6 ms·

There is a german forum thread with a google docs document listing different configurations below 30W[1]. Since there are very different requirements, this migh

There is a german forum thread with a google docs document listing different configurations below 30W[1]. Since there are very different requirements, this might be interesting for many homeserver / NAS builders. For me personally I found my ideal price-performance config to be the following hardware: Board: Fujitsu D3417-B2 CPU: Intel Xeon 1225 V5 (better the also compatible 1275v6, but its way more expensive) RAM: 64GB ECC RAM (4x16GB) SSD: WD SN850x 2TB (consumer SSD) Case: Fractal Design Define Mini C Cooling: Big block no name, passively cooled by case fan Power: Pico PSU 120W + 120W Leicke power supply Remote Administration via Intel AMT + MeshCommander using a DP Dummy Plug I bought this config used VERY CHEAP and I am running Proxmox - it draws 9.3W idle (without HDDs). There are 6 SATA ports and a PCIe port, if anyone would like to add more space or passthrough a dedicated GPU. It may be hard to get, but I paid €380,00 in total. Does not work very well for Media Encoding, here you should go for a Core i3 8100 or above. Alternatively you could go for the following changes, but these might be even harder to get for a reasonable price: Boards: GIGABYTE C246N-WU2 (ITX), Gigabyte C246-WU4 (mATX), Fujitsu D3517-B (mATX), Fujitsu D3644 (mATX) Power: Corsair RM550x (2021 Version) Cheap used Workstations that are good servers are Dell T30 or Fujitsu Celsius W550. The Fujitsu ones have D3417(-A!) boards (not -B) having proprietary power supplies with 16 power pins (no 24pin ATX but 16pin). There are Adapters on Aliexpress for 24PIN to 16pin (Bojiadafast), but this is a bit risky - I'm validating that atm. Ryzen possibilities are pretty rare, but there are reports that the AMD Ryzen 5 PRO 4650G with a Asus PRIME B550M-A Board is drawing about 16W Idle. Hope I could help :-) [1]: https://goo.gl/z8nt3A https://goo.gl/z8nt3A

