Comment by paulmd | Hacker News Reader

Comment by paulmd | Hacker News Reader

Parent Full thread

> I'm curious as to why no vendor has yet tried to push a 24V or 48V PSU rail standard to reduce the amperage pushed over these crappy pins.

people on another HN thread yesterday literally were bitching about the cost (not safety - cost) of a $1 adapter being "forced" on them by NVIDIA and now you want to tell them they're gonna have to buy a whole new PSU? that's gonna be quite the comment thread.

but yeah you're right in general, gotta move to 48V to move forward, pushing infinite amps at low voltages is completely bass-ackwards. The pcie-add-in-card form-factor needs a rethink in general, the GPU is now the largest and most power-hungry part of the system by far, and power isn't the only problem imposed by that ancient AT standard. imo we should move to some standardized "module sizes" for GPUs so that we solve GPU sag and cooling at the same time.

the "pcie card" of the future should be a 12x6x4" prismatic module that slides on rails into a socket that provides a PCIe edge connector as well as a XT90 jag (or similar) for power, at a defined physical location, with 48V. Airflow should happen in a defined direction and manner, whether that's axial flow or blower-style.

also please for the love of god, hot air rises, the AT standard means axial coolers are pushing against convection. But we couldn't even get BTX to take off, which is a trivial modification of ATX to reverse the orientation and fix this problem, the DIY market is deeply deeply rooted in the 1980s and will not abandon ATX without some nudging. Enterprise and OEM are fine, their volumes are big enough to do their own thing, but consumers are spinning their wheels fighting IBM's 1980 mistakes.

It's gonna take government intervention, the EU is gonna have to step in and do it like with USB-C cables. The mess of arbitrary card dimensions slowly growing larger over time is the exact same problem that we had pre-standardization with USB too.

At 500W+ maybe the GPU should have it's own dedicated power supply with it's own power cable. And then the GPU manufacturer could optimize the voltages and amps since they control the whole power chain.

You could have some standardization, a simple power negotiation protocol "I can supply 500W at 24V or 48V" and a standard (big) connector.

A long time ago the display was powered from the main unit, but we moved passed that and now displays have their own power cable.

Well, it didn't go that well [last time around](http://www.x86-secret.com/pics/divers/v56k/finalv5-6000.jpg).

> Forbidden

> You don't have permission to access this resource.

Blocked based on referrer, if you go up to your URL bar and hit enter it should load directly

Cool, that fixes it- thanks!

Take it as a reminder to disable the privacy leak (for Firefox set network.http.referer.XOriginPolicy to 1 or 2 [0]) and demand that browser vendors fix their defaults.

[0] https://wiki.mozilla.org/Security/Referrer

> A long time ago the display was powered from the main unit, but we moved passed that and now displays have their own power cable.

That was just switched A/C, so you didn't need to turn off two switches when you were done.

Why? Transformers scale just fine. Two power supplies is just worse in every way.

Welcome to eGPUs

500W is going to be one hell of a wall wart.

Eh not really, you just need to get two of these and connect the sync pins together: https://hdplex.com/hdplex-fanless-250w-gan-aio-atx-psu.html

Way smaller than SFX PSUs and "gaming" laptop power bricks.

That's a cool capability and I didn't read deeply enough when I saw the new GAN charger come out to notice that.

I haven't used it but from the outside it all appears very thoughtfully and competently engineered. His products have some of the highest TDP support (passive cases and his PSUs) on the passive HTPC market and the hifi aesthetic is great. I'd love to get an H5 or an H1 (although for a HTPC it'd be nice to have the H1.SODD - the H1.TODD doesn't look nearly as nice imo), don't really need it but the cat would love it.

Streacom does some interesting stuff too (like the DB4 cube case). Not quite as high TDPs and I don't like their body/fin aesthetic as much, but the DB4 is cool. And Akasa as well, love the galactico design.

