NVIDIA Vera Rubin Architecture Explained

By VRLA Tech  ·  AI Hardware Architecture Last updated: June 8, 2026

NVIDIA's Vera Rubin platform is the successor to Blackwell — six co-designed chips functioning as a unified rack-scale AI computing system. This page is the technical breakdown of Rubin: GPU, CPU, memory, interconnect, network, and what each chip actually does. Every specification cited here is sourced from NVIDIA's official announcements at GTC 2025, CES 2026, GTC 2026, and GTC Taipei 2026, or from authoritative coverage by Tom's Hardware, VideoCardz, DCD, and NVIDIA Newsroom.

Origin of the Name

Rubin is named after Vera Rubin (1928-2016), the American astronomer whose observations of galaxy rotation curves provided some of the strongest evidence for dark matter. NVIDIA's architecture naming convention honors scientists who advanced foundational fields: Hopper (Grace Hopper), Blackwell (David Blackwell), Rubin (Vera Rubin), and Feynman (Richard Feynman). The Vera CPU is named after Vera Rubin's first name.

The Full Vera Rubin Platform: Six Co-Designed Chips

Vera Rubin is not a single GPU. It is a complete rack-scale architecture comprising six chips developed in coordination as one system. NVIDIA calls this design philosophy "extreme co-design."

The Rubin GPU

The Rubin GPU is the compute foundation of the platform. Per-package specifications confirmed by NVIDIA at CES 2026 and GTC 2026:

• Transistor count: 336 billion, a 1.6x increase over Blackwell's 208 billion
• Process node: TSMC N3 (3nm), a full-node shrink from Blackwell's 4NP
• Package design: Two reticle-sized compute dies plus I/O dies, assembled in TSMC's CoWoS-L advanced packaging
• Memory: 288GB HBM4 across 8 stacks
• Memory bandwidth: 22 TB/s per package
• NVFP4 inference compute: 50 PFLOPS — 5x Blackwell B200's NVFP4 throughput at platform level
• NVFP4 training compute: 35 PFLOPS — 3.5x Blackwell
• Scale-up interconnect: NVLink 6 at 3.6 TB/s bidirectional per GPU, doubling NVLink 5

The dual-reticle die design is structurally similar to Blackwell B200, which also used two dies in a single package. The change from Blackwell to Rubin is the process node shrink (4NP to N3), the memory generation jump (HBM3e to HBM4), and the NVLink generation jump (5 to 6).

The Vera CPU

Vera is NVIDIA's second-generation custom ARM CPU, the successor to the Grace CPU that paired with Blackwell. Vera is significantly more capable than Grace:

• Cores: 88 custom NVIDIA Olympus ARM cores, supporting Armv9.2
• Threads: 176 via spatial multithreading (2 threads per core)
• Transistor count: 227 billion
• Coherent memory link to Rubin: NVLink-C2C at 1.8 TB/s — 7x faster than PCIe Gen 6
• Speed delta vs Grace: NVIDIA claims Vera is approximately 2x the performance of Grace

NVLink-C2C provides cache-coherent memory access between Vera CPU and Rubin GPU, allowing the GPU to read system memory and the CPU to read GPU memory without explicit transfers. This is critical for agentic AI workloads where the GPU needs frequent access to large external data structures (tool call results, RAG retrieval, KV cache offload).

HBM4 Memory

HBM4 is the sixth generation of High Bandwidth Memory and the most consequential memory upgrade in the Rubin platform. NVIDIA's HBM4 specifications for Rubin:

• Per-package capacity: 288GB
• Stack count: 8 HBM4 stacks per GPU package
• Aggregate bandwidth: 22 TB/s per package — 2.75x Blackwell Ultra's 8 TB/s
• Stack height: 12-Hi and 16-Hi configurations supported
• Suppliers (NVIDIA-certified June 2026): SK Hynix, Samsung Electronics, Micron Technology

HBM4 supply is the gating constraint on Rubin volume. SK Hynix holds the majority of allocation (estimated 60-70% by supply chain analysts). Samsung began HBM4 mass production in February 2026. Micron is the third supplier with smaller volume. NVIDIA CEO Jensen Huang publicly urged increased HBM production in June 2026, indicating that demand exceeds supply at launch.

NVLink 6: The Interconnect Doubling

NVLink is NVIDIA's proprietary GPU-to-GPU interconnect that enables tensor-parallel and pipeline-parallel training and inference at bandwidth far exceeding PCIe.

For tensor-parallel inference of large models (70B+), where every transformer layer performs an all-reduce operation across GPUs, NVLink bandwidth is the critical scaling factor. NVLink 6 on Rubin doubles the inter-GPU bandwidth versus B200 — meaningful for trillion-parameter model serving but largely irrelevant for workloads under 70B parameters that fit on a single GPU. For the full NVLink decision framework, see NVLink vs PCIe for AI.

