NVIDIA Unveils GB200 NVL4 with Quad Blackwell GPUs & Dual Grace CPUs; H200 NVL Now Generally Available

NVIDIA has announced two powerhouse solutions at SC24 to drive high-performance computing (HPC) and AI workloads: the Blackwell GB200 NVL4 and the now widely available Hopper H200 NVL.

H200 NVL: Flexible AI & HPC Workhorse

• Key Features:

  • PCIe-Based Design: Connects up to four GPUs with an NVLINK domain, delivering 7x faster bandwidth compared to traditional PCIe setups.

  • HPC & AI Capabilities:

    • 1.5x more HBM memory.

    • 1.7x better LLM inference performance.

    • 1.3x uplift in HPC performance.

  • Specs:

    • 114 SMs (14,592 CUDA cores).

    • 456 Tensor Cores.

    • 80 GB of HBM2e memory.

    • TDP: 350W.

• Applications: Fits seamlessly into data centers with hybrid AI/HPC configurations.

GB200 NVL4: Blackwell’s Next Evolution

• Specifications:

  • Features 2 Grace CPUs and 4 Blackwell GB200 GPUs on a single server module.

  • Memory: 1.3TB of coherent memory.

  • Performance:

    • 2.2x better simulation capabilities.

    • 1.8x improvement in Training and Inference workloads.

  • Power: Estimated 6KW TDP, a testament to its massive computational capability.

• Advantages:

  • Doubles the CPU/GPU configuration of the original GB200 solution.

  • Increased memory bandwidth for accelerated AI and HPC applications.

  • Ideal for enterprises needing cutting-edge performance for simulation, AI training, and inference tasks.

NVIDIA’s Future Vision

NVIDIA’s advancements continue to dominate the AI hardware sector:

• Recent MLPerf v4.1 world records in both training and inference highlight the power of Hopper and Blackwell architectures.

• Accelerating its AI roadmap to an annual cadence, with upcoming infrastructures like Blackwell Ultra and Rubin.