Comment by MrFoof | Hacker News Reader

Comment by MrFoof | Hacker News Reader

Parent Full thread

Indeed, modern hardware is getting ridiculously efficient, and if you go the extra mile to disable hardware you don't use, as well as strip your configuration to run only the software you absolutely need, you can really drive it into the ground.

The later hardware revisions (such as 1.5) of the Raspberry Pi 4B, even with 8GB of memory, can idle at ~1.33W on WiFi with everything unnecessary disabled -- 2GB versions are close to 1.1W. The 8GB version on power-hungry Ethernet (at 100Mbps) can actually be under 1.45W, compared to 1.68W for 1Gbps. PoE is nice, but the hat adds a lot of power -- though it's possible with an 8GB to squeak under 2W via PoE at 1Gbps, though only just barely (~1.98W).

-- -----

For those who go, "but I need real power"... you can have an 8-core Ryzen 7 PRO 5750GE (with GPU cores for transcoding), 64GB of memory, multi-TB NVMe SSD, and two NICs (one 1GbE and one 2.5GbE) plus WiFi and Bluetooth idle at under 9W. No, it's not 4W, but to say it can do nearly anything you can imagine is an understatement. You're at the same CPU performance of an i5-12600, with tons of memory for VMs galore, in a system that has the same idle power as a cable modem.

I believe Apple rates the newest Mac Mini with an M2 Pro, 32GB memory and 2TB SSD at just 7W idle. Again, given it's capability, that's pretty wild for 7W.

> The later hardware revisions (such as 1.5) of the Raspberry Pi 4B, even with 8GB of memory, can idle at ~1.33W on WiFi

How is this measured? I spent _years_ doing experiments on the Raspberry Pi 4B and was never able to get it to idle to anything less than 3ish W at the wall, albeit arguably this was not with the "1.5" hardware revision.

One day I bought an ASUS PN40, which has an x86 Intel N4000 CPU.... and on the first day, just after installing openSUSE, without even customizing _anything_, I measured _1.7 W_ power consumption at the wall. This is for a full x86 server, with GbE ethernet, 8GB of RAM and a 1TiB SATA SSD.

That server is still my one and only homeserver. It has my calendar, media files, talks to my Zigbee/BluetoothLE devices, and it does so at a fraction of the power used by the Raspberry Pi that TFA uses for waking up his homeserver.

So to ultimately validate you can either use pretty pricey USB power meters that not just display the amperage and voltage, but can record that telemetry, which you can also validate further with clamps and multimeters.

There is a ton you can do to drop Pi power consumption into the ground, which largely is turning off all the hardware on the SBC you aren’t using, followed by undervolting (same clocks, just less power to maintain them), and last by using a very lightweight configuration (look at the Diet Pi distribution) to keep CPU and network interface(s) in the lowest possible power state by just running the absolute minimum you need (and lightweight versions at that) to do the job. After that you could further consider underclocking. Disabling cores can get further tiny gains as well.

Then you could consider purely PXE-booting where you only boot from the SD card to then disable it — relying on keeping everything memory-resident — and having permanent storage be available via network storage presented via ISCSI, or block storage. You’d then only turn on the SD card as part of configuration management changes, to turn it back off after the configuration has finished — or just rebuild the image that’s available via the network then reboot the Pi to start using it. You also then can be picky about what SD card you use, by actually tracking the power consumed accessing particular models — yes, there is a measurable difference.

If I went truly “all in” I’m fairly certain I could have usable 2GB Rev. 1.5 4Bs under 1W of power consumption, with 8GB models under 1.2W, though I’ve been focusing my hobbyist time elsewhere. This isn’t even looking at the Pi Zero 2W, or the Pico, which can be dropped to absolutely absurd low levels of power consumption and do some actual useful things.

I really should do some YouTube videos at some point. I’ve a coworker that’s really starting to ramp up his sub count on his homelab shenanigans, and he says what I’m doing relative to the largest homelabbers blows his mind at times.

> For those who go, "but I need real power"

There's also the flip side, which is "well since I already have real power..."

I.e., like many HNers, I own a powerful coding/gaming Linux desktop with plenty of HDD space, which gets regularly backed up to cold storage. It does however use a frankly obscene 90W when idling, due to the GPU, fans, and multiple hard drives and peripherals.

