Tech News: NVIDIA Launches GB200 NVL4 with 2 Grace CPUs and 2 Blackwell GPUs
SMYG LIMITED
Smyg limited is a focus on original spot, ordering electronic components procurement platform.
英伟达 has officially launched two new cutting-edge AI hardware platforms: the Blackwell GB200 NVL4 and the Hopper H200 NVL. These innovative solutions are designed to significantly enhance performance for AI-driven workloads, bringing substantial improvements in training, inference, and simulation tasks.
The Blackwell GB200 NVL4?is an expanded version of NVIDIA’s existing GB200 Grace Blackwell Superchip AI solution. It features a modular design on a larger motherboard, equipped with two Blackwell GB200 GPUs, two Grace CPUs, and four Blackwell B200 GPUs. This configuration forms a robust 4-GPU NVLINK domain with 1.3TB of coherent memory, providing a unified server solution. In terms of performance, the GB200 NVL4 offers a 2.2x improvement in simulation performance and a 1.8x boost in both training and inference performance. This solution is set to be available through NVIDIA's partners in the coming months, bringing powerful AI capabilities to a wide range of industries.
On the other hand, the Hopper H200 NVL, based on PCIe architecture, has officially launched. This platform enables the connection of up to four GPUs via the NVLINK domain, offering seven times the bandwidth of traditional PCIe solutions. The H200 NVL is highly flexible and adaptable, catering to data centers and a variety of mixed HPC (High-Performance Computing) and AI workloads. It offers significant upgrades with 1.5 times more HBM memory, 1.7 times the LLM (Large Language Model) inference performance, and 1.3 times better HPC performance compared to previous models. The Hopper H200 NVL is packed with 114 SMs (Streaming Multiprocessors), a total of 14,592 CUDA cores, 456 Tensor cores, and up to 3 FP8 TFLOPS (FP16 accumulated) of performance. Additionally, the GPU is equipped with 80 GB of HBM2e memory across a 5120-bit interface, with a TDP (Thermal Design Power) of 350 watts.
As for power consumption, the larger GB200 NVL4 solution?is expected to deliver up to 6000W of power, with the Superchip module itself drawing approximately 2700W.
领英推荐
?
Stay tuned for updates: ?? https://www.smbom.com/ ??
If you like this article, please give us?a like! ?? ??