Rubin is NVIDIA’s top-shelf Rubin-generation datacenter GPU to be released in 2H2026.1

  • 2 reticle-limited GPUs
  • 50 PF FP4
  • 288 GB HBM4 at 22 TB/s
  • NVLink 6 at 3.5 TB/s

Performance

Data TypeVFMAMatrixSparseEmulated DGEMM
FP6433 TF200 TFLOPS
FP32130 TF400 TFLOPS
TF322,000 TF
FP16
BF16
FP84,000 TF
FP617,500 TF
NVFP435,000 TF250,000 TF
INT8250 TOPS

Unsorted information

Quote

Other highlighted qualities of the Rubin platform include the speedy NVLink 6 Switch performance of up to 3,600 GB/s, and a touted CX9 SuperNIC component offering up to 1,600 GB/s.

From Nvidia’s Jensen Huang, Ian Buck, and Charlie Boyle on the future of data center rack density:

Quote

In the second half of 2026, Nvidia is promising to deliver the Vera Rubin NVL144, with a new Arm chip and a new GPU. The company has yet to disclose how much power it expects that rack to consume

Quote

From Rubin on, it will instead show the number of reticle-sized GPUs. Both Blackwell and Rubin are each made of two reticle-sized GPUs, which is why the number has doubled to 144, but in reality, the number of GPUs is the same over both generations.

VR200 NVL144

Note

At GTC25, NVIDIA claimed that the NVL nomenclature will change to reflect the number of reticle-limited GPU dies are in an NVL domain rather than the number of GPU packages starting with Rubin-generation GPUs. VR200 NVL144 would still have 72 GPU packages as GB200 NVL72, and each package would have two reticle-limited B200 GPUs.1

However, at CES 2026, NVIDIA continued using the NVL72 nomenclature with Vera Rubin. It is now unclear what the proper branding is.

Each VR200 NVL144 will have:

  • 3.3x performance over GB300 NVL72
    • 3.6 EF FP4 for inference
    • 1.2 EF FP8 for training
  • 13 TB/s HBM4
  • 75 TB “fast memory”
  • 250 TB/s NVLink6
  • 28.8 TB/s ConnectX-9

Fun fact: there are two miles of copper cables (5,000 cables) in a single NVL72 rack.3

It accepts 45C inlet water and share the same Oberon rack as B200.3

DGX SuperPOD with DGX Vera Rubin NVL72

NVIDIA’s SuperPOD for VR200 is a standard block of eight VR200 NVL72 racks with supporting infrastructure.4 Each SuperPOD will have:

  • 576 Rubin GPUs, 28.8 EF FP4 (sparse)
  • 288 Vera CPUs, 88 cores each
  • 600 TB of “fast memory” (LPDDR5X?)
  • 260 TB/s of aggregate NVLink throughput
  • 576 ConnectX-9 NICs?
  • 144 BlueField DPUs

These specifications changed compared to the NVL576 superpod that Jensen presented at GTC25.1

Vera Rubin for HPC

As of June 2025, NVIDIA has already announced two major supercomputers for traditional modeling and simulation that will use the Vera Rubin architecture:

Blue Lion Supercomputer Will Run on NVIDIA Vera Rubin | NVIDIA Blog

Footnotes

  1. Jensen Huang’s Keynote at GTC25 2 3

  2. NVIDIA markets this as “NVFP4 training” and “NVFP4 inferencing.” I assume “NVFP4 inferencing” is sparse.

  3. Jensen’s CES 2026 talk 2

  4. https://blogs.nvidia.com/blog/dgx-superpod-rubin/