Building a separate, power-efficient home server would take a year or two to break even, even compared to running the workstation 24/7. But it's more like 3-5 years when you consider that it's going to run several hours most days anyway, and that it can be safely put to sleep at night. Plus of course it's more effort to manage two servers rather than one, so everything else being equal the single desktop is preferable.

Now if I can use some of the tricks described in this thread to turn it off when I'm not using it, and only temporarily wake it up when I need to stream or upload something, then using the desktop as a home server becomes kind of a no-brainer.

Of course, if you only use a laptop, then a low-wattage home server is a totally sensible complement.

FWIW, everything added together for my homeserver in the parent comment was about €120 ;) With the 0.4€/kWh that I pay here, that’s worth it quite quickly.

How much storage do you have? That's typically the biggest expense as you put more stuff on your server (i.e. media, both upload and download).

Around 2TB with little need for that to ever grow.

My workstation is a Zen 3 Threadripper Pro that mostly serves as a virtualized Spark cluster, an entire virtualized infrastructure stack, and video editing (plus occasional games). Wife lives in CAD tools on similar hardware with a bit less CPU and less memory. We already have real power.

- - - - -

However, we mostly prefer to have those off when we’re not doing work that requires them (I’m mostly on an iPad), plus there’s stuff that the entire family benefits from with the lab. Considering their performance the workstations are actually very power efficient setups for their capability, but they can noticeably change the temperature of a room in an hour when they’re under load, and both of them combined add about $35-40/mo to our electric bill (we pay ~$0.50/kWh in the Boston area) just for light loads, sometimes quite a bit more.

- - - - -

I will never suggest building something new JUST to reduce power consumption of existing stuff. ROI never makes sense. But when you DO build something new, I DO suggest looking hard at power consumption. There the savings can be quite material. Plus never having to hear any of the equipment EVER (barring the workstations going flat out) is a great thing. Silence is golden. You’ll also learn a lot more about your hardware.

> For those who go, "but I need real power"

I think those usually also want real storage space, NVMe either can’t offer that, or it would make the build so expensive, electricity stops mattering. And in those cases, 4-12 24 TB HDDs will out-consume your other hardware anyway ;)

It depends how one uses the storage, but on my home server I didn't need to access the (external) HDD storage often so I created a systemd service that spun it down after set time of inactivity. The response when the HDD is needed again is not instant, but okay enough for what I needed.

/usr/sbin/hdparm -S 120 /dev/sdb

https://wiki.archlinux.org/title/Hdparm

Yep. I have a storage server, also with a 5750GE, 128GB ECC memory, 40Gbps (QSFP+, passive DAC), EIGHT 16TB drives off a modern HBA, two 2TB NVME SSDs (mirrored) as a cache/ingest tier, a SATA DOM to boot off of, and a BMC (which alone draws 3.2W).

132TB raw storage organized into 16TB and 64TB ZFS volumes, plus the 2TB ingest/cache tier. Can saturate 40Gbps up and down.

43-44W at the lowest idle state, about 55W with just the 16TB volume spinning. Though with all the memory and the cache/ingest tier, it doesn’t spin up often, and often it’s mostly spinning up just the two-disk volume. So most of the time it is completely silent — not even a fan whirr.

And that provides the bulk storage and backups (plus pushing offsite) to the PCs, a six-node Pi cluster, and those 9W virtualization servers by providing iSCSI volumes. Also lets any new device PXE boot, and if a device has no OS and no configuration, the hardware will self-install, then self-configure after the first boot. A lot of work, but very valuable experience that paid off at the day job.

A lot of VERY deliberate hardware choices to drive power consumption that low, and probably two weeks measuring power consumption of everything. Something people fail to realize that 10GbE Ethernet is rather power hungry if you’re using copper, but if you’re using fiber or (ideally) passive DACs, you can eke out lower latency, better performance, and drive power consumption into the ground. It matters when you consider you pay for that power consumption for 10GbE copper at every single RJ-45 port in the chain between the devices. You can pay < 10W for all the devices in a 40Gbps connection (four ports) with passive DAC, but with copper it’s more like 40-42W.

Reply on news.ycombinator.com