Hacker News Reader

Comment by nickpeterson | Hacker News Reader

Just an observation from years building computers. Never cheap out on power supplies. Always buy one (don't use one included in a case), and buy one rated 25-30% above what you think you'll need. Some of the most aggravating troubleshooting you'll ever endure can be avoided entirely.

View on HN

Towards that end does anyone know if there any ATX power supplies that will stream out their sensor info over a serial port or something?

It'd be nice to have power monitoring on that side of the mainboard's power regulators and caps, to pick up marginal issues before they get filtered out, and the power supplies these days will have a little ucontroller with all that information anyway without the mainboard having to duplicate that monitoring circuitry in the right place.

For what's such an important piece of a computer there's so little introspection for data while in use that has to be being collected regardless. Yeah, you can always hook up a oscope or what have you, but catching it in the act of being marginal early on would have saved me more hours in my life than I'd like to admit.

Server PSUs support such communication (looks like this is PMBus protocol, variation of SMBus, which is variation of I²C). In servers it's connected to the BCM and usually allows to read input and output power, FAN speed, temperatures and status (good/bad). The BCM often records historical data, including peak power (in last 1h, 24h), etc.

After quick searching is looks there are people reading this info with a MCU [1]. But I think it would be hard to find such PSU in ATX form factor.

[1] https://github.com/sxjack/dps750tb_psu_i2c

Whenever I experience issues which are likely related to a power supply, it is much easier to drop in a known-good replacement than to spend time diagnosing the specific issues. Also better to remove a potentially faulty PSU before it damages much more expensive hardware.

The typical failure modes mean that it'll be marginal for a while, and should be able to give you information before it goes fully bad. Being able to have introspection before it goes fully bad is what I want, as I hope to have some data before it even makes it's way on to a bench.

The power clients (mainboard, GPUs, disk drives, etc) normally know they're getting pretty bad power from a PSU and filter it heavily to smooth it out. It'd be nice to know if the voltage is occasionally dropping and it's just being covered up by downstream caps that are over provisioned to just smooth out not fully rectified AC.

I had a Corsair power supply with this feature. The system I built with it was quite flaky, but I could reproduce the failure and eventually debugged a 12V transient dropout (when running AVX workloads) that was the cause of the flakiness. The built-in monitoring didn't help at all; doesn't update frequently enough, and the computer crashed by the next USB polling cycle anyway. I used an oscilloscope.

I don't really have enough data to outright blame the power supply. I don't think I connected all of the 12V lines to the motherboard (power supply didn't come with enough cables), and I can't be sure that the motherboard was designed properly for a mildly-overclocked 6950X (VRM banks, capacitors near the CPU, etc.)

It was all enough of a pain that I will never buy another Corsair power supply. My current best practices involve only buying Seasonic power supplies, and connecting every possible "optional" power connector to the motherboard. If the PSU doesn't come with enough cables to do that, buy them separately. My current Threadripper motherboard has a 6-pin and 2 4-pin ATX12V connectors, and a Molex connector. Many are marked "optional" in the motherboard's instruction manual, but I have them connected anyway. Probably overkill. Great stability. Without having designed the system or the test scenario, skipping optional things sounds like the sort of thing that's going to cost you a lot of time at some random point in the future. Best to avoid, even if you look like an idiot to the EE that designed the board.

IIRC, Corsair's AXi and HXi series power supplies -- basically, the really expensive ones! -- support PMBus (I2C).

The entry-level and lower-priced ones don't, AFAIK.

The data isn't necessarily collected. Many existing supplies use hardware counters or PWM modules to control the power output at a level very close to hardware.

In both of those cases there's a feedback loop setting the PWM dynamically. Even the error term of the PID or what have you would be super helpful.

There is no PID loop in the sense you mean. It's a hardware counter module that feeds a flip flop. There might be a free running ADC that is used to compare match against the counter module, or sometimes it's just a binary input like the AC signal zero crossing.

You don't need a CPU core to do this, and the older, cheaper parts do not necessarily have a CPU in them. There are even (iirc, very uncommon) designs with no digital logic save the switching.

Corsair AXi and HXi series. They even got upstream kernel support: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/lin...

Here is my output from lm-sensors

    $ sensors
    corsairpsu-hid-3-1
    Adapter: HID adapter
    v_in:        230.00 V  
    v_out +12v:   12.00 V  
    v_out +5v:     5.00 V  
    v_out +3.3v:   3.00 V  
    psu fan:        0 RPM
    vrm temp:     +49.0°C  
    case temp:    +40.0°C  
    power total:  96.00 W  
    power +12v:   80.00 W  
    power +5v:    16.00 W  
    power +3.3v:   8.00 W  
    curr in:          N/A  
    curr +12v:     6.00 A  
    curr +5v:      3.00 A  
    curr +3.3v:    2.00 A

I learned "what not to do" from my work experience, having to wrangle machines with cheap PSUs, and at home having a "supplied with the case" PSU say "pop" to me years ago. Get something good, and you won't have to get it twice.

But they don't last forever, and my old file server's 10-year-old PSU finally started causing random errors about a couple of months ago - recognized when a sustained hard disk operation dimmed the power LED. I had spares, but it's time for me to think about building something new. Whether it's more computing power for the same electrical input, or the same for less, I can do better.

How do you tell a quality power supply?

The 80plus sticker program doesn't mean a lot, but it does mean a little. Being from a brand that's been around for a long time and is not known for cutting all the corners helps too.

You can’t.

Even quality brands have crappy models.

They are not necessarily the cheapest ones, it is possible that the 400w is fine, the 550w is crap, and the 600w is fine again, even if they are in the same series.

They can also change the specs over time, without any indication on the packaging.

You have to look for reviews where they actually disassemble and check the insides of the psu to determine if it is a good buy.

There is a forum local to my language where they keep track of the small number of truly recommended models, there must be an English equivalent.

Probably the closest thing would be the PSU tier list from the LTT forums: https://linustechtips.com/topic/1116640-psucultists-psu-tier... but even then it's based on aggregated reviews etc.

What sorts of problems are caused by power supply issues?

Random reboots, program crashes, kernel panics/BSODs, boot loops, and wavering displays are some of the things I've seen, aside from the system just not powering up.

Yeah, my old PSU would just shut down whenever I play an old game with vsync off. I’m thinking it couldn’t keep up with sudden power requirement. I recently replaced it with a better PSU and I’m not getting any of these shutdowns again.

Reply on news.ycombinator.com