Mini-box is the company that originated the picoPSU design and they do a fair bit of other miniature-computing stuff as well, including for the embedded automotive market. They have 18650 UPS modules for NUCs and other motherboards, shows up as a battery in the OS. The M350 case is also pretty nice for an on-the-desk type mini-pc, clearly to hit $50 you have to minimize costs but it's really the nicest $50 case that I've used in a long time (NZXT S340 non-elite got to $60 sometimes). The first picoPSU died within a day for no reason that I can tell but the second one has been fine, with the one caveat that the laptop brick is not low-noise at all. At one point I tried putting my PCs line-outs into an AV switcher and I was getting a groundloop and I could hear a bunch of noise. It was the M350/picopsu, when I unplugged it the other two PCs (including a dell laptop) were fine. Maybe not a typical use-case but it probably isn't the best choice for your SDR host or whatever lol. But HDPlex does have those linear regs...

It's super amusing that the 5700G (get the 5750G on ebay if you want ECC) is roughly as fast as my 9900K in most things but it does it at like 60W in a SFF case, can go even lower without much hit with eco mode. And it also can sorta game (in non-action stuff or older stuff). One night for fun I decided to play Team Fortress 2 at 3440x1440 on it, and it actually got a reasonably playable framerate... with a maxframes-config it would drop to around 30fps in teamfights but hey, returning to my roots playing it on my thinkpad w510/fx880m lol. I might have been able to push it farther with some res reduction but it was doing OK generally actually. It's also (again, roughly, haven't actually benchmarked) around the same speed as my 9900K in H264 encoding. The power scales very little past stock (at least during CPU workloads), just set it for silent and let it run, eco would probably be even better.

170x55x50mm is a honking huge wall wart. Bigger than most laptop power supplies, which makes sense given that its transforming five times as much power as a typical laptop power supply.

Right that's why I compared it to the bricks that come with boat-anchor laptops, not normal ones.

Tho if you threw 2 of these PSUs side by side in a laptop bag, they'd be noticably thinner than any MacBook charger (29mm vs 25mm) even though each one delivers >250% the power.

I think at these sizes and weights, the GPU needs to become a board that you screw onto the case, like a second motherboard. Sagging solved. Now the noisy fan, replace it with a preinstalled water cooling block to be integrated with your overall cooling solution, voila: noise and weird airflow solved as well.

Exactly, that's what I mean by "on rails". Imagine sliding a server PSU into place and it latches as it seats... do that but with a "GPU module". The case itself needs to be part of the support to fix GPU sag, and that's only possible with dimensional and connectical standardization.

Smaller stuff that still reflects the "add-in-card" intention of the original design is fine, you can still have a bunch of PCIe slots for network cards and USB controllers if you want. But the GPU is not really an "add-in-card" anymore and it's problematic in a design sense to keep acting like it is, let alone to have every single product on the market doing something completely different dimensionally.

That's the "pre-USB-micro" era problem that USB faced.

There's no need for water though, that's a whole can of worms.

AGP part deux

I think we should stop having them as separate cards altogether. GPU RAM already can't be socketed anymore and Apple Silicone has shown massive CPU improvements are possible with soldered ram. It hurts my DIY PC Enthusiast Spirit but buying a mainboard with CPU, GPU and RAM all soldered in for massive gains in efficiency, performance and potentially even price could well be worth it. The interconnection between CPU, GPU and their respective RAM could potentially be massively improved by this.

I am aware that modern gaming consoles are basically this but they make lot's of tradeoffs that a Desktop PC doesn't necessarily have to. Single Die GPU-CPU Combination for example. On the PC world you can reasonably pay twice as much and get twice the silicon for twice the performance, a tradeoff Consoles can't make.

and as for the server this could make essentially large 1U GPU racks possible by just having one big board per U with all the gpus in it.

All in one water loops aren't as great as they make them out to be. You still need fans on a radiator, they can be quieter because they're often larger. Pump failures are more common than fan failures and harder to repair. Traveling with a water loop is a nightmare unless you drain it first.

Who's traveling with an ATX tower?

In any case, I've always used the standard water cooling kits and they never failed me (as opposed to constantly breaking fans and huge temperatures). Maybe it's my environment.

In my LAN days, I traveled with a MASSIVE Antec case with water cooling, and a 21" Trinitron.

people with apartments who move

My PC with a GTX 1060 has survived 4 moves.

