CONCEPT Cited by 1 source
GPU observability¶
GPU observability is the discipline of correlating accelerator health and performance with the compute, memory, network, Kubernetes, and job-level signals that govern a distributed-training run. A GPU-utilisation aggregate alone is insufficient: a performance drop can originate in GPU memory pressure, network throughput, pod placement, node readiness, resource allocation, power or thermal conditions, or hardware health. (Source: sources/2026-08-13-aws-adobe-firefly-simplified-observability-with-amazon-managed-prometheus)
Why it is different from CPU-fleet monitoring¶
A distributed GPU job needs visibility across at least three interacting layers:
- Job layer — GPU utilisation, memory consumption, and network throughput per training job identify bottlenecks in distributed computation.
- Kubernetes layer — pod status, node readiness, and resource allocation influence placement and scheduler decisions.
- Accelerator-health layer — GPU health determines whether a device must be cordoned and replaced.
Those dimensions create many label combinations, so GPU observability often runs directly into metric-cardinality and time-series query-scaling limits.
Operational implication¶
The relevant outcome is a sufficiently long and fast query window to correlate a job's full lifecycle with infrastructure events. In Adobe Firefly, moving critical GPU metrics to AMP grew the practical diagnostic window from six to 24 hours and made 24-hour GPU-utilisation queries 28.8× faster. (Source: sources/2026-08-13-aws-adobe-firefly-simplified-observability-with-amazon-managed-prometheus)
Relationship to GPU failures¶
GPU observability supplies the input signals that make GPU training failure modes detectable. The two concepts are distinct: failure modes classify what breaks, whereas GPU observability specifies the cross-layer telemetry needed to find and explain an active or impending problem.
Seen in¶
- sources/2026-08-13-aws-adobe-firefly-simplified-observability-with-amazon-managed-prometheus — job, Kubernetes, and GPU-health signals for Firefly training on EKS.