GPU, energy and CO2 metrics

Node only. The browser and React Native entries report no GPU, energy or CO2 metric: WebGPU exposes no utilisation or power, and the device's own counters are not reachable.

What is measured, and how

nvidia-smi --query-gpu=... --format=csv,noheader,nounits is run every GENAI_GPU_COLLECTION_INTERVAL seconds (default 5), off the request path, with a 10-second timeout. One line per device. The gauges report the latest sample. flushTelemetry() waits for the first sample, so a run shorter than one nvidia-smi call still exports a GPU reading.

The timeout is the SDK's own, not execFile's. execFile's timeout is a timer: when the application blocks the event loop past it while nvidia-smi runs - an ES-module application evaluating its modules right after instrument() - the timer fires before the child's exit and output are read, the finished child is "killed", and its output is lost with no error. Up to 0.1.0-dev that read as "no GPU on this host", logged at info level, and the process reported no GPU telemetry at all. Here a timer that fires defers to after pending I/O, and a child that has exited is never killed.

Metric	Unit	From
gen_ai.gpu.utilization, gen_ai.gpu.memory.utilization	%	utilization.gpu, utilization.memory
gen_ai.gpu.memory.used, gen_ai.gpu.memory.total	MiB	memory.used, memory.total
gen_ai.gpu.temperature	Cel	temperature.gpu
gen_ai.gpu.power, gen_ai.gpu.power.limit	W	power.draw, power.limit
gen_ai.gpu.clock.sm, gen_ai.gpu.clock.memory	MHz	clocks.current.sm, clocks.current.memory
gen_ai.gpu.fan.speed	%	fan.speed
gen_ai.gpu.performance.state	0 to 15	pstate
gen_ai.gpu.throttle.thermal, .power, .hw_slowdown	1 or 0	clocks_throttle_reasons.*
gen_ai.gpu.ecc.errors.corrected, .uncorrected	count	ecc.errors.*.volatile.total
gen_ai.gpu.aggregate.mean_utilization, .total_memory_used (GiB), .total_power, .max_temperature		over the devices that reported the value; gpu_count says how many

Per-device points carry gpu_id and gpu_name.

From power, between two samples:

Metric
gen_ai.energy.consumed (component = gpu), gen_ai.energy.total	kWh: mean power of the interval's two ends, times its length
gen_ai.power.consumption	W, a histogram of the readings
gen_ai.power.cost	USD: that energy times GENAI_POWER_COST_PER_KWH
gen_ai.co2.emissions, gen_ai.co2.emissions_rate	gCO2e, gCO2e/s: that energy times GENAI_CARBON_INTENSITY. Only with GENAI_ENABLE_CO2_TRACKING=true

What is not measured is not reported

Situation	What happens
No nvidia-smi on the host	No GPU instrument is created. One line at info level. Not asked again
nvidia-smi answers "No devices were found"	As a host with no GPU: no instrument, one line at info level
nvidia-smi present and failing at start (no driver, a timeout, no output at all)	No instrument yet. One warning with what the tool said; the first sample is tried again every interval, and after 5 failures in a row one more warning and no further attempt. An empty answer is a failure, never "no GPU"
A field the device reports as [N/A] (on a laptop GPU: power limit, fan speed, ECC)	That gauge has no point for that device. The others do
An older driver that does not know one of the fields	The core set (utilisation, memory, temperature, power) is used from then on
The tool starts failing later	Nothing is observed for that sample, and the energy of the gap is not integrated. After 5 failed samples in a row collection stops, with one warning
A device with no power reading	No energy, cost or CO2 for it
CPU and RAM energy	Not measured, so not reported. gen_ai.energy.total is GPU energy only, and says nothing else

A collector that stopped shows as a gap, not as a flat line: the exporter asks the metrics SDK for delta temporality for gauges (src/gauge-temporality.ts), so a gauge nobody observes any more is not exported. Under the SDK's default cumulative temporality it would go on exporting its last value. A gauge has no temporality on the wire, so nothing else changes.

CO2 is a factor, not a reading of the grid

gen_ai.co2.emissions is measured energy multiplied by one number, GENAI_CARBON_INTENSITY, in gCO2e per kWh. Its default, 475, is a world average and is the reference implementation's default. It is not the intensity of the grid the host is on. Set it to that grid's figure, or treat the metric as an order of magnitude.

The reference uses codecarbon, which carries regional grid data and also measures CPU and RAM. None of that exists here. The GENAI_CO2_* location and mode variables are accepted and have no effect; setting one says so in the log at start-up.

Differences from the reference

• The reference's own formula for energy from power is 3.6 times too high (it multiplies kilowatts by seconds and treats the result as watt-hours); it affects its gen_ai.power.cost always, and its CO2 when codecarbon is not installed. Here the energy is watts times seconds over 3 600 000. Confirmed, SDK_ADDITIONS.md section 9.
• gen_ai.gpu.pcie.tx / .rx and gen_ai.codecarbon.task.duration are never reported.
• AMD GPUs are not read. The reference reads them through the amdsmi Python binding; the command-line tool's output could not be checked on any machine available here.

What was run

The parser and the collector are tested against the real output of nvidia-smi on one machine (RTX 5090 Laptop GPU, driver 592.02, Windows), and the real tool was run through NvidiaSmiSource there. A two-device case uses a hand-written fixture in the same format. Not run: a multi-GPU host, a data-centre GPU, Linux, an older driver.

Another source of readings

setGpuSource(() => source) replaces nvidia-smi with anything that implements sample(): Promise<GpuSample[] | undefined> - a vendor tool, an exporter already running on the host. A factory that returns undefined switches GPU reading off.