NVIDIA Unveils GB200 NVL4: A Powerhouse AI Accelerator with Quad Blackwell GPUs and Dual Grace CPUs

NVIDIA Introduces GB200 NVL4: A New Era in AI Acceleration

NVIDIA has unveiled its latest high-performance AI accelerator, the GB200 NVL4, marking a significant advancement in the company's Blackwell GPU lineup. This single-server solution integrates four Blackwell GPUs and two Grace CPUs on a single motherboard, pushing the boundaries of AI and high-performance computing (HPC) capabilities 1

1

2

Technical Specifications and Performance

The GB200 NVL4 boasts impressive specifications:

• Four Blackwell GPUs with 768 GB of HBM3E memory
• Two Grace CPUs equipped with 960 GB of LPDDR5X memory
• Combined memory bandwidth of 32 TB/s
• 1.3 TB of coherent memory shared across all four GPUs via NVLink 3

  3

• Total power consumption of 5,400 watts 1

  1

  4

  4

NVIDIA claims significant performance improvements over its predecessor:

• 2.2 times overall performance increase
• 1.8 times faster training speeds compared to the GH200 NVL4 1

  1

  3

  3

• 2.2 times faster for simulation workloads using MILC code
• 80% faster for training the 37-million-parameter GraphCast weather forecasting AI model
• 80% faster for inference on the 7-billion-parameter Llama 2 model 4

  4

NVLink Interconnect Technology

A key component of the NVL4 architecture is the fifth-generation NVLink interconnect technology. This enables efficient communication between all processors on the board, with up to 1.8 TB/sec of bidirectional throughput per GPU. This interconnect is crucial for maintaining optimal performance across multiple processing units, especially during large-scale training operations or when handling multi-trillion parameter models 1

1

2

Cooling and Deployment

Given the GB200 NVL4's high power consumption of 5,400 watts, liquid cooling is necessary to ensure optimal performance. The system is expected to be deployed in server racks for hyperscale customers, who typically utilize custom liquid cooling systems within their data centers 1

1

5

Related Stories

Nvidia Unveils Blackwell Ultra B300: A Leap Forward in AI Computing

NVIDIA Unveils Blackwell Ultra GB300: A Leap Forward in AI Accelerator Technology

HPE Ships First NVIDIA Grace Blackwell GB200 NVL72 System with Advanced Liquid Cooling for AI

NVIDIA Unveils Blackwell AI GPUs: A Leap Forward in AI and Data Center Technology

Industry Collaboration and Availability

NVIDIA is working closely with original equipment manufacturer (OEM) partners to bring various Blackwell-based solutions to market. These include:

• DGX B200
• GB200 Grace Blackwell Superchip
• GB200 Grace Blackwell NVL2
• GB200 Grace Blackwell NVL4
• GB200 Grace Blackwell NVL72 1

  1

Hewlett Packard Enterprise has already announced its Cray Supercomputing EX154n Accelerator Blade, which can pack up to 224 B200 GPUs. This system is expected to become available by the end of 2025 4

4

H200 NVL: A PCIe-based Alternative

Alongside the GB200 NVL4, NVIDIA announced the general availability of its H200 NVL, a PCIe-based Hopper card solution. This offering provides:

• Connectivity for up to 4 GPUs through an NVLink domain
• 7x the bandwidth of a standard PCIe solution
• Flexible server configurations optimized for hybrid HPC and AI workloads
• 1.5x more HBM memory and 1.7x LLM inference performance compared to previous generations 2

  2

  4

  4

The introduction of the GB200 NVL4 and the availability of the H200 NVL demonstrate NVIDIA's commitment to accelerating AI computing and maintaining its leadership in the rapidly evolving AI hardware market.