Hacker News Reader

Comment by rx_tx | Hacker News Reader

For context, the RTX4090's power budget is 450W, and if all 4 plugs are connected to the cable adapter, the system can allow up 600W (at 12V) to the GPU. That's 6A to 8A per pin (there's 6 pins for current, 6 for GND plus 4 signaling pins), so pretty sensitive to resistive heating in case of a bad contact.

View on HN

There isn't an issue with resistive heating. The safety margin is down a few factors from the 8-pin PCIe adapter, but there is still a healthy safety factor.

https://twitter.com/hms1193/status/1585257428291325958

This isn't like running kA of current where the threat of thermal runaway is always present, even with active cooling. I worked on a stellarator that had a coil feed blow (years before I started). Pushing even 10% more current meant a complete overhaul of the feeds. That's what thin margins looks like.

It’s only a healthy safety factor if absolutely nothing goes wrong. Connectors are normally a weak point and internal PC components aren’t supposed to be plugged and unplugged regularly. Simply testing the multiple hardware configurations can quickly result in failures.

More interestingly, modern ASICs run their logic at something like 1V (I don't know the exact voltage of this particular one of course), so the current going into the actual chip (after voltage downconversion on the board) is hundreds of amperes.

AFAIK the top-end for modern Ryzen chips is in the 1.5-1.6V range, usual operating levels are lower. My 3xxx Threadripper is at 1.3V right now. In comparison, my 3xxx-series RTX card is at 1.06V (but only operating at 40% of max power). The EVGA software only allows me to set a +100mV voltage offset, but the default voltage curve looks like it peaks at 1.100V for ~1950mhz.

Who's idea to feed monstrous 450W GPU with power using this small choke point?

Seems like GPUs need more voltage. Something like 120V would be more appropriate.

Yeah, sending 48V would be more appropriate, but unfortunately that's not really backwards-compatible (well, the way to do that is typically to allow the PSU to be farther from the use-site, like having a datacenter with tall adult high cabinets (racks; they have 4 posts for holding the equipment, which is standardized as occupying one-or-more slots of 1¾ inch height and free to occupy the entire length) and putting 4 AC/DC converter blocks into each such cabinet, to power a few layers of servers/equipment both above and below the PSU layer.

They do this so that the PSU can be made up of multiple redundant modules, and to integrate some batteries that can cover a generator starting up. This way they don't need to pay for a dedicated UPS battery with chargers and inverters just to convert it back to lower-voltage DC near the server.

They can make it so that for example 5-8 PSUs spread over different electrical circuits can share the load of the servers, even if 2-3 of them fail (so they, for example, only need 2 extra on top of the 5 that can power the servers and the 1 that recharges the battery (62.5% load if everything works and the servers are running full-power), or even relying on the servers occasionally running below full power to spread the full rated server power over 6 modules with 2 spares (75% load if everything works and the servers are running full-power)).

A big issue in this situation here is though that Nvidia skimps heavily on board space for the 4090 Founders Edition cards, seemingly to use the back for heat sink fins/air-flow cross-section. It would take up some board area, but in theory, they could add a 48V power option that makes it all sleek and pretty, possibly with a 12V->48V adapter for the people who don't already have a 48V supply.

Maybe it's own power supply? Although this would require electric decoupling from the PC because of switching power supplies. I see here an opportunity for Jensen and his marketing people to come up with a fancy name for it.

FSP did this with the Booster X5: https://youtu.be/CF00ez4JNEU

You plug both PSUs in, and DC output appearing on the primary PSU switches this auxiliary one on and off.

Isn’t that a E-GPU basically?

Countries like Australia/China 240v is the standard. You need step-down transformers for 110v devices. China has both voltages/plugs running in its cities - not sure how they manage that.

Most of the world is on 220-240V: https://theverbalized.com/why-do-we-have-two-standards-of-el...

Also the power supply to a GPU is DC anyway, so the wall voltage probably doesn't matter much.

A common misconception about the US is that the US is only 110V. 220V service in homes in America is standard, it's just that standard power plugs around the house get 110V because of split phase breaker boxes. This lets 220V be available for large appliances like Washers, Dryers etc using the full phases, and 110-120V to the rest of the house using only half phase.

In my next house, I think I'll request a 220V outlet in my kitchen with a UK style plug, so I can have my fast tea kettle.

