Recurring mistakes in GPU sizing, storage IOPS, networking, schedulers, and operations staffing
Optiscale Senior Engineering Team · Scheduled for 2026 Q2
Drawing on involvement in RFPs, deployment, and operations across some 30 domestic HPC/AI infrastructure projects, this paper lays out the five most recurring design and operations pitfalls alongside measurable criteria. For each pitfall, it provides checklist items and verification benchmarks that can head off the problem at the RFP stage.
Three-axis sizing per workload: memory bandwidth, NVLink topology, and power efficiency. H100/H200/B200 comparison table.
For AI training datasets, metadata ops are more decisive than sequential bandwidth. The impact of NSD separation.
ECMP collisions on NCCL collectives. The difference rail-optimized + adaptive routing makes.
How backfill can indefinitely defer short jobs. The pitfalls of Slurm priority and fair-share settings.
Cases that omit the 8–12%/year operations-cost assumption. The savings automation delivers.