SIAgentNews.com
← All news
DevelopingChips

NVIDIA announces Rubin HPC GPU with 50,000 cores for 2028

The new chip targets scientific computing and data analytics, featuring a massive core count and new data acceleration for complex simulations.

· First reported by VideoCardz.com

Blueprint drawing: picture for the story 'NVIDIA announces Rubin HPC GPU with 50,000 cores for 2028'
AI-made blueprint drawing for this story. No real people or products are shown.
50,000CUDA cores on a single GPU

The short version

  • Rubin HPC features more than 50,000 CUDA cores, the highest count NVIDIA has ever announced for a single GPU.
  • The chip supports seventh-generation NVLink technology for high-speed data movement between processors.
  • New sparse data acceleration helps the GPU handle scattered data in climate modeling and genomic analysis.
  • The hardware is scheduled for release in 2028, two years after the standard Rubin generation arrives.

Why it matters

This hardware signals a shift toward optimizing GPUs for complex scientific simulations, not just AI training. It will likely accelerate how researchers model climate change and analyze human genomes.

A new focus on science

NVIDIA recently unveiled the Rubin HPC GPU, a specialized processor designed for high-performance computing. While many people associate the company with AI training, this specific chip targets scientific simulations and data analytics. It is part of the broader Vera Rubin platform but is built for general-purpose GPU computation.

The company revealed the hardware at SC26, positioning it as a tool for heavy-duty engineering and research. It is intended to handle workloads that require massive amounts of processing power rather than just running AI inference. This distinction helps clarify that the Rubin HPC is meant for diverse, complex math problems.

By focusing on these specific tasks, NVIDIA aims to support industries that rely on massive data sets. These include fields like financial risk analysis and engineering simulations. The chip is designed to be a workhorse for researchers who need to process information at a massive scale.

Massive core counts and links

One of the most notable specifications for the Rubin HPC is its core count. NVIDIA confirmed that the processor features more than 50,000 CUDA cores on a single GPU. This represents the highest number of cores the company has ever announced for one individual chip.

To keep these cores working efficiently, the chip will support seventh-generation NVLink technology. NVLink is an interconnect technology that allows different GPUs to communicate and share data with each other. This is an upgrade over the NVLink 6 system used in the standard Rubin architecture.

The goal of these hardware choices is to maximize the speed of data movement. Because high-performance computing requires moving massive amounts of information quickly, the interconnect is just as important as the cores themselves. These features ensure the chip does not become a bottleneck during large-scale operations.

Solving the data bottleneck

A key technical addition to the Rubin HPC is sparse data acceleration. This feature is designed to improve how the GPU accesses scattered data during a computation. In many complex simulations, data is not organized in a neat, continuous line, which can slow down processing speeds.

NVIDIA explains that this capability keeps the CUDA cores supplied with information constantly. Without it, the cores might sit idle while waiting for memory operations to complete. By streamlining this process, the hardware can handle operations like sorting, reductions, and stencil computations more effectively.

These specific types of operations are common in professional engineering software. By optimizing for these patterns, NVIDIA is making the hardware more efficient for specialized scientific needs. This helps researchers run simulations that were previously too slow or computationally expensive to complete in a reasonable timeframe.

Real-world applications

The architecture is designed to benefit a wide range of scientific and industrial fields. NVIDIA specifically highlighted climate modeling and genomic analysis as primary use cases. These fields require the ability to process vast amounts of variables simultaneously to produce accurate results.

Engineering simulations and financial risk analysis are also listed as key workloads. These tasks often involve sparse matrix calculations, which the new hardware is specifically tuned to handle. This means the chip can be used to test new materials, predict weather patterns, or model complex economic shifts.

Additionally, the Rubin HPC will work with NVIDIA’s CUDA-X libraries. These libraries include algorithms that AI agents can use when they are running simulations. This creates a bridge between traditional scientific computing and the growing field of autonomous AI agents.

Timeline and missing specs

NVIDIA expects the Rubin HPC to become available in 2028. This launch date is set for two years after the introduction of the standard Rubin generation. This timeline suggests a deliberate rollout to ensure the hardware is ready for complex enterprise and research environments.

While the core count and interconnect generation are known, many other specifications remain secret. NVIDIA has not yet provided the exact memory capacity, power consumption, or specific compute performance numbers. These details are often kept private until closer to the actual release date.

For comparison, the existing Rubin GPU features 288GB of HBM4 memory and up to 22 TB/s of memory bandwidth. However, NVIDIA warned that these figures should not be assumed for the new Rubin HPC. The company has also not announced pricing, manufacturing partners, or specific system configurations for the new chip.

What to watch

  • Will the 2028 launch date be pushed back due to manufacturing delays?
  • How will the sparse data acceleration compare to current software-based solutions?
  • Which manufacturing partners will be chosen to produce the Rubin HPC?

Try it

Look up the difference between 'sparse' and 'dense' data to understand why specialized acceleration is needed for scientific simulations.

Sources and links

First reported by VideoCardz.com. 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