GPU Telemetry & Power-Efficiency Metrics — Dataflow

How per-vendor GPU power/energy signals flow from the collectors to the console. NVIDIA and AMD stay in separate lanes end-to-end: they converge only in the shared TelemetryHierarchy store, then re-split at the per-platform aggregation so each vendor's totals sum only its own GPUs.

NVIDIA lane (nvidia_*, DCGM / pynvml) AMD lane (amd_*, amdsmi) Shared / cross-vendor
1 · Produce
NVIDIA
DCGMTelemetryCollector Scrapes DCGM /metrics; DCGM_TO_FIELD_MAPPING → nvidia_*
PyNVMLTelemetryCollector Local NVML → nvidia_power_usage, nvidia_energy_consumption, …
AMD
AMDSMITelemetryCollector amdsmi → amd_power (W), amd_energy_consumption (MJ)
TelemetryRecord + TelemetryMetrics Each record stamps platform (nvidia/amd) + vendor fields
2 · Converge (shared store)
both lanes flow in →
GPUTelemetryAccumulator
.process_telemetry_record() Routes every record into the hierarchy
TelemetryHierarchy telemetry_source_url → gpu_uuid → GpuTelemetryData time series. Holds both vendors, each GPU tagged by metadata.platform.
3 · Aggregate (re-split per vendor)
compute_efficiency_metrics() Once per profiling phase. Loops _EFFICIENCY_VENDORS (NVIDIA, AMD).
→ per-platform filter splits back into lanes →
NVIDIA _sum_gpu_power_watts / _sum_gpu_energy_joules platform=nvidia, fields NVIDIA_POWER_USAGE_FIELD / NVIDIA_ENERGY_CONSUMPTION_FIELD
AMD _sum_gpu_power_watts / _sum_gpu_energy_joules platform=amd, fields AMD_POWER_FIELD / AMD_ENERGY_CONSUMPTION_FIELD
shared inputs total_output_tokens (records) + concurrency (config) feed tokens/J and J/user for both lanes
4 · Vendor MetricResults
NVIDIA nvidia_total_gpu_power
nvidia_total_gpu_energy
nvidia_output_tokens_per_joule
nvidia_energy_per_user
AMD amd_total_gpu_power
amd_total_gpu_energy
amd_output_tokens_per_joule
amd_energy_per_user
console_group GPU_POWER_EFFICIENCY_NVIDIA / _AMD route each tag to its section
5 · Render
ConsoleGpuVendorDisclaimerExporter Vendor warning banner — heads all GPU sections
ConsoleNvidiaPowerEfficiencyExporter “GPU Power Efficiency (NVIDIA)”, avg-only
ConsoleAmdPowerEfficiencyExporter “GPU Power Efficiency (AMD)”, avg-only
GPUTelemetryConsoleExporter Per-GPU telemetry tables
CSV / JSON exporters All vendor tags persisted to artifacts

Resulting console order

The disclaimer banner renders first, so it's clear everything below it is vendor-specific.

GPU Power Efficiency (NVIDIA)   Total GPU Power · Total GPU Energy · Output Tokens per Joule · Energy per User (avg only)
GPU Power Efficiency (AMD)   Total GPU Power · Total GPU Energy · Output Tokens per Joule · Energy per User (avg only)
AIPerf | GPU Telemetry Summary   per-GPU tables (power, util, mem, temp, …)