. By using the 7th Gen NVLink, the system can move data between the many cores across different chips with minimal latency. This ensures that the massive core counts translate into actual performance gains for the end user.

Timeline and market outlook

NVIDIA has set an expected availability date of 2028 for the Rubin HPC GPU. This means there is more than a year of waiting before the hardware reaches the market. The timeline reflects the complexity of developing high-performance hardware that must meet the rigorous demands of scientific research.

The announcement comes after NVIDIA previously announced the Rubin CPX GPU, which was later cancelled. Given this history, industry observers are watching the Rubin HPC closely to see if it successfully makes it to market. The company must navigate the transition from its AI-centric focus back to the traditional HPC stage successfully.

The success of Rubin HPC will depend on its ability to outperform competitors in FP64 compute, which is a standard for high-precision scientific calculations. If NVIDIA can provide a significant leap in these capabilities, it will secure its position in the scientific community. For now, the 2028 date serves as the primary target for this new hardware generation.

What to watch

  • Whether the Rubin HPC successfully reaches the market following the cancellation of the Rubin CPX.
  • How the Rubin HPC compares to AMD's Instinct MI400 series in FP64 compute performance.
  • The specific performance benchmarks for sparse matrix operations and stencil computations once the hardware is released.

Try it

Check the NVIDIA website for updates on the Rubin architecture and the Vera CPU to see how they plan to integrate with the 2028 hardware launch.

Sources and links

First reported by Wccftech. 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.

More AI news