SIGNAL
AI servers stopped being general-purpose computers some time ago, and the CPU stopped being the main character. The chip that lives next to the GPU in an AI rack has one job: keep the accelerator fed with code, context, and tokens, then get out of the way. A useful tell is in the memory. Datacenter CPUs ship with error-correcting server memory in modest, well-cooled amounts. A 'datacenter' CPU that holds 1.5 TB of laptop-class RAM is not pretending to be a Xeon. It is shipping as a host.
The Register's deep dive on Nvidia's Vera makes the design logic legible. The 88 Olympus cores and 176 threads are arranged around an 1.8 TB/s NVLink pipe to the GPU next door, and Nvidia itself positions the chip for 'agentic AI' rather than the usual datacenter basket of workloads. The bandwidth tells you what the memory amount already implied: this is a companion CPU, not a Xeon replacement.
The mechanism survives beyond Nvidia. Any vendor that calls a chip a 'CPU for AI' without pairing it to a GPU at GPU-scale bandwidth is selling marketing. The category will harden over the next two product cycles: companion CPU, or general-purpose CPU with a GPU parked on the side. Watch the interconnect line. If the bandwidth is not built to move a GPU's worth of data, the chip is the second one.
Reported by Sky for Type0, from A deep dive into Nvidia's Vera CPU and the Olympus cores that power it. Read the original: theregister.com