Operations

Fleet KPIs and What You Are Measured On

The five numbers operations grades you on - availability, realized hashrate, J/TH, reject rate, MTTR - and why a hot miner quietly wrecks efficiency.

The scorecard

Operations does not grade you on how many miners you touched. It grades the fleet on five numbers. Know each one, know its target, and know which unit is dragging it down.

KPI	What it means	Healthy target
Availability / uptime	Percent of deployed time a unit is actually hashing	~95-99%
Realized hashrate %	Rolling-average real hashrate vs nameplate	Close to 100%
Efficiency (J/TH)	Watts / TH/s - lower is better	~29.5 / 17.5 / 15 (S19j Pro / S21 / S21 Pro)
Reject rate	Share of submitted shares the pool throws out	Under ~1% (under 0.5% excellent)
MTTR	Avg time from fault to back-in-service	As low as you can drive it

A unit that is not hashing scores zero on every KPI at once - availability is the metric that eats all the others. Get it back online before you optimize anything else.

Efficiency and heat

J/TH = watts / TH/s. It is the number that decides whether a unit earns its power bill, and heat is what quietly ruins it.

Same hashrate, hotter chips = more watts. Braiins measured an S19j Pro go from ~24 J/TH at 20 °C to ~34 J/TH at 75 °C - a ~42% penalty at constant hashrate.
Firmware throttles hot silicon to protect it, so the hashrate can still look fine while J/TH creeps up. Watch power draw, not just TH/s.
A hot miner quietly costs ~40% more per TH. Airflow, dust, and inlet temp are efficiency problems, not just temperature problems.
Deliberate undervolting improves J/TH (fewer watts per TH). Aggressive overclocking degrades it - you buy hashrate with worse efficiency and more heat.

S19j Pro chip temp	J/TH	vs cool
20 °C	~24	baseline
75 °C	~34	~42% worse

Reject rate and realized hashrate

Two numbers that tell you whether a unit is really producing what its nameplate promises.

Reject rate = stale/rejected shares the pool discards. Under 0.5% is excellent, 0.5-1% acceptable, over ~1% means chase it - flaky pool link, network latency, or overclock instability.
Realized hashrate % = rolling-average actual hashrate vs nameplate. A 100 TH unit averaging 92 TH is realizing 92% - the missing 8% is a dead board, throttling, or short outages.
High rejects burn real work: the chips hashed but the pool paid nothing for it, so it drags efficiency and realized hashrate together.

Work the units that move the needle

Prioritize the units that move these metrics. This is the language leadership speaks - report in numbers, not anecdotes.

One dead 100 TH unit hurts availability and realized hashrate more than ten units each down 2%. Triage the biggest loss first.
MTTR is a team metric: stocked spares, an accurate site map, and fast diagnosis all shrink it. A confident swap beats a slow investigation.
When you report up, talk in KPIs: 'row B back to 98% availability, rejects under 0.4%' - not 'fixed some miners'.

Every ticket you close should move one of these five numbers. If it does not, ask why it is on your list.

The scorecard

Operations does not grade you on how many miners you touched. It grades the fleet on five numbers. Know each one, know its target, and know which unit is dragging it down.

KPI	What it means	Healthy target
Availability / uptime	Percent of deployed time a unit is actually hashing	~95-99%
Realized hashrate %	Rolling-average real hashrate vs nameplate	Close to 100%
Efficiency (J/TH)	Watts / TH/s - lower is better	~29.5 / 17.5 / 15 (S19j Pro / S21 / S21 Pro)
Reject rate	Share of submitted shares the pool throws out	Under ~1% (under 0.5% excellent)
MTTR	Avg time from fault to back-in-service	As low as you can drive it

A unit that is not hashing scores zero on every KPI at once - availability is the metric that eats all the others. Get it back online before you optimize anything else.

Efficiency and heat

J/TH = watts / TH/s. It is the number that decides whether a unit earns its power bill, and heat is what quietly ruins it.

Same hashrate, hotter chips = more watts. Braiins measured an S19j Pro go from ~24 J/TH at 20 °C to ~34 J/TH at 75 °C - a ~42% penalty at constant hashrate.

Firmware throttles hot silicon to protect it, so the hashrate can still look fine while J/TH creeps up. Watch power draw, not just TH/s.

A hot miner quietly costs ~40% more per TH. Airflow, dust, and inlet temp are efficiency problems, not just temperature problems.

Deliberate undervolting improves J/TH (fewer watts per TH). Aggressive overclocking degrades it - you buy hashrate with worse efficiency and more heat.

S19j Pro chip temp	J/TH	vs cool
20 °C	~24	baseline
75 °C	~34	~42% worse

Reject rate and realized hashrate

Two numbers that tell you whether a unit is really producing what its nameplate promises.

Reject rate = stale/rejected shares the pool discards. Under 0.5% is excellent, 0.5-1% acceptable, over ~1% means chase it - flaky pool link, network latency, or overclock instability.

Realized hashrate % = rolling-average actual hashrate vs nameplate. A 100 TH unit averaging 92 TH is realizing 92% - the missing 8% is a dead board, throttling, or short outages.

High rejects burn real work: the chips hashed but the pool paid nothing for it, so it drags efficiency and realized hashrate together.

Work the units that move the needle

Prioritize the units that move these metrics. This is the language leadership speaks - report in numbers, not anecdotes.

One dead 100 TH unit hurts availability and realized hashrate more than ten units each down 2%. Triage the biggest loss first.

MTTR is a team metric: stocked spares, an accurate site map, and fast diagnosis all shrink it. A confident swap beats a slow investigation.

When you report up, talk in KPIs: 'row B back to 98% availability, rejects under 0.4%' - not 'fixed some miners'.

Every ticket you close should move one of these five numbers. If it does not, ask why it is on your list.