NVIDIA Rubin vs. Blackwell: Next-Gen AI Architecture, Features, and Performance

NVIDIA Rubin represents a new generation of AI accelerator architecture, succeeding Blackwell. With Rubin, NVIDIA shifts the focus from individual GPUs to a comprehensive computing platform-NVIDIA Vera Rubin-which integrates graphics accelerators, the new Vera server CPU, high-speed memory, and networking components into a unified system for training and deploying large neural networks.

Key Purpose of the Vera Rubin Platform

The main goal of the platform is to handle the rapidly growing computational demands of modern AI systems. Reasoning models generate huge volumes of tokens and work with long context windows, while autonomous AI agents may execute hundreds of sequential operations per task. In these scenarios, a single powerful GPU is no longer enough-memory speed, data exchange between accelerators, and the efficiency of the entire data center become critical.

Rubin is the next evolutionary step from NVIDIA's Grace Hopper and Blackwell approaches-turning a full server rack into a single, tightly integrated computational complex where dozens of GPUs and CPUs function as one cohesive system.

What Is NVIDIA Vera Rubin and Why Did It Replace Blackwell?

Vera Rubin is not just a single processor; it's a family of technologies and systems built around the Rubin architecture. At its core are the Rubin GPU and Vera CPU, but it also includes NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 networking. NVIDIA designs these components together to minimize bottlenecks between compute, memory, and data transfer.

The flagship product is the Vera Rubin NVL72 server system, which combines 72 Rubin GPUs and 36 Vera CPUs in one rack. These accelerators are linked via NVLink 6, allowing software to utilize them as a massive shared resource rather than isolated cards-crucial for today's neural networks, whose sizes can reach hundreds of billions or even trillions of parameters. The speed of interconnects directly impacts overall performance.

While Blackwell advanced the concept with GB200 NVL72 systems, Rubin takes it further-boosting memory bandwidth, updating interconnects, and replacing the Grace CPU with the new Vera processor. The architecture is optimized not just for training, but also for inference with long context, complex reasoning models, and agentic systems that process and generate vast numbers of tokens.

NVIDIA now describes the data center as the new "unit of computation," with performance increasingly defined by how efficiently dozens of accelerators, CPUs, memory, and networking work together-not just by the specs of a single GPU.

NVIDIA Rubin Architecture: GPU, Vera CPU, HBM4, and NVLink 6

Rubin GPU

The new Rubin GPU is engineered specifically for training and running large-scale AI models. It features 336 billion transistors, 224 streaming multiprocessors, and 896 Tensor Cores-a significant increase from Blackwell's 208 billion transistors. Rubin utilizes two large compute dies connected via a high-speed NV-HBI interface, enabling NVIDIA to build an enormous GPU without the manufacturing challenges of a monolithic die.

A key innovation is the third-generation Transformer Engine, supporting low-precision computing (including NVFP4) and delivering up to 50 PFLOPS of NVFP4 performance in inference and up to 35 PFLOPS in training. These figures refer to specialized matrix computations for neural networks, not standard FP32 gaming GPU metrics.

This focus matches the changing nature of AI workloads: modern reasoning models execute lengthy chains of thought, access tools, process extra context, and generate thousands of tokens, requiring sustained high utilization rather than just peak performance.

HBM4 Memory

Memory is another major upgrade. Each Rubin GPU can be equipped with up to 288 GB of HBM4 memory, with bandwidth up to 22 TB/s-a dramatic leap from Blackwell's ~8 TB/s. This bandwidth is vital not only for large models, but also for real-time access to neural network weights and KV-cache during token generation, especially as dialogue length and user concurrency increase.

As a result, HBM4-next-generation memory for AI and HPC-becomes a key component in modern AI accelerators. More capacity keeps more data close to the GPU, while higher bandwidth reduces wait times between compute operations.

