997
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 26 Jul 2024
997 points (99.5% liked)
Technology
60123 readers
3153 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 2 years ago
MODERATORS
Linux works great on ARM, I just want something similar to most mini-ITX boards (4x SATA, 2x mini-PCIe, and RAM slots), and I'll convert my DIY NAS to ARM. But there just isn't anything between RAM-limited SBCs and datacenter ARM boards.
arm is a mixed bag. iirc atm the gpu on the Snapdragon X Elite is disabled on Linux, and consumer support is reliant on how well the hardware manufacturer supports it if it closed source driver. In the case of qualcomm, the history doesnt look great for it
Eh, if they give me a PCIe slot, I'm happy to use that in the meantime. My current NAS uses an old NVIDIA GPU, so I'd just move that over.
Apparently (from another comment on a thread about arm from a few weeks ago) consumer GPU bioses contain some x86 instructions that get run on the CPU, so getting full support for ARM isn't as simple as swapping the cards over to a new motherboard. There are ways to hack around it (some people got AMD GPUs booting on a raspberry pi 5 using its PCIe lanes with a bunch of adapters) but it is pretty unreliable.
Yeah, there are some software issues that need to be resolved, but the bigger issue AFAIK is having the hardware to handle it. The few ARM devices with a PCIe slot often don't fully implement the spec, such as power delivery. Because of that, driver work just doesn't happen, because nobody can realistically use it.
If they provide a proper PCIe slot (8-16 lanes, on-spec power delivery, etc), getting the drivers updated should be relatively easy (months, not years).
Datacenter cpus are actually really good for NASes considering the explosion of NVMe storage. Most consumer CPUs are limited to just 5 m.2 drives and a 10gbit NIC. But a server mobo will open up for 10+ drives. Something cheap like a first gen Epyc motherboard gives you a ton of flexibility and speed if you're ok with the idle power consumption.
I'm kind of not. I don't need a ton of drives, and I certainly don't need them to be NVMe. I just want 2-4 SATA drives for storage and 1-2 NVMe drives for boot, and enough RAM to run a bunch of services w/o having to worry about swapping. Right now my Ryzen 1700 is doing a fine job, but I'd be willing to sacrifice some performance for energy savings.
Arm servers are slow, and arm laptops are not compatible with Linux.
Eh, it looks like ARM laptops are coming along. I give it a year or so for the process to be smooth.
For servers, AWS Graviton seems to be pretty solid. I honestly don't need top performance and could probably get away with a Quartz64 SBC, I just don't want to worry about RAM and would really like 16GB. I just need to server a dozen or so docker containers with really low load, and I want to do that with as little power as I can get away with for minimum noise. It doesn't need to transcode or anything.
ARM laptops don't support ACPI, which makes them really hard for Linux to support. Having to go back two years to find a laptop with wifi and gpu support on Linux isn't practical. If Qualcomm and Apple officially supported Linux like Intel and AMD do, it would be a different story. As it is right now, even Android phones are forced to use closed-source blobs just to boot.
Those numbers from Amazon are misleading. Linus Torvalds actually builds on an Ampere machine, but they don't actually do that well in benchmarks.
https://www.phoronix.com/review/graviton4-96-core
AWS' benchmark is about lambda functions, not compile workloads, which are quite different beasts. Lambdas are about running a lot of small (so task switching), independent scripts, whereas compiling is about running heavy CPU workloads (so feeding caches). Server workloads tend to be more of the former than the latter.
That said, I'm far less interested in raw performance and way more interested in power efficiency and idle and low utilization. I'm very rarely going to be pushing any kind of meaningful load on it, and when I do, I don't mind if it takes a little longer, provided I'm saving a lot of electricity in the meantime.
Man so many SBCs come so close to what you're looking for but no one has that level of I/O. I was just looking at the ZimaBlade / ZimaBoard and they don't quite get there either: 2 x SATA and a PCIe 2.0 x4. ZimaBlade has Thunderbolt 4, maybe you can squeeze a few more drives in there with a separate power supply? Seems mildly annoying but on the other hand, their SBCs only draw like 10 watts.
Not sure what your application is but if you're open to clustering them that could be an option.
Here's my actual requirements:
Services:
The ZimaBlade is probably good enough (would need to figure out SATA power), I'll have to look at some performance numbers. I'm a little worried since it seems to be worse than my old Phenom II x4, which was the old CPU for this machine. I'm currently using my old Ryzen 1700, but I'd be fine downgrading a bit if it meant significantly lower power usage. I'd really like to put this under my bed, and it needs to be very quiet to do that.
Those are tough requirements to meet, I'm not sure there is a low power CPU that can do it all. You would likely need to cluster some devices but that means you need a separate NAS anyway and that kind of defeats the purpose for your case.
Servers being slow is usually fine. They're already at way lower clocks than consumer chips because almost all that matters is power efficiency.