Hacker News Reader

Comment by cromka | Hacker News Reader

Why do we never see such cheap, low-wattage CPUs with ECC memory controllers? There’s absolutely a market for these, starting with home NAS and generic servers. I just don’t get it, such a missed opportunity here!

View on HN

Intel uses ECC as a differentiator between consumer chips and server chips (and pro-sumer?)

They’d never do that.

They SHOULD. Everyone had too much memory and it’s too dense. We should all have ECC by now.

Unless AMD does it I don’t think it will happen. Even if they do I feel like OEMs wouldn’t want it because memory would suddenly cost 2¢ per GB (or whatever) and refuse.

AMD does it in their embedded low-voltage CPUs, which are not easy to be found in regular applications. V2000 and V3000 series have 10W TDP versions.

Any ATX solutions are super expensive: https://solidrun.atlassian.net/wiki/spaces/developer/pages/5...

Intel actually did add ECC in their Atoms, see C3000 for example:

https://www.intel.com/content/www/us/en/products/docs/proces...

I wonder how ibecc compares with real ecc, now that it's supported for these lower spec chips.

Will it help at all or is it a waste of time?

It reduces the memory capacity and the memory performance, in comparison with real ECC.

Nevertheless, this will rarely matter for the kind of applications typically run on small computers with Atom CPUs, which are rarely more demanding than a router/firewall or a network-attached storage server.

You can see examples of performance decrease with IBECC here:

https://www.cnx-software.com/2024/05/26/odroid-h4-plus-revie...

For most applications the performance decrease should be less than 10%. The Intel IBECC controller includes a special cache memory, to decrease the number of additional memory accesses required by IBECC. Otherwise the memory performance would have been halved.

The recent generations of Atom CPUs for industrial applications include an in-band ECC controller, which works with non-ECC DDR or LPDDR memory.

Besides the Alder Lake N or Amston Lake CPUs that are branded as the Atom x7000 series, some of the models branded as N CPUs, e.g. N97 and i3-N305 also have a functional IBECC controller, if the computer BIOS provides means to enable it.

But DDR5 has inband ECC by default, so it’s hardly any consolation to be honest.

I stand corrected, DDR5 has on-die, not in-band ECC.

Some N100 do have In-Band ECC (which uses the same data lane as regular memory channel, so less usable memory space/effective memory bandwidth), such as the Beelink S12 Pro box (DDR4) and is recognized by EDAC subsystem on Linux.

    [    8.485808] [    T703] caller igen6_probe+0x138/0x780 [igen6_edac] mapping multiple BARs
    [    8.487601] [    T703] EDAC MC0: Giving out device to module igen6_edac controller Intel_client_SoC MC#0: DEV 0000:00:00.0 (INTERRUPT)
    [    8.487625] [     T67] EDAC igen6 MC0: HANDLING IBECC MEMORY ERROR
    [    8.487626] [     T67] EDAC igen6 MC0: ADDR 0x7fffffffe0
    [    8.487664] [    T703] EDAC igen6: v2.5.1

DDR5 has in-band ECC by default on all platforms, though.

DDR5 doesn't have an In-Band ECC. What they have is On-Die ECC, and is not visible to the memory controller (i.e. error information won't be reported by EDAC), and only corrects 1-bit when they're stored/retrieved, where Side-Band ECC (regular ECC) and In-Band ECC can correct 1-bit and detect 2-bits, and can also correct/detect memory corruption during transit.

Ohhh right, right, I had them mixed up! Thanks for the clarification.

I saw this $390 Ryzen NAS enclosure yesterday but it’s never clear with these configurations if the motherboard supports ECC, even if the CPU does:

https://liliputing.com/aoostar-launches-a-ryzen-7-5825u-vers...

True. I've wondered the same thing, and always come back to inexpensive Ryzens on ASRock motherboards that I simply underclock.

We're kind of stumbling into a golden age of home NAS devices and while there aren't any ready-to-go ECC options yet, you can buy lovely little cases to support your own components. I think we'll surely see ECC in the next generation of AOOSTAR and similar devices, nobody making these devices has any motivation to never implement ECC.

How exactly are companies like AOOSTAR going to implement ECC if it's not supported by the CPU vendor (which is Intel)? AMD CPUs are useless for these home NAS devices if your goal is to build a low-power media server, which I think is one of the main use-cases (why else would you want to store tens of terabytes of data at home?).

ECC has been growing in support since ~12th generation Intel (currently 14th) and before that for AMD I think, so it is only a matter of time before this trickles down into the little NAS boxes.

>ECC has been growing in support since ~12th generation Intel (currently 14th)

It is? How so? I don't see it anywhere outside their server and embedded products. The server-grade stuff isn't appropriate for "the little NAS boxes" used at home because of both expense and more importantly power consumption, and the embedded stuff just isn't really available to consumers (so it might be potentially great for a company like Synology that builds consumer-friendly pre-built NAS boxes running on proprietary software, but it's not useful to people building their own DIY NAS boxes running open-source software).