In a full Vera Rubin NVL72 rack, total HBM4 reaches about 20.7 TB with aggregate bandwidth around 1.58 PB/s. This underlines why measuring AI systems solely by single-GPU power is increasingly obsolete.

NVIDIA Vera CPU

Alongside Rubin, NVIDIA introduced the Vera CPU to replace Grace. Built on NVIDIA's own Arm-compatible Olympus cores, it is designed for large AI systems, featuring 88 cores and support for up to 1.5 TB LPDDR5X memory. In the Vera Rubin Superchip configuration, each Vera CPU connects to two Rubin GPUs via NVLink-C2C, providing up to 1.8 TB/s of bandwidth.

The Vera CPU's role is not to replace GPUs in neural computation, but to handle data preparation, task management, standard code execution, and other tasks unsuitable for Tensor Cores-freeing GPUs for the most demanding parallel workloads. Fast CPU-GPU links reduce the need for slow data copying via traditional interfaces, crucial for massive models.

NVLink 6 and Multi-GPU Scaling

Data exchange between accelerators is a key bottleneck when models are distributed across dozens of GPUs. Rubin deploys sixth-generation NVLink, delivering up to 3.6 TB/s per GPU and an aggregate of 260 TB/s NVLink bandwidth in a Vera Rubin NVL72 rack. This allows 72 GPUs to communicate over a high-speed fabric-vital for Mixture-of-Experts models and other distributed neural networks.

Rubin's architecture revolves around four tightly coupled resources: Tensor Core compute, HBM4 bandwidth, fast CPU-GPU interconnect, and scalable NVLink. All components must keep pace for the system to remain efficient, which is why Vera Rubin is designed as an integrated platform-not just as a collection of standalone processors.

Vera Rubin Performance: Meeting Modern AI Demands

Vera Rubin's performance is best measured not by a single FLOPS metric, but by infrastructure throughput-how many tokens can be generated per megawatt, how many accelerators are needed to train a model, and how fast the system handles long contexts.

According to NVIDIA, Vera Rubin NVL72 can deliver up to 10 times more tokens per megawatt than GB200 NVL72 for reasoning models like Kimi K2 Thinking, and reduce the cost per million tokens by about an order of magnitude. However, Rubin is not universally ten times faster than Blackwell-gains depend on the workload, data format, context length, and use case.

Agentic systems see particularly large gains. Unlike simple chatbots, AI agents may sequentially search for information, use external tools, launch additional models, and repeatedly analyze results-dramatically increasing token throughput. NVIDIA notes that agentic tasks can require up to 15 times more tokens than traditional AI applications.

Here, the key bottleneck shifts from raw matrix operation speed to the ability to move large data volumes between HBM4, GPU, and accelerators while serving many users. High memory and NVLink 6 bandwidth allow Rubin to sustain such heavy, continuous loads.

In August 2026, NVIDIA published AgentX benchmark results for agentic programming, showing up to 30x higher throughput per megawatt versus GB300 NVL72 at a target of 160 tokens/sec/user, though these were awaiting independent verification.

For large-model training, Rubin also emphasizes system efficiency-NVIDIA claims that training a 10 trillion parameter Mixture-of-Experts model over a set time period requires about four times fewer GPUs with Vera Rubin NVL72 versus GB200 NVL72, reducing hardware, power, and networking needs.

This is where AI Fabric: how networks for training on thousands of GPUs are built becomes critical, since performance depends on how quickly accelerators can exchange results-big GPUs are ineffective if waiting for data from other nodes.

Vera Rubin is therefore designed as the foundation for massive AI factories: NVLink for scale within a rack, InfiniBand and Spectrum-X Ethernet for scaling across clusters. As of May 2026, NVIDIA announced mass production, targeting infrastructures with huge numbers of accelerators.

The approach also benefits Mixture-of-Experts and long-context tasks, where only a subset of the network or a massive context is processed for each token. Fast interconnects and memory minimize communication delays and enable analysis of massive codebases, documents, or extended video sequences.

