performance

Definition

An overloaded word: in ML research it usually means task quality — accuracy, win rate, benchmark score — while in systems contexts it means throughput and latency. Always check which, and under what conditions, since the two can move in opposite directions. A single headline number also hides variance across seeds and prompts, so treat small gaps between systems with suspicion.

Explainers using this term

• Systolic Arrays — Building the Heart of the TPU From ScratchIn-Datacenter Performance Analysis of a Tensor Processing Unit
• Activation Functions from Scratch — Why Nonlinearity Is Non-NegotiableDelving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification
• How AI Accelerators Are Designed — What Actually Separates GPUs, NPUs, and TPUsIn-Datacenter Performance Analysis of a Tensor Processing Unit