Throttle model v1.0
How every number on this site is produced
This model is deterministic and published in full. Given the same processor and the same board it returns the same verdict every time, on every page, in the calculator and in the database — and you can check its arithmetic, which is the point of writing it down.
By Farrukh Sufyan Sarwar · Last verified 19 September 2026 · Model v1.0 · Build 2026-09-19
The constants
Six numbers decide everything. They are the model's judgement calls, and they are the right place to argue with it.
- Effective Vcore
- 1.25 V
- Converts phase current into watts
- Interior case air
- 38 °C
- Around the socket under load
- Temperature slope
- 65 °C
- Rise per unit of stress ratio
- Moderate threshold
- 0.65
- Stress ratio where clocks start to slip
- Severe threshold
- 0.85
- Stress ratio where the board is the limit
- Temperature cap
- 125 °C
- Above this the model stops reporting a number
Phase design factor
How much of a stage's rating a design actually delivers.
- true phases
- × 1.00
- doubled phases
- × 0.90
- teamed phases
- × 0.85
Heatsink derating
An editorial class, not a manufacturer figure: none means bare MOSFETs, heavy means large finned or heatpipe-linked blocks.
- none
- × 0.45
- light
- × 0.70
- medium
- × 0.90
- heavy
- × 1.00
The formulas
Rule 1
Socket check comes first
If the processor's socket and the board's socket differ, nothing else is computed. Capacity is zero, no temperature is reported, and the page becomes a page about what fits instead. A model that reports a temperature for a pairing that cannot physically exist is not being careful.
cpu.socket ≠ board.socket → INCOMPATIBLE_SOCKET
Rule 2
Sustained power capacity
Vcore power stages only. Vendor specifications like “16+1+2” add the SoC and graphics phases to reach a bigger number, and those phases carry none of the core load, so counting them would flatter every board equally and tell you nothing.
capacity = stages × amps × 1.25 V × phaseFactor × heatsinkFactor
Rule 3
Stress ratio uses sustained draw, not boost
This is the decision that makes the model useful. A 65 W Ryzen has an 88 W boost ceiling it reaches for seconds, and using that ceiling would condemn every entry board for no reason. What matters is what the processor holds for ten minutes.
stressRatio = cpu.sustainedWatts ÷ capacity
Rule 4
Estimated VRM temperature
Interior case air plus the stress ratio scaled by the slope, capped at 125 °C. It is an estimate of a band, not a reading: real numbers move with case airflow, room temperature and how the board's sensor is placed.
temperature = 38 °C + stressRatio × 65 °C, capped at 125 °C
Rule 5
Risk band and expected clock loss
Two thresholds, each with its own loss curve. Below 0.65 the model expects nothing. The severe curve is capped at 28 % because past that point the behaviour stops being throttling and starts being a machine that should not have been built.
ratio > 0.85 → severe, drop = min(28, (ratio − 0.80) × 50) · ratio > 0.65 → moderate, drop = min(10, (ratio − 0.60) × 20)
Rule 6
Capacity score
A single 0–100 figure for ranking boards against one processor. It is a rescaling of the stress ratio, not new information, and it exists so that tables can be sorted.
score = round(max(10, 100 − stressRatio × 60)), or 0 for a socket mismatch
Rule 7
BIOS updates by generation rank, never by name
Each processor carries a generation rank that increases with launch date within its socket, and each board carries the highest rank its shipping firmware supports. An update is needed when the first exceeds the second. Inferring this from a model number — “does the name contain 14?” — breaks the moment a vendor renames something.
requiresBiosUpdate = cpu.generationRank > board.shipsBiosGenRank
Rule 8
Memory generation and speed
Memory generation is a hard check: a DDR4 board and a DDR5-only processor cannot share a kit, and nothing in the BIOS changes that. Speed is a note rather than a failure, raised when a board validates below the platform's sweet spot — AM5 at DDR5-6000, LGA1700 at DDR5-6400, LGA1851 at DDR5-8000.
memoryCompatible = cpu.memoryTypes includes board.memoryType
What this model does not claim
It is not a measurement. Nobody here put a thermocouple on 80 boards, and any site claiming otherwise for a catalogue this size is not telling you the truth. What this is: a consistent calculation from published specifications, applied identically to every pairing, so that comparisons between boards are meaningful even where the absolute temperature is approximate.
It does not model your case. Airflow over the VRM changes real temperatures by ten or fifteen degrees, and a board in an open bench sits cooler than the same board behind a glass front panel. The 38 °C ambient figure is a typical closed case with a working intake.
It does not model overclocking or undervolting. Every figure assumes stock settings with power limits as the board ships them. Raising a power limit moves the processor's sustained draw and therefore the entire verdict; lowering it is the cheapest fix for a severe verdict, and it costs less performance than the throttling would.
Where a specification could not be verified, the product is left out of the catalogue rather than estimated. That is why some boards you expect to see are missing.
Where the data comes from
Processor specifications — core counts, TDP, boost ceilings, memory support — come from Intel's and AMD's own product pages. Sustained package power is the measured all-core figure at stock, which for Intel non-K parts is the base power limit and for K-series parts on a board with unlimited tau is close to the maximum turbo power.
Board VRM configurations come from the manufacturer's published power-stage specification where they publish one, and from VRM comparison tables that identify the exact power-stage part otherwise. Each board's source is recorded in the catalogue. Heatsink mass class is our own classification and is the most subjective input in the model, which is why its factors are printed above rather than hidden.
Prices and availability come from the Amazon Product API and are refreshed nightly. The date of the last refresh is shown next to every price.