Speaking of USB-C and crazy design choices, I'm a little surprised someone hasn't just slapped a bunch of those on a graphics card to provide 100W at 20V each. Arguably the parts should be cheap since they are already available at scale, and 6 ports wouldn't be any larger than the problematic nvidia connector. I'm mostly joking, but for portable/mobile GPUs it doesn't seem that insane, that way there is also some choice as far as power adapter bulk.

The inside of consumer-assembled PCs is a terrible place for government intervention.

> ... and now you want to tell them they're gonna have to buy a whole new PSU?

I saw the connector and did a double take, it has a third row of smaller signal pins which none of my existing power supplies have. Not even my newest which is a 750W from last year. Sone new gpu power connector. So I assume most people had to buy a new psu anyway.

I am all for higher voltage. 24V is the standard in industrial automation as well as off highway machinery and european trucks. 48V is common in telco but positive ground though it is common in industry; I use 48v supplies for small stepper and servo motor amplifiers for motion control.

People haven't had to buy new PSUs. The adapter has the circuitry built in to detect how many of the connections are populated. For each one, the 40xx cards will raise their power limits by 150W.

fun fact - that circuity is totally superfluous. Connector itself has 4 small pins underneath main ones specifically for detection. Nvidia did Apple here inventing stupid shit to make it proprietary. All this circuity does is 4 to 2 binary encoding.

It's not proprietary, the connector is called 12VHPWR and it's specified by the PCI SIG.

Not the connector, additional useless chip Nvidia put in it encoding 4 signals into two pins, and leaving 2 remaining pins unpopulated.

Usefull to have unused pins, if you have to update it to support 1000w in the future

Sometimes its good to stop and think a second longer. This "600W" connector is already melting, and you want to pump 1000W thru it?

The reason BTX failed is that it didn't present compelling arguments for enthusiasts at the time.

Were there high-end enthusiast-class BTX boards, comparable to the best ATX boards of the time? I don't recall seeing them in any numbers. Hell, even BTX cases from recognizable manufacturers are quite rare. It's obvious they were making the pitch to the largest OEMs, the Dells and Acers who can source their own cases and boards.

And the sales pitch was most compelling for them. If you've got a roasting Pentium 4 and a modest graphics card, you can fit them in the BTX "cooling channel" and keep them at a frosty 95c with a dinky heatsink and one or two cheap fans. That's not really exciting to an audience already saying "I can spend $75 on a kilogram of heatsink to get my CPU down to 70c, or $200+ on a custom-loop watercooling setup and get to 45c".

Maybe they needed to find a niche market for hobbyists to work with-- for example, the early "silent computing" crowd might have been more interested in marginal increases in cooling. After all, people are willing to consider other form factors when there's some compelling draw (see Mini-ITX)

I also wonder if the original design thesis still holds up. Once the Core series came out, "how do we cool this angry CPU" became less of a fixation, so it was probably easier to switch back towards commodity ATX designs. On the other hand, we've shot back past the original thermal situation and then some-- would the design still hold up with an unlimited 7950X drawing ~250W and a 4090 drawing 400w all trying to exhaust it into the cooling channel, or would they bake each other to death?

I do agree that today's GPUs are ridiculously long, and nonstandard designs make them difficult to manage. Given that most outward-facing drives are gone, and SSDs have replaced most spinning rust in enthusiast PCs, I wish we'd see more "flat desktop" style cases. The GPU isn't sagging because its weight is directed mostly straight down into the case's bracing, convection is probably a little better, but it doesn't really help the power nightmare. They always make the cheap photoshopped jokes about GPUs with mains power, but maybe that's a stopgap solution-- sell the 5090 with an external power brick supplying, say, 48V/12A, and there's no internal cable gore and the power system can be 48v everywhere it matters. Then there's no need to design the rest of the PSU to guess what nVidia will want tomorrow.

The workstation world sort of broadly standardises card length, allowing for support on 3 sides (PCI Bulkhead, PCI-E Slot, and a card extension with a blade to fit in a rail guide towards the front of the chassis). Even old cheese grater Apple Mac Pro's had this setup.

Reply on news.ycombinator.com