Skip to content
Back to posts

Vera Rubin - Super Simple Explanation and Comparison with Blackwell & AMD

Notes by SDG ·

Instead of being just a chip, Vera Rubin is a whole system designed from the ground up as one AI supercomputer — a rack in a data center becomes the “unit of compute” rather than an individual processor. This means all the pieces are co-designed to work together seamlessly. 

You can think of it like a mega orchestra where every section is custom tuned and synchronized rather than a bunch of solo musicians playing separately.

🧩 Six Major Components — The “Symphony Players”

The Vera Rubin architecture is built from six tightly coupled chips and subsystems that each play a specific role inside a rack-scale AI machine. 

🔹 1. Vera CPU

Role: Orchestrator and data mover

  • Custom ARM-based CPU with 88 Olympus cores and “spatial multi-threading” for high throughput. 
  • Designed not just for general compute, but for AI reasoning workloads and data orchestration across the system.
  • Acts as coordinator for tasks, memory, networking, and security.
  • Can access large shared memory (up to ~1.5 TB per CPU) with high bandwidth. 

👉 This is more than a typical server CPU — it’s aimed at keeping data flowing to GPUs and coordinating reasoning tasks efficiently.

🔹 2. Rubin GPU (Graphics Processor)

Role: The AI engine

  • Built on the next-gen Rubin microarchitecture with HBM4 memory (high-bandwidth memory). 
  • Delivers tens of petaflops of compute using FP4 data precision, which is efficient for large AI models. 
  • Includes a third-generation Transformer Engine for faster AI math and adaptive compression to feed data faster. 
  • In one NVL72 system, 72 of these GPUs work together under a single unified memory and compute fabric. 

👉 GPUs do the heavy lifting for both training and inference (the phase where models answer queries), and Rubin GPUs boost throughput and efficiency massively.

🔹 3. NVLink 6 Switch

Role: High-speed internal data highway

  • A specialized interconnect switch that connects CPUs and GPUs inside the rack. 
  • Provides extremely high bandwidth — ~3.6 terabytes per second per GPU and ~260 TB/s total within the rack. 
  • Enables GPUs to share memory and data directly at speeds rivaling internal GPU memory, which is critical for scaling large models efficiently.

👉 Think of this as the super-fast wired backbone that keeps all parts talking without delays.

🔹 4. ConnectX-9 SuperNIC

Role: Network interface for scaling out

  • A network interface card that connects the rack to external networks and other racks. 
  • Offers huge networking bandwidth (~1.6 Tb/s per GPU), so clusters can scale across data centers without choking on communication. 
  • Helps distribute tasks across multiple racks. This is key for very large AI workloads.

👉 It’s the bridge between the internal supercomputer and the outside cluster ecosystem.

🔹 5. BlueField-4 DPU (Data Processing Unit)

Role: Traffic cop, security guard, and data mover

  • Offloads storage, security, and networking tasks from CPUs/GPUs. 
  • Includes its own processing cores and network engines to handle things like secure isolation (confidential computing) and task dispatching. 
  • Enables inference context memory — very fast shared memory storage for AI models. 

👉 DPUs take care of logistics so the heavy compute units can stay focused on AI math.

🔹 6. Spectrum-6 Ethernet Switch

Role: Cluster-wide traffic director

  • A large-scale Ethernet switch for cluster networking. 
  • Manages traffic between racks and data center infrastructure.

👉 This helps when Vera Rubin systems are linked to build super-clusters or “AI factories.”

📦 How These Parts Fit Physically

Inside a rack (e.g., the NVL72 configuration):

  • 🔌 72 Rubin GPUs
  • ⚙️ 36 Vera CPUs
  • 🛜 18 BlueField-4 DPUs
  • 🔗 NVLink 6 fabric connecting everything internally
  • 🌐 ConnectX-9 and Spectrum-6 handling scaling and external links

All of these elements are laid out in a modular, cable-free tray design so the system can be assembled and serviced quickly. 

📊 Key Technical Benefits

By tying all these parts together in one fabric:

⚡ 1. Unified Memory & Compute

Data doesn’t have to shuttle between separate boxes — GPUs and CPUs share a coherent memory space, making large AI models run more like they’re on one giant processor. 

🚀 2. Massive Bandwidth

With NVLink 6 and ConnectX-9 networking, data moves at blazing rates both inside the rack and out to other racks. 

🔐 3. Security & Isolation

Confidential computing spans CPU, GPU, and interconnect domains — meaning models and data stay protected even when shared across tenants. 

💰 4. Efficiency

Compared to the prior Blackwell architecture, Vera Rubin promises up to 10x lower inference cost per token and 4x fewer GPUs needed for certain training tasks thanks to the integrated design and advanced Transformer engines. 

🧠 In Plain Words

Instead of selling a single engine, NVIDIA built an entire AI car where every component — engine, transmission, fuel system, telemetry, and driver — is tuned to work together smoothly. CPUs handle the planning and coordination, GPUs do the heavy math, DPUs guard and move data behind the scenes, and fast interconnects make sure nothing is ever waiting in line. That’s why Vera Rubin is a platform, not just a processor. 🚗🤖


Rubin vs Blackwell vs AMD


  • Rubin = NVIDIA’s new platform (Vera Rubin CPU + Rubin GPU, etc.)
  • Blackwell = the current big NVIDIA platform (Grace-Blackwell GB200 NVL72)
  • AMD = Instinct MI300 / MI325X family

1. Big Picture: What each one is trying to be

🟣 NVIDIA Vera Rubin (next-gen, 2026+)

  • Think of Rubin as NVIDIA’s “AI factory 2.0”.
  • It doesn’t just upgrade the GPU; it re-designs the full rack: CPU, GPU, networking, storage, and software all co-designed. 
  • Goal: Huge models, lower cost per answer, better for agent-style AI and long conversations.

