Rail-to-Ground Resistance: What a Multimeter Tells You About a GPU's Power Stages

A used GPU can arrive with a shorted power stage. Plug it in and the result ranges from a tripped power supply to a burnt board, and the evidence of what was wrong before power-on is gone. That is why no power is applied during Phase 1 of an Assurd certification. One of the last Phase 1 checks is a resistance reading from each power rail to ground, taken with an ordinary bench multimeter.

This post explains what that reading measures and how the Rules Engine judges it (SOP-GPU-001 v1.1, Step 1.5.2, as amended on 2026-11-01). It also sets out what the reading cannot tell you. The method is deliberately simple. Its limits are part of the method, and they are stated here as plainly as its uses.

What the meter measures

In resistance mode a digital multimeter pushes a small known test current through its probes, measures the voltage that develops, and displays the ratio. Everything between the two probe tips is part of that ratio:

$$R_{\text{meas}} = R_{\text{leads}} + R_{\text{path}} .$$

$R_{\text{leads}}$ is the resistance of the test leads and the probe contacts. $R_{\text{path}}$ is what we want: the resistance the meter's current finds between one power rail and the card's ground. On an unpowered card, that path runs through everything connected between the rail and ground:

• the rail's filter capacitors;
• the low-side switches of its voltage regulator;
• the silicon the rail feeds.

On a healthy card that path is not a short. A shorted capacitor or a failed switch connects the rail straight to ground, and the reading collapses towards zero. That is the failure the method exists to catch.

This is a 2-wire measurement: the same two leads carry the test current and sense the voltage. The bench meter required by the SOP (Uni-T UT61B+ or equivalent) resolves 0.1 Ω. Everything below follows from those two facts.

Zeroing the leads

The leads are not zero ohms. With 0.1 Ω resolution, a few tenths of an ohm of lead and contact resistance is a large share of the readings that matter. So before any rail is measured, the technician shorts the probe tips together, waits for the reading to settle and presses REL. From then on the meter displays

$$R_{\text{shown}} = R_{\text{meas}} - R_{\text{leads},0} ,$$

where $R_{\text{leads},0}$ is the reading captured at the moment REL was pressed. Every recorded value is relative to that zero. Why that matters is the first worked example below.

Probe placement and polarity

The black probe goes on ground: the I/O bracket or a ground pad. The red probe goes on the output side of the rail's inductor, the point where the regulated rail leaves its power stage.

Polarity is fixed because the path is not symmetric. Semiconductor junctions conduct one way, so reversing the probes sends the meter's current through a different path in the regulator and gives a different number. A reading taken the other way round is not comparable with one taken correctly, so the SOP fixes the polarity.

Settle time

When the probes touch, the meter's test current starts to charge the rail's capacitors, and the displayed value climbs as they charge. The technician holds the probes still until the reading stops moving, about 10 s, and records the settled value in ohms to 0.1 Ω.

A reading taken too early is too low. That cannot hide a short, but it can make a healthy rail look suspicious, which is why the Soft Limit below asks the technician to re-check the settle time first.

The six rails

Rail	Review floor	Measured on
GPU core (VCore)	0.3 Ω	every tier
GDDR memory VDD	1.0 Ω	every tier
VDDCI (I/O interface)	1.0 Ω	every tier
1.8 V (PLL / memory control)	10.0 Ω	Premium
PEX (PCIe interface)	2.0 Ω	Premium
5 V (logic / fan / RGB)	50.0 Ω	Premium

The three primary rails are measured on every tier: Standard, Premium and Re-Certification. The three forensic rails are required for Premium Certification.

The two input rails, 12 V and 3.3 V, are checked separately in diode mode (Step 1.5.1). This post covers the regulated output rails only.

Hard Limit: dead short

Before comparing, the Rules Engine rounds each reading to 0.1 Ω, the meter's resolution, and then applies the following rule:

$$R \le 0.1\ \Omega \ \Rightarrow\ \text{dead short}.$$

A dead short on any rail is a Hard Limit. The asset is quarantined, the card is not powered on, and the flag cannot be overridden.

Because the comparison uses the rounded value, a reading entered as 0.14 Ω is compared as 0.1 Ω and is a dead short. That matches what the meter's display would have shown. A negative reading after REL is also at or below 0.1 Ω, so it is treated the same way.

Soft Limit: below the review floor

A reading above 0.1 Ω but below the rail's review floor is a Soft Limit: a possible partial short. Something is pulling the rail towards ground harder than expected, but not hard enough to be a dead short. The technician re-checks REL, polarity and settle time. If the reading stands, they write a justification into the audit trail before any power is applied. A Soft Limit is never escalated to a Hard Limit.

Every other reading is recorded and shown on the certificate without a limit. In particular, a high reading is not flagged: on a 2-wire reading, a high value mostly says how far the capacitors had charged when the reading settled. There are no per-SKU reference values for these rails.

