The Most Powerful DIY NAS I've built (ft. LattePanda SIGMA)
LattePanda have made a name for themselves by making tiny single board computers with their own little twist. Unlike your typical Raspberry, Orange and Banana Pis, the LattePanda v1 was based on x86 It was marketed as a direct competitor to Raspberry Pi 4, with lower power consumption, and the ability to run any x86 operating system you want, including Windows. Fast forward to present day, and a few days ago, LattePanda sent me this. This is LattePanda Sigma. This bad boy comes with an Intel i5-1340p, 12 core, 16 threads 16 gigs of DDR5 memory with in-band ECC support three M.2 slots And two Thunderbolt 4 ports! Now that’s a far cry from the anemic Atom board that LattePanda released in 2018.
But with great specs comes great competition. LattePanda v1 might have been disruptive with its relatively low price, credit-card form factor and an x86 CPU. But LattePanda Sigma is going against a much tougher enemy The burgeoning market of miniPCs from Asia, with companies like Topton, Minisforum, Beelink and Geekom providing a lot of bang for your buck with their offerings.. So today I’m going turn this little computer into a super compact home server/NAS ...and also turn you into a mathematical genius! by telling you about today’s sponsor, Brilliant.org
Brilliant.org is a learning site unlike many others. Instead of boring you with monotonous lectures or walls of texts Brilliant breaks down super complex and intimidating subjects into very fun and engaging interactive courses. Always wanted to understand quantum mechanics? Neural networks? Special relativity? Brilliant is the way to go – No PhD required! And if you have a more practical goal Let’s say, learning math, data analysis or computer science for work or school Brilliant has got you covered too, with whole learning paths dedicated to educate you without making you fall asleep. Now I’m not a math person. So my favorite math related course on Brilliant is obviously Math History, in which you’ll learn about Pascal’s triangle, and how it’s actually existed in many other cultures for thousands of years. I’m not gonna spoil the whole course, but it was also very fascinating to learn that some antique civilizations actually used base-60 for their calculations.
So go ahead and visit brilliant.org/wolfgang , to get your free 30-day trial. The first 200 people to sign up will also get a 20% off their annual subscription. Thank you Brilliant for sponsoring this video, and now, let’s get back to our mini PC! So, what does LattePanda Sigma have that distinguishes it from dozens of generic Intel-based MiniPCs? And before we go any further – yes. Despite LattePande’s marketing calling Sigma a “hackable single board computer”, I think it’s more of an Intel NUC than it is a Raspberry Pi, if that makes sense. First of all, let’s look at the I/O. On one side, we’ve got a 3.5mm audio jack,
two USB 3 ports, two 2.5 Gigabit Ethernet ports, an HDMI port, a Thunderbolt 4 port with support for power delivery, a 12V barrel jack and an 8-pin power connector, which can also deliver 12V to the board, or even be used as a 12V out when you use a barrel jack for power. And on the rear, there’s another Thunderbolt 4 port, a microSIM slot, two USB 2 ports, and a power button. However, apart from your usual ports, Sigma also features 20-pin GPIO, which goes directly into the built-in ATMega microcontroller.
It’s an Arduino-compatible chip, which means that with this little board, you get the best of both worlds The performance and high-speed PCIe connectivity of x86, and the features normally found on the ARM-based “maker” boards, like GPIO and eDP. But don’t get too excited though. Just because this board has a built-in Arduino chip, doesn’t mean that you’ll be able to use an NVMe SSD to store your Arduino scripts, or blast through them using the 12-core Intel CPU.
As Jeff from Craft Computing helpfully pointed out in his review of the board, the x86 part and the Arduino part are completely isolated, apart from the power and serial. Finally, when it comes to other maker/industrial features, the board also has an eDP connector and a touch panel ouput. Back to some more traditional I/O, Sigma offers two full-size NVMe slots, one Gen 4, and one Gen 3.
There’s also an E-key slot for Wifi and Bluetooth, and a B-key slot for M.2 SATA SSDs. And if you need to connect a 2.5" or 3.5" SATA drive to the board, there’s also a single lonely SATA port on the other side of the computer, together with a 4-pin SATA power connector.
Now, when it comes to RAM, the bad news is that the RAM on the Latte Panda Sigma is not upgradable. My model comes with 16 gigs of memory, and there’s also a 32 gig model, but that’s basically it. Considering how beefy the CPU is, I could definitely see some people using it as a VM host, and for that kind of a use case, 64 or even 128 gigs would definitely be better. The good news however, is that the built-in memory is pretty fast.
