One replica in a serving fleet returns tokens slower than its siblings. Find out why.

Identical replicas serving identical traffic should produce identical numbers, so a difference is a defect and the fleet median is the tool that finds it. Five causes, the three that are about the replica and the two that are about what it was sent.

Updated Sep 2026 · Grounded in real AI infrastructure interview loops and written to a senior-engineer editorial bar, with every number worked and every diagram hand-built.

20 answers per topic instead of 10, plus saved progress and bookmarks · no cardor unlock all 283 remaining answers · ₹2,000 / $25

The concepts behind this question

Ranked by how closely each one overlaps this question's topic, so the first card is the thing to read if the answer above moved too fast.

Foundational

💻 Coding for Infra

Cache-Friendly Data StructuresA cache line is 64 bytes and it is the unit of coherence, so where data sits decides how fast code runs more often than which algorithm it uses. Two consequences dominate infrastructure code: a lookup that chases a pointer pays two dependent memory stalls instead of one, and two threads updating adjacent variables contend for a line they do not logically share. Both are layout problems with layout fixes.

Foundational

🩺 Fleet Reliability & Observability

Alert Design and On-Call LoadAn alert exists to change what a human does, so any alert that fires without a decision attached is a false alarm regardless of whether its condition was true. GPU fleets generate a specific set of noisy signals that look serious and are not, and separating those from the ones that need a person at three in the morning is what keeps a rotation sustainable. The measure of an alerting system is the fraction of pages that led to an action.

Advanced

🩺 Fleet Reliability & Observability🔒 Premium

Thermal, Power and Cooling EventsA GPU that gets too hot or is denied power does not fail; it slows down, and on a synchronous job a slow GPU is a slow job. Thermal and power events are the most common cause of the 'nothing failed but the run is 15% slower' ticket, and they are the incidents that scale from one node to a whole hall when a cooling distribution unit or a power feed has a problem. This page explains how throttling works, derives the step-time cost of a clock reduction, walks the failure modes of air and liquid cooling, and covers the power behaviour peculiar to training: thousands of GPUs going idle and busy in lockstep.

Core

🩺 Fleet Reliability & ObservabilitySign in

DCGM and GPU TelemetryNVIDIA's Data Center GPU Manager reads a GPU's counters, runs its diagnostics and exports both to the monitoring stack, and nearly every fleet's dashboards and alerts are built on it. The skill is knowing which of its hundreds of fields carry signal: the profiling metrics that say whether the tensor cores are busy (not the utilization number everyone reads first), the error counters that predict a failure, the throttle reasons that explain a slow step, and the diagnostic levels that decide whether a node returns to the pool. This page walks those fields, derives an MFU estimate from them, and gives a fleet's alert thresholds.

Explore the full curriculum →

UP NEXT ON YOUR JOURNEY

Next in this trackThe model advertises a million tokens of context. What actually breaks when you serve it?Next in this trackThe model authors publish a new revision. How do you roll it out without a quality regression?Next in this trackDesign a platform that serves twenty open-weights models of varying size to internal teams.

FEDITOR'S NOTE

Scored on comparing against the fleet median, on separating replica-side causes from traffic-side ones, and on the routing skew that makes a healthy replica look slow.

DISCUSSION · 0

No comments yet — be the first to share your approach.