Ultimately, Vera Rubin's performance is defined by the synergy of a more powerful GPU, expanded HBM4 bandwidth, and faster interconnects, all in a scalable server architecture-meeting workloads where the main resource is billions of generated tokens and massive data flows.

NVIDIA Rubin vs. Blackwell: Key Differences

Rubin continues the trajectory set by Blackwell, but nearly every major component gets a significant upgrade-more GPU compute, adoption of HBM4, doubled NVLink bandwidth, and a new Vera CPU. Thus, generational differences extend beyond individual accelerator specs to the server system as a whole.

Performance metrics are for server-class NVIDIA accelerators, with actual compute performance depending on data format and system configuration.

The most visible change is the GPU itself: Rubin packs 336 billion transistors versus Blackwell's 208 billion, and up to 50 PFLOPS NVFP4 inference compared to 10 PFLOPS. Third-generation Transformer Engine further optimizes for modern transformer and reasoning models.

However, boosting compute alone is not enough-the faster the GPU, the more data must be fed to its compute blocks. Rubin's switch to HBM4 raises memory bandwidth from 8 to 22 TB/s per GPU-especially beneficial for long-context and mass inference workloads, which are often data-transfer-bound, not compute-bound.

Differences are even starker with interconnects: Blackwell's fifth-gen NVLink offers up to 1.8 TB/s bidirectional bandwidth per GPU, while NVLink 6 doubles this to 3.6 TB/s, and total NVLink Switch bandwidth in NVL72 systems jumps from 130 to 260 TB/s.

This is critical for large Mixture-of-Experts models, where different parts of the network operate on different accelerators and must exchange data intensively. Doubling interconnect bandwidth reduces waiting time and better utilizes compute resources.

For a deeper look at the previous generation and the reasons for this transition, see NVIDIA B200 and the Blackwell Architecture: The New AI Accelerator Standard, where NVIDIA first made rack-scale NVL72 systems core to its strategy-Rubin takes this even further.

The central processor has also changed: Blackwell used the Arm-based Grace CPU, while Rubin introduces the 88-core NVIDIA Olympus-based Vera CPU, designed for modern AI factory workloads: environment launch, data handling, KV-cache management, orchestration, and AI agent code execution.

CPU-GPU interconnect bandwidth has doubled as well: NVLink-C2C in Blackwell offered up to 900 GB/s, while Vera and Rubin deliver up to 1.8 TB/s, accelerating data transfer and reducing GPU idle times.

Thus, Rubin should not be seen as just a speed bump over Blackwell. Blackwell already turned a server rack into a tightly integrated compute system; Rubin upgrades every critical element-GPU, memory, CPU, and networking-at once.

Architecturally, the crucial distinction is what workloads the system targets. Blackwell was a universal platform for training and inference of large models, while Rubin is focused on the era of reasoning and agentic AI-handling long operation chains, massive context, and vastly more tokens per user request.

This is why NVIDIA increasingly compares generations by tokens per megawatt, inference cost, and required accelerator count for model training-metrics that directly impact data center operational costs, and where Vera Rubin is designed to provide the biggest advantage over Blackwell.

What Is Rubin CPX and the Future of Vera Rubin?

NVIDIA Rubin CPX

Alongside the main Rubin GPU, NVIDIA has introduced the Rubin CPX accelerator, built on the Rubin architecture but optimized for massive-context inference-ideal for AI-assisted programming, large codebase analysis, generative video, and other tasks involving huge volumes of information.

Rubin CPX is not simply a more powerful Rubin GPU. While the main Rubin is for general-purpose training and inference with HBM4, CPX is designed for massive-context processing, using 128 GB GDDR7 memory for a better cost-performance balance in specialized applications. The chip is a monolithic die, delivering up to 30 PFLOPS NVFP4. NVIDIA claims CPX can triple attention operation performance over GB300 NVL72 in suitable workloads.