ECC is supported on any higher end 12th/13th Gen, but it's locked behind a chipset feature lock. If you can find any board that uses W680 (e.g. W680D4U-2L2T which can be had on Amazon for $569), then it's possible to use ECC UDIMM on Intel Core and have ECC detected/reported.

>e.g. W680D4U-2L2T which can be had on Amazon for $569

Yeah, that cost puts it out of reach for the typical home users building relatively inexpensive little DIY media server boxes, and places it firmly in "server" territory. A standard motherboard for these CPUs (typically using a B760 chipset) is a very small fraction of that cost.

It's good to know this factoid, but it still basically supports my prior assertion that on Intel, ECC support is basically only for servers.

The very old HP Microserver Gen7's used a low power AMD cpu, as did the original Gen10 model.

The Gen7 things would never be useful for transcoding on cpu though, waaaaaay too underpowered.

But they did have a PCIe slot for add-in cards (ie small gpu or network card).

Of course you can add a GPU card on the PCIe slot, but this 1) takes away a slot, and PCIe lanes, that you might want to use for connecting storage devices (such as a SATA HBA card), and 2) this will probably poorly affect your system's overall power consumption, which is a big factor for home users building NAS/media server boxes. Intel's QuickSync performs extremely well, with very low power (esp at idle) for 4k transcoding, for very low cost, while not needing any PCIe lanes; it's why these CPUs are so popular for this application. Unfortunately, Intel stupidly refuses to allow ECC support on non-Xeon chips, so that's a big drawback, as ECC is really useful for a storage server (or really anywhere).

> Intel stupidly refuses to allow ECC support on non-Xeon chips

They at least used to support ECC on some models of i3 chips as well.

My understanding was that was specifically to support ECC on lower power NAS systems.

Random examples:

• https://ark.intel.com/content/www/us/en/ark/products/90729/i... (6th gen i3)

• https://ark.intel.com/content/www/us/en/ark/products/97455/i... (7th gen i3)

• https://ark.intel.com/content/www/us/en/ark/products/129944/... (8th gen i3)

• https://ark.intel.com/content/www/us/en/ark/products/190886/... (9th gen i3)

• https://ark.intel.com/content/www/us/en/ark/products/203899/... (10th gen i3)

Looking over random i3 pages from 11th gen onwards though, I'm not even seeing ECC mentioned on the pages at all.

It's not a "yes/no" item like on the older generation ones pages, it's completely absent (from the ones I just checked). :(

It looks like these CPUs were specifically targeted at embedded devices; they aren't things you can normally purchase yourself easily like the regular desktop chips, and if you look at the desktop variants (the ones without any letter suffixes), they don't support ECC.

You can still see it even in 13th generation CPUs: https://ark.intel.com/content/www/us/en/ark/products/232105/...

But as you can see here, this is an embedded variant; you're not going to find it at Amazon, as it's probably only sold directly to manufacturers.

I own one if these i3s from 6th gen. Bought at retail. Nothing special.

> they aren't things you can normally purchase yourself easily like the regular desktop chips

That's not really relevant to the point of them only allowing ECC on Xeon chips though. They do (or at least did) support ECC on non-Xeon chips.

These days though, people are probably better off buying an appropriate Ryzen processor and motherboard that explicitly supports ECC. That's what I did anyway. :)

>These days though, people are probably better off buying an appropriate Ryzen processor and motherboard that explicitly supports ECC.

As I pointed out before, you can't do that for a media server. Transcoding is necessary for many (most?) users.

>That's not really relevant to the point of them only allowing ECC on Xeon chips though. They do (or at least did) support ECC on non-Xeon chips.

That's a distinction without a difference: those other non-Xeon chips aren't useful if you can't even buy them because you're not an OEM.

They look pretty easily available to me, or are you going to exclude Ebay now as well?

https://www.ebay.com/itm/305342386279 (note: I don't know the seller)

Yes, I'd exclude anything on the secondhand market, especially things you have to buy from overseas sellers where there's no possibility of support or returns. This is not like a normal new part where you can purchase it firsthand and expect it to perform as advertised or else get a refund. Furthermore, the (Intel) chipsets on motherboards sold to consumers don't normally support ECC either, so it seems you'd have to get something that's not designed to Intel specifications, probably on Aliexpress, so it's questionable if ECC would even work, and you can certainly forget about any warranty or returns.

Seriously, if you can't just buy it at Best Buy/NewEgg/Amazon (not from 3rd-party overseas sellers), then it's not applicable to this discussion at all.

> if you can't just buy it at Best Buy/NewEgg/Amazon (not from 3rd-party overseas sellers), then it's not applicable to this discussion at all.

Maybe the problems you were having with sourcing the cpus you wanted have something to do with the arbitrary limitations you were working under?

Meanwhile, other people (as per sibling comment) were able to buy these same chips without issue. ;)

> if you look at the desktop variants (the ones without any letter suffixes), they don't support ECC.

That's not correct. I literally linked to some non-suffixed processors that support ECC in the comment you're replying to.

I don't see any currently-sold CPUs there. Everything is years out-of-date, maybe even a decade.