manmal 3y ago
For anybody reading this - I think it’s a great config, but would be careful around pico PSUs in case you want to run a bunch of good old spinning disks. HDDs have a sharp power peak when spinning up, and if you have a couple of them in a RAID, they might do so synchronously, potentially exceeding the envelope.
- LargoLasskhyfv 3y ago
  In times long gone there's been a setting. Staggered spin-up it was called. Turned on one disk after another. So the PSU didn't bother, or stalled. Done.
- agilob 3y ago
  to go deeper, depending on a file system, some FS won't let HDDs go to sleep, so they always consume power for max RPM.
  - scns 3y ago
    > some FS won't let HDDs go to sleep Which ones?
    - agilob 3y ago
      btrfs and zfs at least, probably most frequently used FSs for RAID
      - tomatocracy 3y ago
        You can do it with ZFS at least. I have a ZFS setup which idles drives. If I don't access the datasets on the drives, they don't spin up. You do have to be careful with things like auto snapshotting tools though - that will typically wake drives up. I had to script mine to only run on relevant datasets when drives are spun up and also had to edit the tool (sanoid) to prevent it from enumerating the list of snapshots on the idle datasets not being snapshotted.
      - PBondurant 3y ago
        I'm running a ZFS NAS with mirrored 2 HDD vdevs. I've set the HDDs to spindown after 20mins using hdparm -S and this works fine under zfs. The drives (Toshiba N300s) take ~15s to spin up when something accesses them, and ime zfs has always handled this gracefully. This vdev only contains data; it's not running the OS / root.
- sandreas 3y ago
  You're right. 6 spinning disks and a discrete GPU won't work with 120W. I would recommend to get the Corsair RM550x 2021 (or similar) when using a NAS with more than 2 disks or a discrete GPU. I personally don't need more than 10TB space, so 2 10TB Seagate EXOS in ZFS mirror would just work fine with 120W as long as you don't run Prime95 the whole time. You might go up to Pico PSU 150 then.
- tomatocracy 3y ago
  Even if you have no issues with spinup, insufficient power can result in silent data corruption, which can be quite frustrating to diagnose (even if ZFS will help you see it). The difference between a picoPSU and a decent "real" PSU in terms of power consumption at the wall is maybe a couple of watts in my experience - not worth the risk for most people.
  - sandreas 3y ago
    Is there a source for this?
    - tomatocracy 3y ago
      Not sure. I experienced this personally using a picoPSU with spinning drives on one machine - no issues with spinup but silent data corruption on writes (which then got picked up by ZFS). Switching to a traditional PSU cured the problem.
      - sandreas 3y ago
        Ok, I have running multiple machines with PicoPSU and never experienced this. In my opinion it is a pretty reliable consumer device. However, thank you for pointing that out. I did not know that.
    - justsomehnguy 3y ago
      Can't say anything about picoPSU, except writing anything on RAID involves all the drives to seek and write and without a a solid power reserve you can receive a momentary dip of voltage which can lead for a malformed write, but back in mid-late 2000s I've seen enough HDDs failing and corrupting data which miracluosly behaved when moved to another machine or when you replaced PSU (with a decent one, not another $15 chinesium). A lot of bad reputation of WD from that time is actually from the shitty PSUs, not the drives themselves.
      - sandreas 3y ago
        Well, that's really interesting. If I ever experience these issues somewhere, I know where to start...
ksjskskskkk 3y ago
b550m with a amd5 pro from 2023 (will double check models and post on that forum) i get 9w idle and amd pro cpus have ECC support which is a requirement for me on any real computer. i disable most components on the board. it's bottom tier consumer quality. best part is when i need to burn many more watts the integrated gpu is pretty decent
- bb88 3y ago
  This is the setup I was looking for, the ECC support is really a requirement for a NAS.
- Jedd 3y ago
  The same way backups are most rigorously performed by people who've lost data, ECC is a non-negotiable requirement for people who've suffered slow data corruption via silent memory failures. It surprises me that people are happy with 64GB+ builds of non-ECC, especially for NAS (ie. very long term storage, where corruptions probably wouldn't be noticed for years). Periodically I look at replacing my small fleet of HP micro Gen8's, which use Xeons, have 4 x 3.5" bays (with proper h/w RAID1), but max out at a frustratingly low 16GB. They're quite robust, but because of their age - and HP - a horse veterinarian approach to component failures is usually indicated. A Ryzen + ECC whitebox build is massively appealing, but almost everyone's build-out includes caveats like 'check the datasheet of the mobo and CPU' (because series aren't consistent), and about half the time a terrifying disclaimer to the effect of 'ECC is present / enabled in the kernel, but I can't tell if it's actually functioning properly'.
  - sandreas 3y ago
    This. Additionally, there is the price point of modern AMD "workstationish" systems vs. used Intel "real server" systems. Efficiency together with more performance is great, but if I had to pay double the price for a system that offers lots of performance I won't need anyway I'm not willing to pay the price. Used Intel Systems are just way cheaper, because nobody seems to want them anymore... Everyone wants a >= 8 core ryzen :-) I personally don't need this for a little Proxmox / NAS kind of system. However, I'm hoping for frame.work to announce official ECC support[1] on their Notebook boards (pretty likely this will never happen). I would love to just use the coolermaster case with small efficient modern notebook hardware for 600 bucks just because I don't need additional harddisks and I would just buy one to support the company. [1]: https://community.frame.work/t/ecc-error-correcting-code-defaultly-supported-already-ryzen-amd/1716 https://community.frame.work/t/ecc-error-correcting-code-def...
kogepathic 3y ago
> The Fujitsu ones have D3417(-A!) boards (not -B) having proprietary power supplies with 16 power pins (no 24pin ATX but 16pin). There are Adapters on Aliexpress for 24PIN to 16pin (Bojiadafast), but this is a bit risky - I'm validating that atm. They work just fine. The pinout is well known [1]. You can also adapt a normal ATX PSU if you boost 5VSB to 11V. Fujitsu boards are great, and very inexpensive to purchase in the EU. Someone has even reverse engineered the license for the KVM features of their remote management (iRMC S4/S5) [2] [1] https://web.archive.org/web/20200923042644/https://sector.biz.ua/docs/fujitsu_16pin_to_atx_24pin_power_adapter/fujitsu_16pin_to_atx_24pin_power_adapter.phtml https://web.archive.org/web/20200923042644/https://sector.bi... [2] https://watchmysys.com/blog/2023/01/fujitsu-irmc-s4-license/ https://watchmysys.com/blog/2023/01/fujitsu-irmc-s4-license/
- sandreas 3y ago
  Oh this is pretty interesting, thank you very much. You mean that the Bojiadafast adapters work fine? If so, I wonder if there is a hit in efficiency because of the required step-up / step-down converters in that adapter. However, on my board there seems to be an unsoldered 24pin connector, that could be just used as is with a little soldering, but since it is on-hold my replacement system if my ...-B variant dies, I'm not willing to risk too many experiments :-)
  - kogepathic 3y ago
    > You mean that the Bojiadafast adapters work fine? I have no experience with this particular vendor/brand but the adapters are not complex. If you're concerned, verify that it has 11VSB before you plug it into the motherboard. Otherwise it's entirely passive (passing through PS_ON#, PWR_OK, and 12V) > If so, I wonder if there is a hit in efficiency because of the required step-up / step-down converters in that adapter. It should be quite nominal. The boost converters have a very high efficiency (>85%) and standby power is typically below 10W max.
    - sandreas 3y ago
      Cool thank you... I also thought about soldering a 2$ female 24pin plug to the spot where it is on the (-B) variant... seems that they just left the plug out but the connections are there. We'll see :-)

