Hacker News new | past | comments | ask | show | jobs | submit login

Worth noting:

> By default, this server is turned off. When I want to do some work, I turn the server on using wake-on-lan. I do this with all my equipment except for my router/firewall.

If it is only on when it is doing work, idle power doesn't matter _too_ much.




The Intel ark site says the TDP for that xeon is at 115W, and that's two of them. When you're running at full power (which could happen) you're looking at the CPUs alone drawing over 200W. Another pointer to high power consumption is the 750W PSU. All in all we're not looking at a particularly efficient machine (even if it's unfair comparing it to 2019 hardware).


The PSU's rating is for its potential to supply power, whilst draw components (CPUs, disk, etc.) are rated by use.

The actual efficiency is defined by capacity, load, and miscellaneous other factors, but should be in the neighbourhood of 20% of delivered power.

If the system idles at ~100 watts, the PSU might add another 20-30 to the mix, not 750.

https://www.howtogeek.com/129667/what-exactly-does-the-watta...


Agree with all of that, but I think OP was referring to the fact that higher-rated PSUs will use more power than a lower rated PSU when idle.


Running a PSU at ~50% of its rated max is generally recommendable. At 115W * 2, plus other draws (disk, mobos, blinkenlights), that's around 300W.

Again: peak draw, when on.

I'm not arguing that this system is particularly efficient, only that you don't want to add 750W for the PSU to the draw.

I've not specced out low-power systems myself. I doubt you could get 40 hyperthreads running way below this, though there are definitely some low-power systems which might have a total budget below 50W. Reddit's HomeLabPorn may have some more useful guidance: https://old.reddit.com/r/HomeLabPorn/

(Not my area of expertise, I've never really stayed current in HW. Though I'm aware power/thermal budget has been a major focus, both mobile and server, for the past decade or so.)


The 24 cores/48 threads EPYCs that come with 128mb L3 cache have a TDP that's between 155W and 180W. If you want to go with a lower core count (because you have a more modern architecture), you could even go with the 16c/32t EPYC 7282 (64MB L3) that sits at 120W. All in all, you're looking at almost half the power draw (AMD measures it differently than Intel, so aboutish the same) for the same performance.

I specifically mention the L3 cache size because, while not being substantial, a large cache can get you 10 to 20% improved performance due less CPU stalling from cache misses. For comparison, the Xeon in question has 25MB L3, so we'd be looking at 50MB cache, split in two dies (so it doesn't quite work as a whole block of 50mb cache).

greatjack613 also mentioned the fact that an AMD Zen 3900x matches the multicore performance (and has better single core perf) for a fraction of the TDP, see https://news.ycombinator.com/item?id=20687932.

All in all that was the general idea: yes, you can get a used server for a good price, but we shouldn't forget the efficiency aspect of it, compared to new hardware.


Good points, and again, not something I keep current in, generally, as the field changes so quickly, particularly relative to my own replacement cycles.




Applications are open for YC Winter 2020

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: