NVIDIA presents its own server, the NVIDIA GB200 NVL72, highlighting its processing power and sufficient memory to train artificial intelligence models up to 1.8T in size. This server is sold as a complete unit, with the GB200 compute tray as the heart, containing two sets of Grace CPUs paired with four Blackwell CPUs, cooled by a water cooling system. Each node boasts 80 petaflops of processing power and 1.7TB of memory. Configuring the unit includes NVLink Switch System to interconnect each node. Each tray holds 18 compute sets, totaling 72 GPUs.
While emphasizing performance in LLM, NVIDIA mentions that the NVIDIA GB200 NVL72 machine accelerates database operations and can simulate physics, such as computational fluid dynamics (CFD), faster than x86 CPUs by up to 22 times.
TLDR: NVIDIA showcases the powerful NVIDIA GB200 NVL72 server, designed for AI model training and accelerated database operations, outperforming x86 CPUs in simulations like CFD.
Leave a Comment