manmal 3y ago

For anybody reading this - I think it’s a great config, but would be careful around pico PSUs in case you want to run a bunch of good old spinning disks. HDDs have a sharp power peak when spinning up, and if you have a couple of them in a RAID, they might do so synchronously, potentially exceeding the envelope.

LargoLasskhyfv 3y ago
In times long gone there's been a setting. Staggered spin-up it was called. Turned on one disk after another. So the PSU didn't bother, or stalled. Done.
agilob 3y ago
to go deeper, depending on a file system, some FS won't let HDDs go to sleep, so they always consume power for max RPM.
- scns 3y ago
  > some FS won't let HDDs go to sleep Which ones?
  - agilob 3y ago
    btrfs and zfs at least, probably most frequently used FSs for RAID
    - tomatocracy 3y ago
      You can do it with ZFS at least. I have a ZFS setup which idles drives. If I don't access the datasets on the drives, they don't spin up. You do have to be careful with things like auto snapshotting tools though - that will typically wake drives up. I had to script mine to only run on relevant datasets when drives are spun up and also had to edit the tool (sanoid) to prevent it from enumerating the list of snapshots on the idle datasets not being snapshotted.
    - PBondurant 3y ago
      I'm running a ZFS NAS with mirrored 2 HDD vdevs. I've set the HDDs to spindown after 20mins using hdparm -S and this works fine under zfs. The drives (Toshiba N300s) take ~15s to spin up when something accesses them, and ime zfs has always handled this gracefully. This vdev only contains data; it's not running the OS / root.
sandreas 3y ago
You're right. 6 spinning disks and a discrete GPU won't work with 120W. I would recommend to get the Corsair RM550x 2021 (or similar) when using a NAS with more than 2 disks or a discrete GPU. I personally don't need more than 10TB space, so 2 10TB Seagate EXOS in ZFS mirror would just work fine with 120W as long as you don't run Prime95 the whole time. You might go up to Pico PSU 150 then.
tomatocracy 3y ago
Even if you have no issues with spinup, insufficient power can result in silent data corruption, which can be quite frustrating to diagnose (even if ZFS will help you see it). The difference between a picoPSU and a decent "real" PSU in terms of power consumption at the wall is maybe a couple of watts in my experience - not worth the risk for most people.
- sandreas 3y ago
  Is there a source for this?
  - tomatocracy 3y ago
    Not sure. I experienced this personally using a picoPSU with spinning drives on one machine - no issues with spinup but silent data corruption on writes (which then got picked up by ZFS). Switching to a traditional PSU cured the problem.
    - sandreas 3y ago
      Ok, I have running multiple machines with PicoPSU and never experienced this. In my opinion it is a pretty reliable consumer device. However, thank you for pointing that out. I did not know that.
  - justsomehnguy 3y ago
    Can't say anything about picoPSU, except writing anything on RAID involves all the drives to seek and write and without a a solid power reserve you can receive a momentary dip of voltage which can lead for a malformed write, but back in mid-late 2000s I've seen enough HDDs failing and corrupting data which miracluosly behaved when moved to another machine or when you replaced PSU (with a decent one, not another $15 chinesium). A lot of bad reputation of WD from that time is actually from the shitty PSUs, not the drives themselves.
    - sandreas 3y ago
      Well, that's really interesting. If I ever experience these issues somewhere, I know where to start...

