GPJuice - Experimental device to improve PSU transient response

msystems

King of Cable Management
Original poster
Apr 28, 2017
781
1,366
A little while back I saw the GN video regarding "the GPU transient problem" and came across people mentioning their experiences with this problem. Like (flex) PSUs tripping overcurrent protection when loading a benchmark, even though the PSU is rated above the sum of the tdp of their components.

So that got me thinking about adding capacitance to the 12v output. Just a bunch of caps really. I found a few examples of this:
1) Evga Powerlink integrated a few caps into their device without much detail on the results. The marketing just seems to mention this as "power filtering"

2) A user on Hackaday added a lot of caps in parallel to their 12v rail and reported positive results. Nice! But looks really crazy/messy.

3) RC Racing enthusiasts seem to be crazy about their capacitor banks. The application is different but it is basically working to stabilize the load of their RC car engine which is a PWM electric motor. My understanding is the battery (very small LIPOs) can't supply the brief instantaneous current demanded by these motors, so the cap bank works to smooth out the voltage drop.


Design I came up with:

-very small SFF friendly design
-connected by daisy chain or Y-splitter to GPU
-3d housing to immobilize the solder joints in a fixed position relative to the connector

Parts - $30
-12 x 470uf Aluminum caps (5,640uf) P/N: 25SEK470M Panasonic
-1 x 12v 30ma Led (power on indicator / bleeder)
-30 x 70mm proto board
-Molex minifit jr 8-pin housing, female
-2 x m3 screw
-3d housing



The plan - for simplicity only use 4 wires of 14ga copper. Enough to thicken the traces, but not too absurd so could still crimp molex terminals on it.



Assembled result



Since "daisy chain" cables are common, I tried this approach for simplicity instead of trying to build two connectors into the device for it to be "inline". I didn't notice a difference whether it was positioned before or after the gpu on the "daisy chain".



Unfortunately I don't have much conclusive evidence about the results of the experiment at this point. I don't have the electrical tools to measure the performance in a scientific way, nor a setup that can even produce the right kind of transients that this thing could possibly mitigate, so not much results to report.

When testing with a HDPlex 250w GAN, it already has robust power delivery and can handle the 3060ti at 100% power limit so there was no observable difference whether using GPjuice or not. Triggering OCP was only caused when running over spec anyway (so, OCP was doing it's job).

I hope to eventually get some quantitative result of how it might affect the performance. Its just a research project at this point. If anyone has an oscilloscope setup like in the GN video, perhaps they would be interested in scientifically testing it.
 
Last edited: