Small form factor PC, with ECC - Not NAS - 2026 edition

About 2 years ago I researched replacing my old ASRock DeskMini A300 desktop, with something that supported ECC RAM. Never got a Round Tuit, so that did not happen. Here is a link:

Recently I found that ASRock started making a new mini PC with ECC RAM capability, called DeskSlim X600. This is smaller than the DeskMeet X600, (or the older AM4 socket X300). Still larger than my existing DeskMini A300. Yet not too big.

Well, even with the AI Apocalypse affecting both memory and storage availability and cost, I bit the bullet, (it felt like .50 caliber, on full auto :frowning: ), and bought one. Hope this was a good decision. And lasts for as long as my existing DeskMini A300, (which I bought in 2019, 7.5 years ago).

All the parts arrived except the CPU, and I started building it. Including 16GB ECC EC4 RAM, (at too high a cost…). In the short term I will run a single NVMe, partly to give it a different wear amount than a second NVMe for Mirror. The real CPU with ECC support is at least a month away. But, I found a cheap deal for a temporary CPU, (without ECC :frowning: ) This arrives Friday.

This DeskSlim seems like a nice PC. It has:

  • AMD AM5 CPU support, both with internal graphics and PCIe slot graphics. Up to 120watt CPUs / APUs.
  • ECC RAM, up to 256GB
  • Triple builtin video out
  • 2.5Gbps Ethernet
  • 10 USB external ports, (and 2 x USB 2 internal ports on header)
  • Quad storage: 2 x NVMe 2280 and 2 x 2.5" SATA
  • Half height PCIe, double width slot
  • M.2 slot for WiFi & Bluetooth, (which my X600 included a dual band WiFi card)

There is an Intel version of the DeskSlim, but it does not support ECC RAM. Neither do any of the Intel DeskMeet or DeskMini models…

With the temporary CPU I can start loading the software, (including Linux root on ZFS), which will take time. So, if anyone is interested, they can read my follow on post with the results. Can’t really test the ECC RAM feature without paying for the Passmark version of MemTestx86.


PS: All unbuffered ECC RAM appears to be EC4 variety, which means 72 bit. ECC DDR5 also supports 80 bit, which I would guess would be called EC8. The "EC" part implies how many ECC bits each 32 bit sub-channel would get. So EC4 is half of the 8 ECC bits of 72 bit RAM, so 36 bits for sub-channels.

I find myself only every using 1 pciex16 slot for my main desktop & always wanting an endless amount of pcie for my homelab crap. Any reason for only 1 pcie slot that doesn’t support bifurcation?

I do have to admit that the $200 for motherboard/case/psu isn’t a terrible combo.

For me, I don’t even need 1 PCIe slot, it is just a desktop. If the built in APU graphics are good enough, then I don’t want to bother with discrete graphics.

In fact, if a DeskMini, (which has no PCIe slots, just 2 x M.2 NVMe slots + 2 x SATA 2.5" bays), had ECC memory, I’d buy an updated DeskMini. These are small desktops but still can take reasonably powerful 65 watt CPUs, (which must have builtin graphics).

The goal was ECC memory on my next desktop.

Now I am somewhat glad that the DeskSlim X600 has a PCIe slot. With only 2.5Gbps Ethernet builtin, the slot can be used for a 10Gbps Ethernet PCIe card. I am SO over the slow speed of 27 year old 1Gbps Ethernet. Even 10Gbps was approved in 2002, 24 years ago.

There is a PCIe card that supports both copper 10Gbps Ethernet and 2 x M.2 NVMe:

Synology 10Gb Ethernet and M.2 Adapter card E10M20-T1, RJ-45; 1 port

It’s half height, which is needed for the DeskSlim X600. Not sure yet if it requires bifurcation. Or is too long. Just something I ran across and would consider if it seems like it will work.

Well, I’ve put the DeskSlim together and am now running 24 hour’s worth of memory tests. First pass was successful, which I’d hope it should be, considering the arm, leg and other body parts a single measly 16GB, (EC4 ECC), DIMM cost.

Seems the Gentoo Admin ISO image includes MemTest86+. Originally tried the standalone MemTest86+ ISO image but it was not recognized.

