"At the AI Infra Summit this week, simulation savant Cadence added Nvidia's biggest iron, the GB200 NVL72 Superpod, to its digital twin fire. At roughly a megawatt, the compute cluster is among the GPU giant's most complex compute platforms to date. Each Superpod consists of eight 120 kilowatt NVL72 racks containing more than 500 Blackwell GPUs and 288 Grace CPUs capable of churning out a combined 11.5 exaFLOPS of the lowest-precision compute money can buy."
"However, getting the most out of the systems requires a datacenter specifically designed not only to handle the intense thermal load that comes with packing a megawatt of compute into eight racks, but one whose power draw can jump from idle to 100 percent and back in a fraction of a second. The last thing bit barn operators want to do is find out the hard way that they've under-specced their facilities and can't fully utilize the machines their customers have borrowed billions to buy."
Cadence integrated Nvidia's GB200 NVL72 Superpod into its Reality Digital Twin Platform to model operational, thermal, and power dynamics of high-density GPU clusters. The Superpod includes eight 120 kW NVL72 racks with over 500 Blackwell GPUs and 288 Grace CPUs, delivering about 11.5 exaFLOPS of low-precision compute. Datacenters must handle intense thermal loads and rapid power swings from idle to full load. Running simulations on spare GPUs can reveal whether power delivery, cooling, and capacity planning will support full utilization before purchasing large GPU systems. Nvidia's Omniverse is being used to gamify datacenter design, and collaboration with Cadence began in March.
Read at Theregister
Unable to calculate read time
Collection
[
|
...
]