UNCOVERING THE LAYERS
Marc Cram , Director of New Market Development at Legrand ,
It ’ s a safe assumption that each software workload has its unique power consumption requirements . If form follows function , then the application drives architectural choices for hardware and its environment . Hyperscalers provide a roadmap for adding more space and more racks for more servers when we think we ’ ve reached , or are about to hit , our power consumption caps . But supercomputing wants everything physically close together to maximise throughput , while AI wants to be on specialised processors , and by its very nature , Edge Computing is inherently distributed .
MareNostrum is the central supercomputer at Barcelona Supercomputing Centre . Its generalpurpose block has 48 racks with 3,456 nodes . Each node has two Intel Xeon Platinum chips , each with 24 processors , amounting to 165,888 processors and a main memory of 390 terabytes . All this is sitting in the Chapel Torre Girona , built in the 1920s . As one can imagine , placing the data centre without disturbing the chapel ’ s structure is an ultimate challenge of making the most with existing conditions .
The space dedicated to processing and CRACs ( computer room air conditioners ) leaves little room for distributing power going into the units . A situation like this has potential challenges for the deployment of PDUs , necessitating a customised solution :
• There is possibly little or no room at the back of the rack for a zero-u PDU , indicating it might have to sit in racks ’ sides
• The likelihood of little or no airflow to cool the PDU suggests convection outside the rack cools the PDU
• Taller racks with more servers generate high outlet density situations
• The need for high power density for the racks may necessitate PDUs with monitoring capabilities
AI poses possible predicaments for PDUs
Artificial Intelligence regularly produces incredible accomplishments with www . intelligentdatacentres . com
63