[pja]

> I think next gen we will begin seeing 24V power supplies to deal with this.

May as well go the whole hog & jump to 48V.

(50V is as high as you can go whilst still being inside the “low voltage” electrical safety regime in most countries IIRC.)

[eqvinox]

General SELV limit is 60V, that's why PoE is 54≈56V at the source (it's calculated at roughly 10% tolerance so it can be built cheaply.)

[Schiendelman]

Then the graphics card would have to have a transformer on it to step down to the voltage that the chips can handle.

[dcan]

They already do - most of the components buck the 12V down to the 1.3ish volts that the GPU core needs

[xxs]

They are not transformers, though. The coil/chokes are not galvanically isolated which makes them (more) efficient. Stepping down from 48V to 0.8V (with massive transient spikes) is generally way harder than doing it from 12V. So they may ended with multi step converters but that would mean more PC with more passives.

[eqvinox]

3.3V from 48V is a standard application for PoE. (12V intermediate is more common though.) The duty cycle does get a bit extreme. But yes, most step-down controllers can't cover both an 0.8V output voltage and 48-60V input voltage. (TI Webench gives me one - and only one - suggested circuit, using an LM5185. With an atrocious efficency estimate.)

You'd probably use an intermediate 12V rail especially since that means you just reuse the existing 0.8V regulator designs.

[xxs]

Aside the step down, the transients can be quite crazy, which might make the power consumption higher (due to load line) calibration. 48V fets would have much worse RDSon compared to lower voltage spec'd ones. So it does make sense no single smart power stage to have such transistors (presently).

There are other issues, too. 48V would fry the GPU for sure, 12V often time does not even with a single power stage failure.

In the end we are talking about a stupid design (seriously 6 conductors in parallel, no balancing, no positive preload, lag connectors, no crimping, no solder) and the attempted fix is a much more sophisticated PCB design and passives.

[Schiendelman]

So then it would need to be significantly larger.

[lazide]

Likely smaller actually.

[SV_BubbleTime]

This isn’t how it works.

Your SMPS needs sub-2V output, cool. That means it only needs to accept small portions of the incoming.

But, if the incoming is 48V, it needs 48V tolerant parts. All your caps, inductor (optional typically), diodes, the SMPS itself.

Maybe there isn’t a sides difference in a 0603 50V capacitor and 10V 0603 capacitor, but there is a cost difference. And it definitely doesn’t get smaller just because.

Your traces at 48V likely need more space/separation or routing considerations that they would at 24V, but this should be a quickly resolved problem at your SMPS is likely right next to your connector.

[lazide]

Yes. And it also doesn’t need to handle 40+ AMPs on input, with associated large bus bars, large input wires, etc.

Extra insulation is likely only a mm or two, those other components are big and heavy, and have to be.

It’s the same reason inverters have been moving away from 12v to 48v. Larger currents require physically larger and heavier parts in a difficult to manage way. Larger voltages don’t start being problematic until either > 48v or >1000v (depending on the power band).

[deelowe]

No one uses transformers anymore. VRRs are basically mini PCs now. They run firmware and report telemetry and are crazy efficient.

[Schiendelman]

I'm not familiar with this! I've tried to investigate but I just get variable refresh rate. Tell me more?

[deelowe]

Voltage regulators. Voltage regulation technology is extremely advanced as even very small efficiency gains can save billions for hyperscalers. Unfortunately, I don't know of any specific products to share as power isn't my domain. I'm only familiar with the space because we sometimes have to pull telemetry directly from the VRs when doing system level RCAs. Some of our BMCs can do this directly via I2C.

[Schiendelman]

No that's okay, thank you for the pointer in the right direction!

It doesn't look to me like anything out there can take voltage from 48v to 2-3v; at least not obviously.

[deelowe]

https://www.digikey.com/en/products/detail/texas-instruments...

There should be plenty. 48-54 VDC is the standard for OCP powershelf designs. Hyperscalers such as Google have been working for nearly two decades now to eliminate voltage conversion steps. When I left, the power plane within the server PCB ran at the busbar voltage, which could float up to 54VDC. Given this, I'd expect them to convert from 48-54VDC down to 3.3 directly or at most something like 5VDC and then use smaller VRs near components such as ram and cpu.