NVIDIA Unleashes AI Revolution with H200 NVL GPU and GB200 NVL4 Superchip Debut
1min read In a major announcement at the SC24 supercomputing conference, NVIDIA introduced two groundbreaking AI hardware products: the H200 NVL PCIe GPU and the GB200 NVL4 superchip. This launch underscores NVIDIA's commitment to pushing the boundaries of artificial intelligence technology.
The H200 NVL PCIe GPU is specifically designed for environments that offer limited power and cooling capabilities. It caters to enterprise racks that typically provide less than 20 kW of power and rely on air cooling, offering a lower-power AI computing option. Developed in a PCIe AIC form factor, the H200 NVL has a dual-slot thickness and reduces maximum TDP power consumption from 700W to 600W, compared to the H200 SXM model. Although the processing power has slightly decreased, notably by around 15.6% for INT8 Tensor Core operations, it shares the same HBM memory capacity and bandwidth as the H200 SXM, with 141GB and 4.8TB/s respectively.
In terms of connectivity, the H200 NVL supports dual or quad configurations with up to 900GB/s NVLink interconnect bandwidth for each GPU. NVIDIA highlights that the H200 NVL boasts 1.5 times the memory capacity and 1.2 times the bandwidth of its predecessor, the H100 NVL, alongside a 1.7-fold improvement in AI inference performance. Moreover, it delivers a 30% performance boost in high-performance computing (HPC) tasks.
Additionally, NVIDIA revealed the GB200 NVL4 superchip, aimed at single-server solutions. This module combines two Grace CPUs with four Blackwell GPUs, amassing an HBM memory pool of 1.3TB, equating to two sets of GB200 Grace Blackwell superchips with an overall power consumption of 5.4 kW. Compared to the previous-generation GH200 NVL4 system, comprising four Grace CPUs and four Hopper GPUs, the new GB200 NVL4 enhances simulation performance by 2.2 times, AI training performance by 1.8 times, and AI inference performance by 1.8 times.
NVIDIA's GB200 NVL4 superchip is scheduled for market release in the latter half of 2025, poised to make significant strides in AI performance capabilities. This new hardware suite not only reinforces NVIDIA's position at the forefront of AI advancements but also provides substantial improvements in efficiency and performance for a range of applications.
