NVIDIA GB200 NVL4

Type: Platform Tags: NVIDIA, GPU, Hardware, NVLink, Blackwell, Grace, Data Center, Single-Node, AI Related: NVIDIA-GB200-NVL72, NVIDIA-Blackwell-Architecture, NVLink, NVIDIA-Grace-CPU, NVIDIA-HGX, NVIDIA-DGX, NCCL Sources: ServeTheHome, Tweaktown, Tom’s Hardware, VideoCardz — November 2024 coverage of NVIDIA announcement Last Updated: 2026-04-10

Summary

The NVIDIA GB200 NVL4 is a single-server Grace Blackwell configuration combining four Blackwell B200 GPUs and two Grace CPUs on an enlarged motherboard, connected via fifth-generation NVLink. It targets data centers that need Blackwell-generation AI performance without the full rack-scale footprint of the NVL72. It delivers up to 1.3 TB of coherent shared memory and ~6 kW system power — positioned as the entry point into the GB200 ecosystem.

Detail

Purpose

Bridges the gap between single-GPU or dual-GPU Blackwell deployments and the full rack-scale NVL72. Enables organizations to deploy Blackwell AI compute in a standard server form factor without liquid-cooling infrastructure or full-rack commitment.

Key Specifications

  • GPUs: 4x NVIDIA Blackwell B200 GPUs
  • CPUs: 2x NVIDIA Grace CPUs (Arm Neoverse V2, 72 cores each — 144 total)
  • Coherent memory: Up to 1.3 TB (HBM3E GPU memory + Grace LPDDR5X unified via NVLink-C2C)
  • NVLink bandwidth: 1.8 TB/s bidirectional per GPU (5th-gen NVLink)
  • System power: ~6 kW (full server with NICs, SSDs, and components)
  • Form factor: Single server, enlarged motherboard (4-way NVLink domain)
  • Availability: H2 2025, via OEM partners (MSI, ASUS, Gigabyte, Lenovo, HPE, Wistron, Pegatron, ASRock Rack)

Performance vs. Prior Generation (GH200 NVL4)

  • 2.2x simulation performance
  • 1.8x training performance
  • 1.8x inference performance

Key Features

  • 4-way NVLink domain — all four B200 GPUs fully connected via 5th-gen NVLink
  • NVLink-C2C connects Grace CPUs to B200 GPUs at 900 GB/s per link
  • Unified coherent 1.3 TB memory pool across CPUs and GPUs
  • Single-node deployment — no NVLink Switch required (unlike NVL72)
  • Standard server chassis compatible — no dedicated liquid-cooling rack needed

How It Differs from NVL72

NVL4NVL72
GPUs4x B20072x B200
CPUs2x Grace36x Grace
Memory~1.3 TB13.4 TB
NVLink4-way domain72-way all-to-all
Form factorSingle serverFull rack (liquid-cooled)
NVLink SwitchNot requiredRequired
Power~6 kWRack-scale

Use Cases

  • Mid-scale LLM inference (models that fit within 1.3 TB)
  • Enterprise AI deployment without full rack infrastructure
  • HPC simulation workloads at departmental scale
  • Development and fine-tuning of large models
  • Organizations stepping into Blackwell without full NVL72 commitment

Target Customers

Enterprises, research institutions, and cloud providers wanting Blackwell-generation AI in a standard data center footprint. OEM ecosystem: Lenovo, HPE, ASUS, MSI, Gigabyte, ASRock Rack, Wistron, Pegatron.

Connections

Resources