Glossary › in-datacenter
GLOSSARY
in-datacenter
appears in 2 paper titles
Definition
Signals that measurements were taken inside a production datacenter under real workloads, not on a benchmark rig. Google's TPU paper used it in exactly this sense, reporting utilisation and the actual mix of serving jobs alongside raw throughput. The point of the qualifier is the gap it exposes: peak specification numbers routinely overstate what a deployed accelerator delivers once traffic patterns, memory bandwidth, and idle time are counted.
Explainers using this term
- Systolic Arrays — Building the Heart of the TPU From ScratchIn-Datacenter Performance Analysis of a Tensor Processing Unit
- How AI Accelerators Are Designed — What Actually Separates GPUs, NPUs, and TPUsIn-Datacenter Performance Analysis of a Tensor Processing Unit