85 Percent Goodput Is the Number That Decides a Gigawatt Cluster

A hyperscaler just published the number that decides whether a gigawatt cluster is worth what it cost. It is not the GPU count.

• Meta converted a data center that had served web traffic two years ago into training infrastructure, and the first completed phase ran above 85 percent goodput on pre-training and above 90 percent on reinforcement learning.
• A two rack pair draws about 120 kilowatts and forms one 72 GPU NVLink domain, so a single degraded host or switch can stall an entire job.
• Getting from about 47,000 to more than 54,000 GPU hours between interruptions took memory scrubbing, tighter HBM firmware, a better LPDDR error correction algorithm and predictive retirement of devices before a job found them.
• The stated target is two interruptions per 8,000 GPUs per day, and the next platform draws more again with Vera Rubin above 170 kilowatts per rack.
• Almost none of that appears on a quote, which is the part most infrastructure plans still miss.

Read the full analysis at Why 85 percent goodput is the number that decides a gigawatt cluster.