Test report DSG-7038 · Rev E · tested November 17, 2026

Processors & AcceleratorsDevice under test

NVIDIA Confirms Rubin HPC: 50,000+ CUDA Cores, Landing in 2028

NVIDIA's Rubin HPC GPU arrives in 2028 with 50,000+ CUDA cores and NVLink 7, its first dedicated science part since Hopper, targeting FP64 workloads.

Read
4 min
Words
712
Node
3nm
Operator
Amara Osei

Spec summary

  1. Rubin HPC will ship in 2028 with over 50,000 CUDA cores, roughly twice the current Rubin flagship.
  2. The chip supports NVLink 7, one generation beyond NVLink 6 on the current Rubin server GPU.
  3. NVIDIA's FP64 has plateaued near 40 TFLOPS since Hopper; Blackwell Ultra delivers under 2 TFLOPS.
  4. AMD's HPC-focused Instinct MI430X launches in 2027; Intel is preparing the PCIe-based GPU D-Series.
  5. Rubin HPC supports precisions from NVFP4 up to FP64 and likely uses HBM4e memory.
NVIDIA Announces Rubin HPC: Science-Focused Server GPU to Arrive in 2028
Fig. ANVIDIA Announces Rubin HPC: Science-Focused Server GPU to Arrive in 2028 — Brecht Corbeel / unsplash

NVIDIA will ship an HPC-optimized GPU with more than 50,000 CUDA cores in 2028, the company announced at SC26, marking its first dedicated high-performance computing part since the Hopper generation almost four years ago.

The chip, named Rubin HPC, is an HPC-tuned version of NVIDIA's Rubin architecture. Roughly twice the CUDA core count of the current flagship Rubin SoC signals where NVIDIA is aiming: vector and SIMD processing, not tensor-heavy AI math.

VP of Hyperscale and HPC Ian Buck presented the announcement, which served largely as a teaser. NVIDIA disclosed few details beyond the core count, the 2028 release window, and NVLink 7 support.

Why does a dedicated HPC chip matter?

Server GPU demand has shifted toward AI workloads and low-precision math over the past half-decade. Vendors responded by packing designs with tensor cores optimized for matrix multiplication, leaving traditional science and engineering users with fewer suitable options. For most of the prior decade, server GPUs had treated HPC as a first-class citizen, if not a top customer.

NVIDIA then all but sat out the HPC market with its Blackwell generation. That raised the question of whether the AI leader would bother returning to the smaller HPC segment. The answer, delivered at SC26, is unambiguous: yes, but not until 2028.

The move signals NVIDIA does not intend to permanently cede the HPC and supercomputer markets to AMD and Intel, who have been splitting that business in recent years. In short: NVIDIA is back in the science GPU business.

What do the specifications show?

Confirmed and inferred figures for Rubin HPC against current and prior HPC-class parts:

  • CUDA cores: >50,000, versus 18,944 on B200 and 16,896 on H200
  • Interconnect: NVLink 7, one generation beyond NVLink 6 on the current Rubin server GPU
  • Memory: HBM4e expected, given the 2028 timeline, versus HBM3e on B200
  • FP64 vector performance: NVIDIA quotes only "a whole lot" — the figure to watch
  • Precision support: full range from NVFP4 up to FP64

NVIDIA did not share expected performance figures or memory subsystem details. Instead, the company described a "balanced design" that grows compute and data movement capabilities together, rather than ramping compute throughput alone.

The NVLink 7 detail carries weight. NVIDIA's current Rubin server GPU is the first chip to offer NVLink 6, so Rubin HPC's scale-up networking will be a full generation ahead at launch.

The specification profile points to a connection with Rubin Ultra, NVIDIA's mid-generation refresh due in about a year. Rubin Ultra will be the first chip to ship with NVLink 7, which supports the view that Rubin HPC is closer to a Rubin Ultra derivative than to the original Rubin server chip.

A chip render in NVIDIA's SC26 presentation resembles current Rubin GPU renders — 8 stacks of HBM spread across multiple dies, matching the current Rubin server chip. NVIDIA renders this far from launch tend toward the illustrative rather than the technical.

What about FP64 performance?

NVIDIA's FP64 output has plateaued at just under 40 TFLOPS since Hopper: 37 TFLOPS on B200, 34 TFLOPS on H200. Blackwell Ultra performs especially poorly here at under 2 TFLOPS.

Rubin HPC's much larger CUDA core count gives NVIDIA the means to lift FP64 substantially. How far depends on two unanswered questions: how many FP64 CUDA cores the architecture carries, and how much die space NVIDIA takes from tensor cores to make room. Tensor core counts on the chip remain unconfirmed.

NVIDIA also briefly touted a "sparse data acceleration" capability to keep the GPU cores fed, without explaining how it works. The company's tensor cores have long supported sparse matrices, but this feature appears to be something different, given the emphasis on vector math and CUDA cores.

Who will NVIDIA be fighting?

AMD plans to launch its HPC-focused Instinct MI430X accelerator in 2027, a year before Rubin HPC arrives. Intel is preparing its PCIe-based GPU D-Series. NVIDIA will face entrenched competition on both fronts.

For NVIDIA, the Rubin generation offers a reset after the AI-focused Blackwell era, plus the chance to reuse technologies built for it. Whether the 50,000-core bet retakes the classic supercomputer market will only be answerable in 2028.

via ServeTheHome (Source)

Filed under

  • nvidia
  • rubin-hpc
  • cuda
  • nvlink
  • high-performance-computing
Share this article:

More from Amara Osei

Amara Osei

Show full bio

Staff writer covering money, funding and market shifts around hardware layer of artificial intelligence at Die Signal.

698 articles

Same lot · LOT-362D

« Previous articleNext article »