The problem with those inexpensive Ryzens is that they don't have iGPUs with Intel QuickSync so they can't transcode video with high efficiency. This makes them unsuitable for use for media servers, which is what a lot of people are doing with these chips these days: building home NAS/media servers that combine both functions into a single, small, low-power unit. You put a bunch of HDDs in it so you can store tens of terabytes of movies using RAID to protect against drive failures, and run Jellyfin or Plex on it. Plus you can use the box for other regular NAS tasks too, as well as many other convenient applications. Intel chips are virtually required here since AMD chips are useless for transcoding, and require external GPU cards which destroy the system's power efficiency which was one of the main points of getting one of these low-power Intel chips in the first place.

Doesn't AMD make Ryzens that also have iGPUs that work with jellyfin/plex transcoding?

From what I've read, the AMD offerings here are not at all competitive with Intel's QuickSync.

While Intel QuickSync remains faster in CPUs of the same generation with AMD, the AMD laptop CPUs starting with Phoenix (Ryzen 7440/7540/7640/7840/7940) have made a great jump in the video encoding/decoding speed, exceeding the older Intel CPUs, and they should be suitable for any such application.

There are a huge number of models of cheap small NUC-like computers with AMD Phoenix or Hawk Point (Ryzen 8000 series) CPUs.

Good to know!

I run Jellyfin on an 8-year-old Intel NUC. Works great. Don't need to worry about transcoding on the fly because I make sure all video files are in H264 going in, which can be direct-streamed to pretty much all clients.

https://jellyfin.org/docs/general/clients/codec-support/#vid...

>Don't need to worry about transcoding on the fly because I make sure all video files are in H264 going in, which can be direct-streamed to pretty much all clients.

That doesn't work if you use subtitles. With many formats, the subtitles have to be transcoded on-the-fly into the video, even if the source and destination resolution and codec are the same.

Also, the Intel NUC might be fine just running Jellyfin (without transcoded subtitles) using separate file storage, but if you try to combine that with a NAS box running 4-8 drives using ZFS and perhaps a few other services as well, it's probably just not enough.

Subtitles, or bitrate conversion for streaming devices.

Subtitles aren’t transcoded into any video. They’re remuxed into the media container. Totally different.

Transcoding using hardware acceleration is one of those things that some people insist upon, like wanting "gaming" everything and overclocking everything, but most people don't need.

First, the quality of output of hardware accelerated transcoding is in most instances noticeably worse than CPU transcoding, as in most people can discern the difference watching two videos side-by-side.

Second, CPU transcoding is more than enough to keep up with most streams. For instance, 1920x1080 h.264 streams can be done in real time by even a modest 4 core processor. People who want 4K h.265 can buy a fancier CPU.

Third, most people are served perfectly fine by pre-transcoded files. Your comment about subtitles isn't correct - most streaming devices can accept subtitle metadata, and even if subtitles need to be burned in, they can be burned in to a static file and/or generated in real time by a CPU.

What's interesting is that the same people who want Intel Quicksync claim they want to transcode to the best quality that a device can support, then say they don't want to transcode to a static file because they don't want to transcode to a lower resolution / bitrate than what their best device can support, but want to be able to view things at the highest quality wherever they want. They've been tricked, because if they really cared about quality, they'd care about the difference in quality between QuickSync and CPU encoding.

That's why I tell people that if the really want highest quality real time transcoding and have sources that are ultra high quality, they'd be much happier with a Ryzen 7900 than with Intel QuickSync. Side-by-side comparisons would bear this out every time.

> People who want 4K h.265 can buy a fancier CPU.

Or they can just get an N95

...and have worse quality, which makes the idea of caring about using real-time transcoding to give the best quality pretty silly.

You are comparing a cpu that (... checks google ...) retails for $368 _by itself_ against one that comes in entire-system-minipcs at $149. Yes, the CPU perf is obviously better. In general HTPCs are setup as: 1) direct play or remux for all the primary usecases 2) optional transcoding

The cpu doesnt generally matter for #1. For #2 you are already making tradeoffs so its a bit of a non-argument that "you are losing quality". Its not the primary usecase (either its other users, or streaming on the go).

I too don’t understand the need for transcoding if the goal is to get best quality. Any transcoding will reduce quality, and given the capability of modern players, the best quality is just the original media file.

The Alder Lake N CPUs have in-band ECC memory controllers, because they are also sold as Atom CPUs for industrial applications.

I do not know if the IBECC controller is functional in all models, but it is usable at least in the models that are marked by Intel (in Intel Ark) as suitable for embedded use, like N97 and i3-N305.

The IBECC controller requires BIOS support. Among the computers known to have BIOS support for IBECC is ODROID H4/H4+/H4 Ultra (the 3 models differ in the CPU model or the number of Ethernet ports and SATA support).

IBECC works with non-ECC DDR5 or LPDDR5 memories, but it reduces the available memory capacity and performance in comparison with traditional ECC. Because no memory vendor offers LPDDR memory with ECC, IBECC is the only way to implement ECC in the CPUs like Alder Lake N or Amston Lake, which are intended to be used with LPDDR5 memory.

I'd rather start seeing ipmi-like controllers rather than ecc memory on high-end/prosumer hardware.

Reply on news.ycombinator.com