The memory tests worked fine. Even got the proper CPU installed:

AMD Ryzen 7 PRO 8700G w/ Radeon 780M Graphics

Rebuilt the kernel with memory encryption and the USB4 features. (Though I seemed to have missed one. Oh, build another kernel tomorrow.)

This thing is fast, even with a 2 year old type CPU. Compiling all the software or kernel was so fast, it prompted me to add more compile jobs. But, one package really did not like that. Using 23 compile jobs, (even though I only have 16 threads), worked for most packages. Yet one wanted 57GBs of memory! Yikes, 2GBs per compile job. I have 16GBs of memory and 32GBs of swap, so it failed. Dropping the amount of compile jobs to 15 worked. And was still quite fast.

A couple of minor problems. At one point, my new DeskSlim X600 would not always see the boot media. That seems to have been solved, at least 10 reboots without issues. Another problem is that the network device changed names a few times after reboot. Now I have to fight that problem, (which should not exist because their is exactly 1 Ethernet port, and no WiFi).

As for recommending this for others, at present it seems like a good, small desktop, (which supports ECC RAM). The feature set is good, 4 x full size DIMMs, 2.5Gbps Ethernet, place for WiFi & it’s antennas. Lots of USB ports, (2 x C, 4 x 2.0 and 4 x 3.x). Dual, (potentially triple), video from iGPU, (HDMI & Display Port). PCIe slot, double width, half height. 2 x NVMe and 2 x 2.5" SATA bays.

As for the video, apparently the 8000 series CPUs have USB4 / Thunderbolt, (I have not confirmed this). But, other CPUs, (7000 & 9000 series), allow Display Port over USB C, giving yet 1 more video out.

This means I now have 2 online computers with ECC RAM. And a third waiting to be rebuilt. Wish I could get ECC RAM for a laptop… not that I can afford a new one now.

Edit: When I first starting running this DeskSlim, I had to keep checking the CPU fan. It’s one of those Noctua coolers with a quiet fan. So, it was too quiet. But, whence I started loading software, the CPU fan ramped up and made enough noise that I knew it was working.

Well, the booting problem returned with a vengeance. This was so frustrating, as devices would disappear in the BIOS. Even all of them, (NVMe, laptop HDD and USB flash drive…). It was so irregular that I was about to give up on Grub and see if EFI booting would work.

Eventually I found and loaded something called “refind”, which is an EFI boot & menu system. Installing it was simple and the first boot used the Grub menu system. This was perfect, as this kept the ZFS alternate boot environments I painstaking developed over 10 years.

So “refind” acts as the initial booting software, the thing that Grub was not doing correctly. But then launched the Grub EFI menu system. This even works for the USB flash drive, and old laptop HDD installed.

There was some minor configuration needed in “refind” so that it was easy to tell which boot device to select. That is because it uses icons instead of names. However, a simple FAT FS label caused “refind” to print that label, thus easy to tell NVMe from SATA from USB, (assuming the labels are appropriate, like “NVME_efi”).


To make an upside of this, I get to rant, complain and hate on Grub, in my blog… (Not that I needed another subject…)

GRUB is grubby. Nomen est omen.
Thankfully, you could refine it with rEFInd. :wink:

When I said this thing is fast, compiling a new Linux kernel, (after “make clean”, so it’s everything selected), that’s less than 5 minutes. And yes, this is very repeatable. Due to new hardware and needing new features, (like Thunderbolt / USB4), I’ve have had to make 10 kernels over the first 2 weeks I’ve owned this desktop.

I remember in the past it could take up to a hour to compile the kernel. (Obviously on slower hardware… but sometimes more memory! Like 32GB verses the 16GB I have now…)

The root pool uses:

# zfs list rpool
NAME    USED  AVAIL  REFER  MOUNTPOINT
rpool  26.2G  70.4G    96K  legacy

And scrubs take 14 SECONDS!

Scrub started  Tue Nov 10 17:31:00 2026
Scrub finished Tue Nov 10 17:31:14 2026

To be fair, this is using a PCIe 4.x, 4 lane NVMe SSD. But still, scrubs averaging more than 1GB per second?!

Eventually when I Mirror the root pool to another NVMe, (if they don’t cost an arm and leg), I expect it to slow down some. Who knows, maybe if the Mirror NVMe is as fast, it will take the same amount of time.