ksjskskskkk 3y ago

b550m with a amd5 pro from 2023 (will double check models and post on that forum) i get 9w idle and amd pro cpus have ECC support which is a requirement for me on any real computer. i disable most components on the board. it's bottom tier consumer quality. best part is when i need to burn many more watts the integrated gpu is pretty decent

bb88 3y ago
This is the setup I was looking for, the ECC support is really a requirement for a NAS.
Jedd 3y ago
The same way backups are most rigorously performed by people who've lost data, ECC is a non-negotiable requirement for people who've suffered slow data corruption via silent memory failures. It surprises me that people are happy with 64GB+ builds of non-ECC, especially for NAS (ie. very long term storage, where corruptions probably wouldn't be noticed for years). Periodically I look at replacing my small fleet of HP micro Gen8's, which use Xeons, have 4 x 3.5" bays (with proper h/w RAID1), but max out at a frustratingly low 16GB. They're quite robust, but because of their age - and HP - a horse veterinarian approach to component failures is usually indicated. A Ryzen + ECC whitebox build is massively appealing, but almost everyone's build-out includes caveats like 'check the datasheet of the mobo and CPU' (because series aren't consistent), and about half the time a terrifying disclaimer to the effect of 'ECC is present / enabled in the kernel, but I can't tell if it's actually functioning properly'.
- sandreas 3y ago
  This. Additionally, there is the price point of modern AMD "workstationish" systems vs. used Intel "real server" systems. Efficiency together with more performance is great, but if I had to pay double the price for a system that offers lots of performance I won't need anyway I'm not willing to pay the price. Used Intel Systems are just way cheaper, because nobody seems to want them anymore... Everyone wants a >= 8 core ryzen :-) I personally don't need this for a little Proxmox / NAS kind of system. However, I'm hoping for frame.work to announce official ECC support[1] on their Notebook boards (pretty likely this will never happen). I would love to just use the coolermaster case with small efficient modern notebook hardware for 600 bucks just because I don't need additional harddisks and I would just buy one to support the company. [1]: https://community.frame.work/t/ecc-error-correcting-code-defaultly-supported-already-ryzen-amd/1716 https://community.frame.work/t/ecc-error-correcting-code-def...

kogepathic 3y ago

> The Fujitsu ones have D3417(-A!) boards (not -B) having proprietary power supplies with 16 power pins (no 24pin ATX but 16pin). There are Adapters on Aliexpress for 24PIN to 16pin (Bojiadafast), but this is a bit risky - I'm validating that atm. They work just fine. The pinout is well known [1]. You can also adapt a normal ATX PSU if you boost 5VSB to 11V. Fujitsu boards are great, and very inexpensive to purchase in the EU. Someone has even reverse engineered the license for the KVM features of their remote management (iRMC S4/S5) [2] [1] https://web.archive.org/web/20200923042644/https://sector.biz.ua/docs/fujitsu_16pin_to_atx_24pin_power_adapter/fujitsu_16pin_to_atx_24pin_power_adapter.phtml https://web.archive.org/web/20200923042644/https://sector.bi... [2] https://watchmysys.com/blog/2023/01/fujitsu-irmc-s4-license/ https://watchmysys.com/blog/2023/01/fujitsu-irmc-s4-license/

sandreas 3y ago
Oh this is pretty interesting, thank you very much. You mean that the Bojiadafast adapters work fine? If so, I wonder if there is a hit in efficiency because of the required step-up / step-down converters in that adapter. However, on my board there seems to be an unsoldered 24pin connector, that could be just used as is with a little soldering, but since it is on-hold my replacement system if my ...-B variant dies, I'm not willing to risk too many experiments :-)
- kogepathic 3y ago
  > You mean that the Bojiadafast adapters work fine? I have no experience with this particular vendor/brand but the adapters are not complex. If you're concerned, verify that it has 11VSB before you plug it into the motherboard. Otherwise it's entirely passive (passing through PS_ON#, PWR_OK, and 12V) > If so, I wonder if there is a hit in efficiency because of the required step-up / step-down converters in that adapter. It should be quite nominal. The boost converters have a very high efficiency (>85%) and standby power is typically below 10W max.
  - sandreas 3y ago
    Cool thank you... I also thought about soldering a 2$ female 24pin plug to the spot where it is on the (-B) variant... seems that they just left the plug out but the connections are there. We'll see :-)