Post is modified on an ongoing basis, last updated 202507
March 2020. The onset of the coronavirus results in a sudden demand for computer graphics cards as people are confined to their homes and seek entertainment from their computers. While most are locked out of the new GPU market with prices rising more than 200%, a few in the repair community saw a chance to start creating some value out of junk on the used market. From here begins my long journey into the electronics repair rabbit hole.
PCIe Knowledge Check
Standard PCIe Supply Voltages
On most PCIe cards, regardless of their peripheral type, a series of pins on the PCIe gold finger provide voltages direct from PC (ATX or otherwise) power for:
- 12V (PCIe finger)
- 3.3V
High power peripherals generally receive majority of their 12V power via the PCIe 6-Pin or 8-Pin finger but there are other defined standards for additional power:
- PCIe 6-Pin
~75W(as the middle pin is not guaranteed to provide power.) - PCIe 8-Pin
~160-252W- Note the last 2 pins for the extension are sense pins to indicate that power can actually be drawn from the middle 12V wire
- EPS 12V
~336W, often found on datacenter GPUs (DO NOT CONFUSE THIS WITH the PCIe 8-PIN!!)
PCIe Lanes
PCIe uses 1 to 16 lanes initialized from the leftmost side of the slot to the rightmost. Bifurcation and lane reversal is another rabbit hole which I will need a whole new article to cover.
GPU Power Up Basics
On board voltage rails
On standard PCI Express based graphics cards, the modern GPU die is rather complex with multiple voltage rails. These rails generally power up sequentially and generally only after signals from the other voltage rails are received.
We established above that there are only 2 voltages coming into the GPU: 12V and 3.3V. From these, generally auxillary voltage rails are generated as below
- 1.8V (through 3.3V rail)
- 5V (through the 12V rail)
For critical GPU voltage rails we have:
- PCIe Rail (PEXVDD)
1.2V - Memory Voltage (Frame Buffer FBVDD)
~1.5V - Core Voltage (NVVDD)
~1V
Resistance measurements
All voltage rails mentioned above should generally have a resistance above 50 ohms EXCEPT for NVVDD which has a resistance low enough that there isn’t a useful reading out of a multimeter.
Notes from typical resistance measurements I’ve seen:
- 12V
1-3kohms - 3.3V
1kohms - 5V
>1kohm - PEXVDD
100ohms - FBVDD
30-100ohms - NVVDD
0-1ohms
Since core resistance is tiny, it’s much more useful to measure a short between 12V and the NVVDD VRM which we will cover below.
Core and Frame Buffer Voltage Generation
GPU cores have such low resistances that the power going through them needs to be supplied with a large amount of current by ohm’s law to get the power targets for high intensity workloads like gaming. This means currents could be anywhere from 50-100A at 1V through the core. To supply such current, we need optimized VRMs (voltage regulation modules) usually of 1 or more phases.
Per Phase Design
A single phase in a PWM is basically a circuit which takes in 12V and switches it at such a frequency (usually a 10-1 duty cycle) as to create pulses of 12V. This is then smoothed out through an inductor which essentially averages out the signal into a 1.2V supply. Of course a single phase is not going to have consistent power.
For the switching segment of a VRM phase, generally you need:
- 1 high side mosfet for 12V to output
- 1 (or more) low side mosfets for output to ground
- 1 MOSFET driver which can connect and switch both the high and low side mosfets with a PWM signal input
And for X amount of phases, they all need to connect into a PWM controller capable of driving the amount of phases specified. Generally more phases will be added for more stability in the votlage supply, as well as a large amount of capacitors close to the GPU core to smooth out the current.
«To be continued»
>> Home