Training reads huge datasets, and GPUs sit idle waiting for data. What should you check?
Why D is the answer
Storage throughput
Slow storage starves GPUs.
In detail
- Slow storage starves GPUs of data.
- Low GPU use can point to storage.
Use case
An illustrative scenario, not a real incident.
Expensive GPUs sit idle, because storage cannot deliver data fast enough.
See it in different ways
mindmap
root((AI cluster))
Compute
GPU nodes
NVLink
Fabric
InfiniBand
Ethernet
Storage
High throughput
Management
Out-of-band networkflowchart LR S0["Low GPU utilization"] --> S1["Check storage throughput"] --> S2["Fix the bottleneck"]
Key point: Slow storage starves GPUs.
Why the other options are wrong
- A. The management network speed
- That network is not in the data path.
- B. The GPU driver version
- A driver does not cause waiting for data.
- C. The BMC firmware version
- BMC firmware is not in the data path.