All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Teach me physics [colo pre-considerations about power consumption]
Ok, I need some help/ideas from anyone that is either hosting servers or using colo.
How do you pre-plan power consumption and what are your real world examples of power consumptions of rack servers?
E.g. I'm considering a system with
- a dual xeon (scalable or E5v4) with about a TDP of 150 each.
~ 8 * 64GB ECC RAM (is it correct that is better to have rather less amount of modules, but bigger RAM modules in terms of total power consumption?)
two to four NVME disks (U.2 or U.3 or M.2?)
four to eight HDDs
Now the load should be in the range of <10% normally with regular spikes into the 20-40% realm and than still be able to do some bad peaks into high loads.
So I think this is a build for a ~1kW PSU, redundancy would be better. I stopped a bit when I saw that most (refurbished) server providers only offer ~80% efficiency on those PSUs as imho this would be way too much lost power...
So I still have no clue about power consumptions of NVMEs, but would consider less than e.g. HDDs.
Eight HDDs should be about 100w.
8 RAM modules about 50 to 100W?
So can I translate TDP into power usage at the CPU level? Meaning 300w for those CPUs under load with like 100W for HDDs and 50 to 100W for RAM? With the power loss at the PSU this would be a total of 600W, not regarding the fans, etc.?
Or are all of these considerations way off compared to reality?
Comments
TDP, especially intel's aren't really true. Better add at least 25% to be safe.
I think that's not significant. What gobbles up power is refreshing or writing the memory which, seen from an electronics perspective, means recharging (tiny but lots and lots of them) capacitors. So, x bytes of DRAM require (nearly) the same amount of power, no matter whether distributed over 2 or say, 8 modules.
mostly depends on size.
Quite diverse. Generally 2.5" disks consume significantly less power than 3.5".
Reason?
Indeed. But I guess that's at least to a significant degree due to sellers assumption that PSUs somehow loose efficiency over the years. From an electronics perspective though I strongly doubt that - but have bad news: Some elements, in particularly electrolytic capacitors do have a limited lifetime and do loose capacity over the years which also translates to "lower PSU efficiency", but that's not the really bad part of the story, which is that those caps tend to die and in an ugly way (ruining the whole PSU unless looked at, taking care of, (properly!) cleaning PSU which should be done by professionals).
Unless you happen to have a decent electronics lab and (quite expensive) test equipment allowing you to properly test 2nd hand PSUs, my advice is to simply buy sufficient spares ...
P.S. At the very least have (or get) a "kill-a-watt" (cheap and simple power guesstimator) and carefully look at it. Although more expensive I guess, even better try to get a power guesstimator which also stores and shows the max. value over some time.
Yep, but details depend on technology (e.g. M.x vs U.x) and of course size (and more).
Nope. (a) don't guess but look it up! and (b) they tend to have nasty spikes, so generously add min. 25% to manufacturers specs (unless they explicitely provide 'max'/'peak' numbers). Also see above.
To quite some degree depends on what you aim for. If reliability then go for sufficient or even plenty reserves. And of bloody course always have a hot spare PSU (meaning one that's not in some storage but physically located in the server)!
IMO, the CPUs, HDDs, and fans will generate the major load.
NVMe and RAM should consume way less, but probably still worth considering just to be on the safe side.
Hence your calculations are generally correct, especially if you have checked the datasheets (definitely exist e.g. for HDDs/SSDs, let alone CPUs).
So make sure that the PSU(-s) you're getting is of high quality and actually delivers the declared wattage.
I think your RAM watts is way high..
But capacity can/will make a difference too.
400-500w for that server.
According to our experience, you can expect average consumption of 350-450W average depends on choosen system and CPU load. But it will consume up to 600W during start or peak load, so you have to consider it if you have limits in rack's socket.
Thank you for all your insights, this really helps.
Now if I turn to a concrete example like e.g. this configuration done at a refurbished dealer here:
https://www.etb-tech.com/supermicro-superstorage-ssg-6019p-acr12l-configure-to-order-cse-802-chassis-with-x11ddw-nt-motherboard-bto6019p-acr12l.html?cyoShareId=5559
(link is only valid for 30 days, so until Nov 12th 2024).
This is a 1U example with dual xeon platinum at 205TDP, quite some RAM, four NVMEs and eight HDDs.
Why does it only offer (redundant) 600W PSUs, but it does allow e.g. dual xeon 82xx cpus with 205TDP each? That does not make any sense, or am I wrong? This would be severely under-powered imho?
The mobo x11ddw-nt claims to support up to 205TDP (no headroom?) but than why only 600w PSU?