Data Center Compute Supply Efficiency

Data Center Engineer — Resource Efficiency (Compute Supply)

When to Use

• Measure and improve utilization of racks, kW, and compute (CPU/GPU/memory)
• Forecast compute supply: how many nodes/GPUs needed by quarter
• Find stranded capacity (power allocated but unused, empty U, low CPU%)
• Plan hardware refresh, standard builds, and end-of-life decommission
• Consolidate workloads to free racks or defer capex
• Set power caps and placement rules for efficiency without breaching SLAs
• Report efficiency KPIs to finance, sustainability, and engineering leadership
• Compare efficiency of keeping workloads on-prem vs shifting burst to cloud

When NOT to Use

• New hall design, MEP, commissioning → data-center-design-execution-lead
• Helm, cluster upgrades, pod debug → cluster-deployment-engineer
• VPC, Terraform, managed cloud architecture → infrastructure-engineer
• AI inference token/cost roadmap → ai-token-improvement-plan-engineer
• AI production ops cadence → ai-lead-ops
• Multi-vendor DC construction program → technical-program-manager
• Multi-site DC roadmap and capex prioritization → data-center-portfolio-planning-execution-lead

Related skills

Core Workflows

1. Baseline efficiency metrics

Establish dashboards for:

• Facility: PUE, total IT kW, cooling kW
• Supply: rack count, kW committed vs kW used, GPU/CPU inventory
• Demand: avg/peak utilization, useful work per kW (define numerator per org)
• Waste: idle hosts, powered empty U, oversubscribed cooling margin

See `references/efficiency_metrics.md`.

2. Compute supply planning

1. Demand — workload growth, new products, GPU training vs inference mix 2. Supply — on-hand, on-order, lead times, standard SKUs 3. Gap — quarter-by-quarter surplus or deficit 4. Actions — buy, refresh, cloud burst, or defer

See `references/compute_capacity_supply.md`.

3. Utilization and consolidation

• Inventory hosts below utilization threshold for 30+ days
• Plan migration windows; validate performance tests post-move
• Target: raise average utilization without violating HA or latency SLOs
• Virtualization or K8s density changes → coordinate with cluster-deployment-engineer

See `references/utilization_optimization.md`.

4. Power and thermal efficiency

• Align rack kW nameplate with actual draw; recover stranded breaker capacity
• Power capping policies (OS/firmware/IPMI) where SLA allows
• Match GPU trays to cooling class (air vs liquid)

See `references/power_thermal_management.md`.

5. Hardware lifecycle

See `references/hardware_lifecycle.md`.

6. Reporting and targets

• Monthly: utilization, PUE trend, supply vs demand
• Quarterly: refresh plan, capex avoidance from consolidation
• Tie narratives to sustainability goals without greenwashing

See `references/reporting_targets.md`.

Output standards

• Supply/demand table by quarter (nodes, kW, GPUs)
• Top 10 stranded assets with recommended action and risk
• Efficiency initiative backlog with estimated kW or capex saved
• Assumptions explicit (utilization window, SLA exclusions)

When to load references

• KPIs and formulas → references/efficiency_metrics.md
• Forecast and procurement → references/compute_capacity_supply.md
• Consolidation and right-size → references/utilization_optimization.md
• Power caps and cooling fit → references/power_thermal_management.md
• Refresh and decomm → references/hardware_lifecycle.md
• Dashboards and targets → references/reporting_targets.md

Compute Capacity and Supply Planning

Supply chain view

Demand forecast → Required capacity → Supply plan → Deploy → Measure → Adjust

Demand inputs

Supply inputs

Standard builds

Reduce supply fragmentation:

• 2–4 server SKUs (general, storage, GPU tier)
• Document: CPU, RAM, NIC, GPU, PSU kW, form factor
• Spare parts aligned to SKUs

Forecast template (quarterly)

GPU / accelerator supply

• Separate training (burst, high power) vs inference (steady, latency-sensitive)
• Track allocation queue — projects waiting for GPUs
• Model: effective GPUs = physical × utilization × scheduling efficiency

Capex vs efficiency tradeoffs

Handoffs

• New kW or halls → data-center-design-execution-lead
• Workload placement on K8s → cluster-deployment-engineer
• Hybrid cloud capacity → infrastructure-engineer

Efficiency Metrics

Facility layer

Compute supply layer

Demand / usefulness layer

Define useful work per organization—examples:

• vCPU-hours consumed (virtualization)
• GPU-hours scheduled (training/inference)
• Completed batch jobs per kW
• Inference requests per watt (requires app telemetry)

Derived efficiency KPIs

• kW per useful GPU-hour — lower is better (same SLA)
• Servers per rack at target kW — density efficiency
• Capex per unit useful capacity — supply efficiency
• Refresh gain — perf/watt delta gen-over-gen

Data sources

• DCIM, PDU, environmental sensors
• vCenter, hypervisor, bare-metal inventory
• Kubernetes metrics (node allocatable vs requests) — coordinate with cluster team
• CMDB / asset tags (owner, environment, criticality)
• Cloud billing for hybrid comparison (not primary for this skill)

Anti-patterns

• Optimizing average CPU% alone while latency SLOs fail
• Ignoring committed kW when colo charges by allocation
• Mixing dev idle hosts with production peaks in one average

Hardware Lifecycle

Stages

Refresh decision matrix

Decommission checklist

• [ ] Workloads migrated or retired
• [ ] Data destroyed per policy
• [ ] Licenses reclaimed
• [ ] Removed from monitoring and CMDB
• [ ] Powered off; PDU port documented free
• [ ] Spares or resale disposition
• [ ] kW and U added back to supply pool

Spares and inventory

• Spares are insurance, not efficiency — cap % of fleet
• Rotate spares through test lab to avoid stale firmware

GPU-specific

• Track firmware, driver, and CUDA stacks per generation
• Pool allocation: shared vs dedicated per team
• End-of-life: resale market, donation policy, destruction certs

E-waste and compliance

• Vendor takeback or certified recycler
• Chain of custody for storage media
• Environmental reporting → compliance-engineer if regulated

Power and Thermal Management

Power hierarchy

Utility → UPS/PDU → Rack PDU → Server PSU → Components

Efficiency work focuses on matching IT load to committed capacity and perf per watt.

Stranded power recovery

1. Measure actual draw per rack (not nameplate alone) 2. Compare to breaker / colo contract kW 3. Identify racks below policy (e.g. < 40% of committed) 4. Options:

• Add workloads (consolidation)
• Lower committed allocation with colo (contract change)
• Rebalance across rows

Power capping

Document SLA exclusions where caps are forbidden.

Thermal and density

Mismatch → throttling, reduced useful work per kW.

Coordination with facility

• Cooling setpoints — don't lower PUE on paper while IT throttles
• Row balance — avoid hot spots from uneven load
• New density → data-center-design-execution-lead if beyond existing class

Tariff and carbon (optional)

• Time-of-use: shift batch to off-peak
• Carbon intensity by grid hour — report separately from PUE
• On-site renewables or PPAs — sustainability narrative

Reporting and Targets

Audience matrix

Monthly operational pack

1. Supply — inventory, incoming POs, GPU queue depth 2. Demand — utilization trends, new requests 3. Gap — surplus/deficit next 2 quarters 4. Actions — consolidations completed, kW reclaimed 5. Risks — lead times, cooling limits, SLA conflicts

Target setting

Targets must be achievable and excluded where SLAs forbid consolidation.

Initiative backlog fields

Narrative quality

• Lead with business outcome (deferred rack purchase, shorter GPU wait)
• Quantify uncertainty (± range on savings)
• Separate one-time reclaim vs run-rate efficiency

Related comms

• Major sustainability claims → communication-lead for review
• Multi-site program → technical-program-manager

Utilization and Optimization

Identify candidates

Always exclude hosts with:

• HA pairs where split would violate policy
• Latency-sensitive dedicated inference
• Regulatory isolation requirements

Consolidation workflow

1. Inventory — CMDB + monitoring + owner tag 2. Classify — prod / staging / lab / unknown 3. Model target — destination host or cluster capacity 4. Test — load test or canary migration 5. Migrate — maintenance window, rollback plan 6. Verify — SLO, power draw, license compliance 7. Decomm or repurpose — free kW and U

Virtualization and containers

• VM density — vCPU:pCPU ratios; avoid extreme oversubscription on latency workloads
• K8s — requests/limits vs node allocatable; bin-packing efficiency
• Bare metal — often lowest utilization; justify with performance or compliance

Coordinate implementation with cluster-deployment-engineer or virtualization admins.

Scheduling for efficiency (DC-aware)

• Batch jobs in off-peak power windows (if tariff or PUE varies)
• GPU time-slicing or MIG where applicable
• Queue training jobs to fill idle accelerators

Risk controls

Success metrics

• kW reclaimed
• Racks deferred (capex avoidance)
• Utilization p50/p95 improved
• No increase in incident rate post-change