The Vera Rubin NVL72 Rack-Scale System

The NVL72 is the flagship rack-scale Rubin platform. It's the direct successor to NVIDIA's GB200 NVL72 and GB300 NVL72 (Blackwell Ultra). NVL72 specifications:

• GPUs: 72 Rubin GPU packages
• CPUs: 36 Vera CPUs (1 CPU per 2 GPUs)
• NVLink fabric: 260 TB/s aggregate scale-up bandwidth
• HBM4 capacity: 20.7TB total
• HBM4 aggregate bandwidth: 1.6 PB/s
• System memory: 54TB LPDDR5x
• NVFP4 inference: 3.6 exaflops
• NVFP4 training: 2.5 exaflops
• Cooling: 100% liquid cooled, cable-free modular tray design
• Installation time vs Blackwell: ~5 minutes vs ~2 hours (per NVIDIA), due to modular tray architecture

The NVL72 is built for hyperscaler-scale AI factories. Microsoft will deploy Vera Rubin NVL72 in Fairwater AI superfactory sites. AWS, Google Cloud, Oracle, CoreWeave, Lambda, Nebius, and Nscale are all confirmed first-wave deployment partners.

The Vera Rubin NVL144 CPX Platform

The NVL144 CPX is a specialized Rubin variant announced at the AI Infra Summit in September 2025 and detailed at GTC 2026. It pairs standard Rubin GPUs with a separate class of GPU called Rubin CPX, purpose-built for million-token context inference.

Rubin CPX uses GDDR7 memory (128GB per card) rather than HBM, optimized for the compute-bound prefill phase of long-context inference. Standard Rubin GPUs with HBM4 handle the memory-bandwidth-bound decode phase. The disaggregated architecture is designed for million-token coding assistants, generative video, and other workloads where context length is the constraint.

NVL144 CPX rack specifications:

• Rubin CPX GPUs: 144
• Rubin standard GPUs: 144
• Vera CPUs: 36
• Total compute: 8 exaflops NVFP4 (7.5x GB300 NVL72)
• Memory: 100TB of fast memory per rack
• Memory bandwidth: 1.7 PB/s
• Availability: Expected end of 2026 per NVIDIA

AI partners exploring Rubin CPX include Cursor, Runway, and Magic. Note that the CPX product positioning has been subject to industry discussion about NVIDIA's broader inference strategy, including the Groq partnership announced at GTC 2026. For buyers, the practical takeaway is that CPX is a specialized rack variant; standard Vera Rubin NVL72 remains the mainline Rubin product.

Beyond Rubin: Rubin Ultra, Feynman, Rosa Feynman

NVIDIA's roadmap extends through 2030. Each generation arrives on an annual cadence.

• Rubin Ultra (H2 2027): Four reticle-sized compute dies per package (vs two on standard Rubin). ~100 PFLOPS NVFP4 per GPU package, 1TB HBM4e per GPU, ~32 TB/s memory bandwidth. 3.6 kW per GPU package. Kyber rack architecture (NVL576) with 576 compute dies per rack, 600kW per rack, 800V DC distribution. NVLink 7 at ~10.8 TB/s per GPU. 15 exaflops FP4 inference, 5 exaflops FP8 training per rack.
• Feynman (2028): Successor to Rubin Ultra. Vera CPU continues. Introduces advanced 3D stacking technology, LP40 memory, BlueField-5, NVLink-8.
• Rosa Feynman (2029-2030): Per VideoCardz and OC3D coverage of NVIDIA's RTX Spark roadmap shown at Computex 2026. Rosa is a next-generation CPU paired with Feynman GPUs. Less is publicly known about this generation as of mid-2026.

What This Means for VRLA Tech Customers

VRLA Tech builds custom EPYC GPU servers and AI workstations for on-premise and data center deployment. The Rubin generation affects VRLA Tech's product line in three ways:

1. Rubin SXM availability for VRLA Tech EPYC GPU servers begins H2 2026. VRLA Tech (EPYC GPU servers) will configure Rubin-based 4U and 8U server builds as SXM modules become available to the channel. Allocation queues will affect delivery timelines through 2027.
2. No Rubin-based RTX PRO workstation card is announced. The VRLA Tech Threadripper PRO workstation with RTX PRO 6000 Blackwell remains the current top-end workstation configuration. The earliest realistic Rubin workstation card timeline is late 2027 to 2028 — and this is speculation, not an NVIDIA-announced date.
3. Blackwell remains production hardware through 2027 minimum. Customers buying Blackwell workstations and servers today are not buying obsolete hardware. Blackwell continues running every AI workload through the platform's full operational lifespan.

For the buy-or-wait decision framework, see Blackwell vs Rubin: Should I Wait?. For the full NVIDIA roadmap, see NVIDIA GPU Roadmap 2026-2030.

VRLA Tech updates this page as NVIDIA releases new Rubin technical specifications. Last updated June 8, 2026.

VRLA Tech builds custom AI workstations and GPU servers in Los Angeles for on-premise and data center deployment. Every system is burn-in tested and backed by a 3-year parts warranty plus lifetime US-based engineer support. Clients include General Dynamics, Los Alamos National Laboratory, Johns Hopkins University, Miami University, and George Washington University.