For example, AI programming assistants can analyze far larger codebases at once, understanding dependencies and performing complex changes without splitting projects into fragments. Similarly, analyzing an hour of video may require around a million tokens-Rubin CPX combines video encoding/decoding with high-throughput long-context inference, supporting video search, analysis, and extended content generation.

The largest announced configuration, Vera Rubin NVL144 CPX, integrates Vera CPUs, standard Rubin GPUs, and Rubin CPX accelerators, distributing inference stages across specialized processors. NVIDIA claims up to 8 exaflops AI performance per rack, nearly 100 TB of fast memory, and up to 1.7 PB/s aggregate memory bandwidth.

This illustrates NVIDIA's strategy: instead of one GPU optimized for everything, multiple processor types handle distinct workload stages-Rubin for core model computation, Vera for CPU-heavy tasks, and CPX for massive-context processing.

Where Will Vera Rubin Be Used?

The primary market for Vera Rubin is large data centers. The platform targets cloud operators, foundational AI model developers, scientific computing centers, and companies building their own AI factories. This is a different class of hardware from GeForce consumer GPUs-here, thousands of connected accelerators, megawatts of power, and maximum infrastructure utilization matter most.

By 2026, Vera Rubin entered mass production, with NVIDIA collaborating with major server makers and global supply chains, and the architecture designed for further scaling up to infrastructures with huge numbers of GPUs.

One major application is cloud AI. Microsoft plans to use Vera Rubin NVL72 in the next-generation Fairwater infrastructure, while CoreWeave will be among the first cloud providers to offer NVIDIA Rubin to customers. This means developers will have access to the architecture without needing to buy their own server racks-compute resources will be available via the cloud.

Another market is supercomputers. Vera Rubin is designed for both generative AI and traditional high-performance computing (HPC) tasks requiring high numerical precision. NVIDIA supports native FP64 operations and positions the platform for climate modeling, computational fluid dynamics, quantum chemistry, energy research, and more. New supercomputers based on Vera Rubin are planned for Los Alamos National Laboratory, NERSC, and the Leibniz Supercomputing Centre.

The concept of specialized components within the AI factory is also evolving. Alongside Rubin GPU and CPX, NVIDIA is adding dedicated accelerators and systems for inference, context storage, networking, and CPU workloads. In August 2026, NVIDIA announced mass production of Groq 3 LPX-a specialized accelerator for rapid sequential token generation, expanding Vera Rubin infrastructure capabilities.

Vera Rubin is thus becoming a modular platform for diverse AI workloads-some companies need maximum training speed, others prioritize cheap high-volume inference, long context, or minimal response latency. NVIDIA addresses these scenarios with a variety of compute blocks, unified by a common software and networking platform.

Rubin CPX exemplifies this shift: new accelerator generations no longer simply mean a more powerful GPU, but rather increasing specialization, distributing workloads among processors optimized for each computational stage.

Conclusion

NVIDIA Rubin is not just another GPU after Blackwell, but the next step in NVIDIA's evolution toward rack-scale computation. The Vera Rubin platform combines new Rubin GPUs, Vera CPUs, HBM4 memory, NVLink 6, and advanced networking to minimize latency between components and scale modern AI systems more efficiently.

Rubin's main strengths are most apparent in tasks requiring vast computation and constant data exchange: large-model training, reasoning, AI agents, long-context, and mass inference. Here, increased memory and interconnect bandwidth are as crucial as raw Tensor Core compute power.

Compared to Blackwell, the new architecture emphasizes overall infrastructure performance-tokens generated, power efficiency, inference cost, and the effectiveness of large-scale multi-accelerator operation. The introduction of Rubin CPX further highlights NVIDIA's move toward specialized processors for different neural network stages.

For everyday users, Vera Rubin will remain behind the scenes-as core infrastructure for data centers, cloud services, and supercomputers. But on these very platforms, the next generation of AI services will run, making neural networks faster, more complex, and more accessible than ever before.