Doesn't seem like a misconception. It seems like y'all have 110v power outlets. Having 220-240v somewhere else is kind of moot... The power socket IS the interface, so if the interface is 110v, then the switcher box, the power line, the high voltage power line, the power plants voltage aren't really relevant.

How can that be true? I plug a kettle cable from my a/c socket to the GPU... is there a rectifier involved in a standard PC setup?

You plug the AC cable into the PC's power supply, right? That converts to 5V and 12V DC and feeds it to the GPU internally. There's no way to plug external power directly into the GPU.

The United States is also a 240V country. It’s just split phase. But yes, mains at the wall would be 120V.

Often, anything above 48 or 50 volts is considered high voltage and in some places might be subject to regulation. Not to mention just being more dangerous to handle.

Houses should come with a separate "Gaming PC" circuit breaker and power plug just like your dryer or oven. /s

The problem is not external, 500 or 1000 W is not that much in the grand scheme of things (a euro plug allows up to 625W, unearthed).

The issue is internal: because the top PSU rail is only 12V, high-power devices (mostly GPU) need to push very high amperage totals. And because PC internal layouts, they can’t do that through large wires with huge connectors.

Also a legacy of cheap friction fit.

Exactly. This is why high wattage USB-C cables rely on the power supply negotiating a higher voltage in order to work. Sending 100W over 5V would be ridiculous (20A). Instead the device tells the power supply to switch to 20V and then the cable only needs to be rated for 5A. Something like that for GPUs would make sense.

It is KINDA external. 1000+w is starting to get into “trip the breaker” pretty easily if you have anything else on that circuit with any kind of transient load - like say a minifridge.

> 1000+w is starting to get into “trip the breaker” pretty easily

I was wondering where you'd get that from, and then I remembered that you're probably american so your baseline is only 1800W, in which case... yeah that's true. If you're installing a really powerful PC in the US you probably want either a dedicated circuit or a 20A circuit (or both).

> I was wondering where you'd get that from, and then I remembered that you're probably american so your baseline is only 1800W, in which case..

In North America, you cant pull 1800W continuous either, it is limited to 1440 by regulation.

"This regulation is called the maximum continuous load, which is 80% of the total wattage computed. "

This is why space heaters and such are 1200 watt.

So while 120v * 15A is 1,800W no electrician would wire a setup like this.

if you need 1800W they would use probably use a 20A circuit and wiring rated for 20A. 80% load of this would be 1920W.

NOTE: Not an electrician.. if i am mistaken please let me know.

I mean that's half the power of a decent kettle...

A gaming PC uses maybe 20-30% of power of an average kitchen appliance like a kettle or a blender. Those are plugged into standard wall sockets just fine.

Depends on how many GPUs are in the PC. Circa 2011 I had quad-Radeon 6970 machines pulling ~1000w each from 1200 watt power supplies, measured with a kill-a-watt.

Those machines ran 24/7 for at least 36 months before I had to start servicing stuff like GPU fans and heat sinks.

This stuff can be designed NOT to melt the dang 12v rail

So you went up to cca 40-50% power use of a single kettle. Whoop.

Depends on how many GPUs are in the PC. Circa 2011 I had quad-Radeon 6970 machines pulling ~1000w each from 1200 watt power supplies, measured with a kill-a-watt.

Those machines ran 24/7 for at least 36 months before I had to start servicing stuff like GPU fans and heat sinks.

This stuff can be designed NOT to melt the dang 6 pin 12v rail

Yeah, a kettle is gonna pull like 2200 W when it's switched on. I'll be impressed if a gaming PC can pull half of that during a gaming load.

>Yeah, a kettle is gonna pull like 2200 W when it's switched on

...in areas that use 220/240V. In north america appliances are limited to around 1500W, because the mains voltage is 120V.

Might depend on where you are and how seriously they take tea. Our very ordinary British kettle is 3kW.

Ffs don't give them ideas.

Ha ha only serious. There were cases on local news when homes burned down because PC caught fire. Not only circuit breaker — please give me something (IR sensor?) that will trip and cut power if anything in the room and in the PC enclosure exceeds, say, 120°C

Or a simpler setup would be a smoke detector that turns off a power strip when triggered.

Kind of like some 3d printers have.

Not sure how useful that would be considering the fire has probably already started.

Things generally smoke before they burn

Your comment reminds me of this meme:

https://static01.nyt.com/images/2016/08/05/us/05onfire1_xp/0...

Reply on news.ycombinator.com