Versus Engine
NVIDIA Vera CPU vs NVIDIA Vera Rubin NVL72
Specs, price and the one trade-off that actually decides it — NVIDIA Vera CPU against NVIDIA Vera Rubin NVL72, side by side.
Cheaper to start
Tie
Both start at a similar price.
Best ecosystem
Tie
Neither lists native integrations.
Standout
NVIDIA Vera CPU
Up to 1.2 TB/s LPDDR5X memory bandwidth
NVIDIA Vera CPU
NVIDIA Vera CPU for AI factories and HPC workloads with high single-threaded performance and massive memory bandwidth.
Where it wins, where it doesn't
Pros
- High single-threaded performance for fast software execution.
- Massive LPDDR5X memory bandwidth (up to 1.2 TB/s) for efficient data-intensive workloads.
- Second-generation NVIDIA SCF for fast, consistent data access across cores.
Cons
- Specialized design may limit applicability for general-purpose computing tasks.
Specifications
- Configuration—
- CPU Core Count—
- CPU Memory—
- FP16/BF16³—
- FP32—
- FP64—
- FP8/FP6 Training³—
- GPU Memory | Bandwidth—
- GPUs in 100 MW¹—
- Host CPUNVIDIA Vera delivers system-level efficiency as the host CPU for AI factories, including NVIDIA Vera Rubin NVL72 and HGX™ Vera Rubin NVL8 platforms. Vera feeds GPUs for large-scale AI while running the CPU work that keeps the factory operating, including ETL, key-value (KV) cache management, and orchestration. With high single-threaded performance, massive memory bandwidth, and a single compute die design that avoids cross-chiplet latency, Vera delivers predictable performance while keeping GPUs fully utilized across accelerated AI and HPC systems.
- Inlet Temp—
- LPDDR5X Memory SubsystemNVIDIA Vera delivers up to 1.2 terabytes per second (TB/s) of LPDDR5X memory bandwidth, providing 2x the bandwidth at half the power of traditional CPU memory. This keeps thousands of parallel software environments responsive while supporting faster RL iterations, efficient KV-cache management, and data-intensive agentic workflows. With up to 1.5 TB of memory, Vera provides the capacity and efficiency for AI factories, analytics, and HPC workloads.
- Networking Bandwidth (Scale-Out)´—
- NVFP4 Inference²—
- NVFP4 Training³—
- NVIDIA NVLink—
- NVIDIA NVLink-C2CNVIDIA NVLink-C2C delivers up to 1.8 TB/s of coherent bandwidth between Vera CPUs and NVIDIA GPUs. When paired with NVIDIA Rubin GPUs, Vera creates a unified memory architecture that helps CPUs and GPUs work together on complex AI and HPC workloads, large datasets, and KV-cache offload. NVLink-C2C reduces data-transfer bottlenecks, simplifies optimization, supports secure isolation for sensitive data and code, and enables high-speed connectivity in dual-socket Vera CPU systems.
- NVLink Bandwidth—
- NVLink-C2C Bandwidth—
- Second-Generation NVIDIA SCFNVIDIA Vera uses second-generation NVIDIA SCF to connect all 88 cores, cache, memory, input and output (IO), and NVLink-C2C across a single compute die. With 3.4 TB/s of bisectional bandwidth and a unified cache architecture, SCF gives cores fast, consistent access to data even when the CPU is fully utilized. By avoiding cross-chiplet communication, Vera maintains predictable latency and throughput for agentic workloads, analytics, and AI factory infrastructure at scale.
- Single-threaded performancehelps software environments, tool calls, and evaluation loops complete faster, while NVIDIA Spatial Multithreading creates 176 threads with partitioned core resources for predictable throughput at scale.
- Standalone CPUFor agentic AI, reinforcement learning, data processing, and analytics, NVIDIA Vera delivers leading per-core performance and massive memory bandwidth to run thousands of parallel sandbox environments, tool calls, code executions, evaluation loops, and data workflows. Faster CPU execution means agents wait less, RL systems generate more feedback per training step, and AI factories produce more tokens per dollar. As a standalone CPU platform, Vera also supports hyperscale cloud, enterprise, and HPC workloads and extends to storage infrastructure with NVIDIA Vera BlueField™-4 STX. Available as a dense, liquid-cooled NVIDIA Vera CPU rack or in standard dual- and single-socket configurations, Vera fits any data center.
- TF32³—
- Total NVIDIA + HBM4 Chips—
NVIDIA Vera Rubin NVL72
High-performance computing system with 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs.
Where it wins, where it doesn't
Pros
- Offers up to 9,360 TFLOPS FP32 and 288 PFLOPS FP16/BF16 for high-performance computing.
- Equipped with up to 54 TB of LPDDR5X CPU memory, ensuring ample resources for complex tasks.
- Supports up to 20.7 TB HBM4 GPU memory with 1,400 TB/s bandwidth, ideal for data-intensive applications.
Cons
- High power consumption and cooling requirements make it unsuitable for environments without robust infrastructure.
- Limited networking bandwidth in smaller configurations restricts scalability and flexibility.
- Complex NVLink requirements necessitate a specific setup, complicating deployment and maintenance.
Specifications
- Configuration72 NVIDIA Rubin GPUs | 36 NVIDIA Vera CPUs | 2 NVIDIA Rubin GPUs | 1 NVIDIA Vera CPU | 1 NVIDIA Rubin GPU
- CPU Core Count3,168 custom NVIDIA Olympus cores | 6,336 Threads | 88 custom NVIDIA Olympus cores | 176 Threads | -
- CPU MemoryUp to 54 TB LPDDR5X | Up to 1.5 TB LPDDR5X | -
- FP16/BF16³288 PFLOPS | 8 PFLOPS | 4 PFLOPS
- FP329,360 TFLOPS | 260 TFLOPS | 130 TFLOPS
- FP642,400 TFLOPS | 67 TFLOPS | 33 TFLOPS
- FP8/FP6 Training³1,260 PFLOPS | 35 PFLOPS | 17.5 PFLOPS
- GPU Memory | Bandwidth20.7 TB HBM4 | 1,400 TB/s | 576 GB HBM4 | 38.5 TB/s | 288 GB HBM4 | 19.2 TB/s
- GPUs in 100 MW¹40K GPUs
- Host CPU—
- Inlet Temp45°C
- LPDDR5X Memory Subsystem—
- Networking Bandwidth (Scale-Out)´32.4 TB/s | 0.9 TB/s | 0.45 TB/s
- NVFP4 Inference²3,600 PFLOPS | 100 PFLOPS | 50 PFLOPS
- NVFP4 Training³2,520 PFLOPS | 70 PFLOPS | 35 PFLOPS
- NVIDIA NVLinkSixth Generation
- NVIDIA NVLink-C2C—
- NVLink Bandwidth216 TB/s | 6 TB/s | 3 TB/s
- NVLink-C2C Bandwidth65 TB/s | 1.8 TB/s | -
- Second-Generation NVIDIA SCF—
- Single-threaded performance—
- Standalone CPU—
- TF32³144 PFLOPS | 4 PFLOPS | 2 PFLOPS
- Total NVIDIA + HBM4 Chips1,296 | 30 | 12