Why the VCore floor is 0.3 Ω

At 0.1 Ω resolution the meter can only display 0.0, 0.1, 0.2, 0.3 Ω and so on. The Hard Limit already covers 0.0 and 0.1. A Soft Limit band needs at least one displayable value strictly between 0.1 Ω and the floor:

• with a floor of 0.2 Ω, the band is $0.1 < R < 0.2$, which holds no displayable value, so the Soft Limit could never fire;
• with a floor of 0.3 Ω, the band holds exactly one value, 0.2 Ω.

So 0.3 Ω is the lowest VCore floor that gives the Soft Limit anything to do. The floors are set low on purpose, below the readings expected from a healthy card, so that a healthy card is rarely sent for review.

Worked examples

These are test vectors, not measurements of a real card. The "healthy" rail values are hypothetical numbers chosen to show the arithmetic, not reference values.

1. A short hidden by the leads. The technician's leads read 0.3 Ω when shorted together. The VCore rail is dead-shorted, $R_{\text{path}} = 0$.

• Without REL, the meter shows $0.3 + 0 = 0.3$ Ω. That is exactly the VCore floor, so the reading raises no flag at all: a dead short recorded as an unremarkable pass.
• With REL, the meter shows $0.3 - 0.3 = 0.0$ Ω: a dead short and a Hard Limit.

The zeroing step is the whole difference between the two outcomes.

2. Partial shorts in parallel. A leak to ground, such as a damaged capacitor, sits in parallel with the rail's normal path. Two resistances in parallel combine as

$$R = \frac{R_h\,R_l}{R_h + R_l} ,$$

where $R_h$ is the healthy path and $R_l$ the leak. Take a hypothetical memory rail with $R_h = 4.0$ Ω (floor 1.0 Ω):

Leak $R_l$	Combined $R$	Shown	Result
none	4.000 Ω	4.0 Ω	no flag
20 Ω	3.333 Ω	3.3 Ω	no flag
0.5 Ω	0.444 Ω	0.4 Ω	Soft Limit
0.05 Ω	0.049 Ω	0.0 Ω	Hard Limit

The last two rows are what the method is for. The 20 Ω row is its limit: a leak 5 times the healthy path moves the reading from 4.0 to 3.3 Ω. That is clearly visible on the display, but with no reference value to compare against, nothing in it says "fault".

3. A typed value between display steps. A reading entered as 0.14 Ω is compared at the meter's resolution as 0.1 Ω: a dead short.

What the method cannot see

A 2-wire reading at 0.1 Ω resolution detects dead shorts and gross partial shorts before power is applied. That is the whole claim. In particular, it does not tell you the following:

• Per-phase health. A GPU's core regulator has several phases in parallel. The reading sees the rail as a whole. One weak phase among several healthy ones does not stand out.
• Milliohm-level degradation. One display step is 0.1 Ω, which is 100 mΩ. Suppose some part of the power stage doubled its resistance from 2 mΩ to 4 mΩ. The reading would move by 0.002 Ω, a fiftieth of one display step. A change that small is invisible to this method.
• A comparison with a reference card. There are no per-SKU reference values yet. A reading is judged only against the dead-short limit and its rail's review floor, never against "what this model should read".
• Behaviour under load. The card is unpowered. How the regulator behaves at full current is measured later, in Phase 3, through the board-power trace.

An Assurd certificate does not claim any of these.

Calibration status

The 0.1 Ω dead-short limit and all six review floors are engineering priors pending fleet calibration (n ≈ 300). They are set conservatively from the physics of a 2-wire reading, not fitted to a population. We do not describe them as empirical or statistically validated.

Once the fleet holds roughly 300 certified cards, the floors will be reviewed against the observed readings per rail. Any change will be versioned in the SOP.

See it on a certificate

On a certificate, the Electrical continuity tile shows each rail's reading to 0.1 Ω, and its detail view adds each rail's review floor. The sample certificate shows that layout. Its readings are illustrative: no card was measured.

Phase 1 is the pre-power filter. What a card does once it is powered and loaded for 30 minutes is covered in Hotspot Delta and TIM Pump-Out and The Thermal Time Constant and the 1800 s Trace.

Notation

Symbol	Meaning	Unit
$R_{\text{meas}}$	What the meter measures between its probe tips	Ω
$R_{\text{leads}}$	Lead and contact resistance	Ω
$R_{\text{leads},0}$	Lead reading captured when REL is pressed	Ω
$R_{\text{path}}$	Rail-to-ground path resistance on the card	Ω
$R_{\text{shown}}$, $R$	Displayed (recorded) reading, compared at 0.1 Ω resolution	Ω
$R_h$, $R_l$	Healthy path and parallel leak path (worked example)	Ω