🟢 NVIDIA Blackwell (current flagship)

  • Blackwell is today’s workhorse AI engine, already powering giant clusters at Microsoft, OpenAI, etc. 
  • Also a rack-scale system (GB200 NVL72), but:
  • Less integrated “six-chip” story,
  • Less focus on extreme cost-per-token reduction than Rubin.

🔴 AMD Instinct (MI300 / MI325X)

  • AMD’s answer to NVIDIA in data center GPUs. 
  • Focus:
  • Lots of memory per GPU (good for big models),
  • Open-source stack (ROCm),
  • Better price/performance in some scenarios.
  • More “GPU + software” than “full rack platform” (though AMD does partner with system vendors).

2. Performance & Cost: Who’s faster, who’s cheaper?

Rubin vs Blackwell (NVIDIA vs NVIDIA)

From NVIDIA & early coverage:

  • Raw speed per GPU (inference)
  • Rubin GPU: about 5× Blackwell’s FP4 inference performance. 
  • Training speed
  • Rubin: roughly 3.5× faster training than Blackwell for certain workloads. 
  • Number of GPUs needed (for MoE models)
  • Rubin can train mixture-of-experts models with ¼ the number of GPUs compared to Blackwell. 
  • Cost per token (inference)
  • Rubin target: up to 10× lower cost per token than Blackwell. 

What this means in normal language:

If Blackwell is already a Ferrari for AI, Rubin is a whole faster racing team that also uses 10× less fuel per lap for many workloads.

Blackwell is still very powerful (and shipping today), but Rubin is NVIDIA saying:

“Okay, now that AI is huge, we optimize not just speed, but the power bill and hardware footprint too.”

NVIDIA (Rubin/Blackwell) vs AMD (MI300 / MI325X)

AMD strengths:

  • Big memory per GPU
  • MI325X: up to 256 GB HBM3E per GPU, ~6 TB/s bandwidth. Good for large models without splitting them across many GPUs.
  • Open software ecosystem (ROCm)
  • Attractive to clouds & companies that don’t want to be fully locked into NVIDIA and CUDA.
  • Price / performance
  • Independent analyses show AMD can be competitive or better per dollar in some inference scenarios, but:
  • Software ecosystem and tooling still lag NVIDIA in many areas.
  • Adoption is catching up but still behind in terms of AI framework maturity.

NVIDIA strengths (Blackwell & Rubin):

  • Ecosystem & maturity
  • CUDA, TensorRT, cuDNN, Triton, plus huge amounts of code already written & optimized.
  • System-level design
  • GB200 NVL72 (Blackwell) and Rubin NVL72 are both rack-scale “one big GPU” machines, with 72 GPUs linked via NVLink in a single domain.
  • Extreme codesign in Rubin
  • Six chips (Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6) are designed as one AI factory platform, not just parts thrown together.

So, in simple terms:

  • AMD: “Here’s a very strong GPU with lots of memory and open software, great value if you’re willing to optimize and invest in the stack.”
  • NVIDIA Blackwell: “Here’s the current standard AI engine, fully integrated, massive ecosystem.”
  • NVIDIA Rubin: “Here’s the next generation, where we optimize entire data centers for AI cost and throughput, not just single chips.”

3. Architecture & Design Philosophy (super simple framing)

🟣 Rubin: “AI Factory in a Rack”

  • Goal: Treat each rack like one giant AI brain.
  • Key ideas:
  • 72 Rubin GPUs + 36 Vera CPUs per NVL72 rack. 
  • 600 TB-class memory pool per rack, ultra-fast NVLink 6 fabric, advanced Ethernet/InfiniBand.
  • Tight integration with BlueField-4 & Spectrum-X for storage, security, and networking. 

🟢 Blackwell: “Monster GPU Cluster”

  • Goal: Make huge models run in real time (trillion-parameter LLMs, etc.).
  • GB200 NVL72:
  • 72 Blackwell GPUs + 36 Grace CPUs. 
  • 5th-gen NVLink, massive FP4 performance, liquid cooling.
  • Focus on speed & density; cost per token is good, but Rubin aims to crush it.

🔴 AMD Instinct: “Big GPU with Big Memory”

  • Goal: Provide a powerful GPU with tons of memory, using an open ecosystem.
  • MI325X:
  • 256 GB HBM3E, 6 TB/s bandwidth. 
  • Great for big models on a single GPU or small cluster.
  • System-level story (whole “factory”) is less vertically integrated than NVIDIA’s Rubin narrative.

4. If you’re “not very technical”, how should you think about it?

Imagine you’re choosing between AI trucking fleets:

  • AMD MI300 / MI325X
“Here’s a really big, efficient truck with a huge cargo space (memory). You can buy lots of them and build your own logistics network.”
  • NVIDIA Blackwell
“Here’s a very fast, premium truck fleet with its own optimized highways (NVLink), fueling stations (software), and driver training programs (tooling).”
  • NVIDIA Rubin
“Here’s an entire automated logistics city designed end-to-end for moving goods (tokens) as cheaply and quickly as possible, with robots, smart roads, traffic control, and energy management all built together.”

Rubin is less about “one chip is faster than another” and more about:

“What if we co-designed all the pieces so AI is not just fast, but economical at crazy scale?”