Expand description
2026-09-26: HardwareState, what the box was doing at one instant, and
HardwareStateDelta, what changed between the captures before and after a run.
Owner: bench hardware. Invariants:
- A reading that does not parse is
None, never a default number (the parsers insuper::parse). - A throttle counter that went backwards between captures yields
Nonein the delta, never zero.
super::Hardware answers “which box”; this answers “what state was that box in”.
Two boxes with identical fingerprints can differ in chassis temperature, throttle
time and clock under load, and only the state shows it.
Memory comes from /proc/meminfo: nvidia-smi reports GB10’s framebuffer usage as
N/A (fixtures/gb10_performance.txt). Cached is kept beside MemAvailable
rather than folded into it.
Structs§
- GpuCompute
App - 2026-09-26: One GPU-resident compute process, as
--query-compute-appsreports it. - Hardware
State - 2026-09-26: Everything the collector could read about the box at one instant.
The executor captures it before a run and again when the run’s terminal frame
is ready;
HardwareStateDeltais the difference. - Hardware
State Delta - 2026-09-26: What changed across the run.
- Machine
Identity - 2026-09-26: Which individual box this is.
super::Hardware::gate_keygives the silicon class ("gb10") that baselines are keyed by; this tells two boxes of one class apart. - Thermal
Zone - 2026-09-26: One
/sys/class/thermal/thermal_zone*reading. - Throttle
Active - 2026-09-26: Throttle reasons asserted at the instant of capture.
- Throttle
Counters - 2026-09-26: Cumulative throttle time per reason, in the microseconds
nvidia-smi -q -d PERFORMANCEreports underClocks Event Reasons Counters.