We’re looking at LPDDR5 chips running at 6400 Mhz And to add to that, LattePanda Sigma also comes with in-band ECC support. And no, in band ECC is not the same thing as the DDR5 “on-die ECC” spec. On-die ECC only offers 1-bit error correction, cannot detect any errors, and is a purely hardware implementation. In-band ECC on the other hand, can correct 1-bit errors, detect 2-bit errors and also offers features like error monitoring and memory scrubbing. But it’s also not quite the same as regular ECC, which we’ve come to know and love. Instead of having dedicated DRAM chips just for ECC data, the in-band ECC uses existing memory channels.
This makes the memory modules cheaper, since you don’t have to ship an additional physical chip just for ECC, but it also comes at a cost of performance. In their review of an Asrock industrial NUC with Intel i7-1360p, Anandtech compared the performance of the machine before and after enabling the in-band ECC function. They saw a pretty big performance hit across the board, especially when it comes to GPU-intensive tasks. Apart from that, when it comes to OS-level support, Intel claims that in-band ECC is only supported on quote on quote “Chrome designs, but not Windows designs”. Whatever that means. I’m gonna touch on Linux support later on in the video, but long story short, it’s not quite there yet, at least for the 13th gen chips.
To sum it up, even though in-band ECC isn’t quite the same as the regular ECC, the option is there, and you can decide for yourself if the performance penalty is worth it for you. So now that we’ve talked about the specs, what are we going to do with this little board? Well, since it’s got Thunderbolt ports, I want to turn it into a super compact NAS using this Thunderbolt DAS enclosure from Terramaster. I’ve already talked about it in my M1 Mac server video, and you might remember me complaining about its built in RAID controller, which was slow, power hungry and forced you to use a tacky WebUI to configure the drives. I’ve since replaced the built-in RAID controller with an ASMedia ASM1166 card. Which is more power efficient, and also gives us full access to individual drives, including things like SMART and drive spindown. Since both the Sigma and the Terramaster DAS use a 12V barrel plug for power, I’m gonna use a 12V power supply from Leicke, and a barrel jack Y-splitter to get two DC outputs from one power supply.
The power brick itself can handle 156 W, so it should be more than enough to power both the hard drives and the board itself. So I’m going to put some hard drives into our DAS, and then we can connect everything together. ...And look at this little set up! Sure, you could probably make it look even better with a 3D-printed enclosure of some sort, but this is good enough for me.
Now, unfortunately, this DAS enclosure has one fatal flaw, which could prevent you from running it in a homelab scenario. It doesn’t turn on automatically on power loss. Which means that every time you unplug it from the wall, or lose power, you’ll have to press the power button to turn it on. Now let’s talk about the software setup. For my operating system, I initially wanted to use Unraid. But unfortunately, after booting into the OS, none of my hard drives seemed to be recognized.
I’m not sure if the Unraid kernel just doesn’t include the Thunderbolt drivers, or maybe I did something wrong, but after like half an hour of trying, I still couldn’t get it to work So, my next option was Proxmox. And, after installing the latest version of Proxmox, all of my drives seemed to be connected and working! Sort of. I still had a lot of situations where the enclosure wouldn’t be detected properly, and only came back up after replugging the cable and rebooting the machine several times.
Now whether that’s the fault of the Linux kernel, the Terramaster DAS or Proxmox itself - I’m not sure. Thunderbolt issues are not unheard of, and in general, it’s just not as robust of a protocol as something like pure PCIe. But one thing’s for sure – using Thunderbolt for your “permanent” storage is probably not a great idea.
So, after managing to make Proxmox kind of work with the DAS, I decided to run Unraid in a virtual machine. Now, is that a good idea? No. Unraid can act as a hypervisor all by itself, and you’ll probably be better off just running it bare metal.
But I was very curious if we can actually pass through our Thunderbolt SATA controller to the Unraid VM and make it see the drives that way So I created the VM with pretty standard settings, went to the Hardware tab, and started adding my devices. First up, we need to pass through our Unraid USB drive. You can’t run Unraid from an internal drive, since your license is basically tied to the GUID of a flash drive.
So emulated storage is obviously a no go. Then, we’re going to pass through our SATA controller And finally, I’m also going to pass through our Intel Xe graphics card Since I want to test hardware transcoding in Jellyfin. After the initial setup, things seemed to run pretty smoothly Unraid detected the drives right away, and I was able to create the storage array.
And then I noticed that my drives were getting very toasty. The Terramaster DAS actually has two fans on the back. But no matter how hot the drives got, the fan speed never changed.
My guess is that the built-in RAID card had some role in controlling the fan speed based on drive temperatures. And with the card gone, the DAS didn’t know when to ramp the fans up. Now keep in mind, these are Western Digital Red Pro drives, which get very hot in general, So maybe the situation would’ve been different if I used 5400 RPM drives. Anyway, moving on to hardware video transcoding I’ve set up the Jellyfin Docker container using Unraid’s amazing app store, and made sure to pass through our rendering device. Then, I copied our usual benchmark movie – a 4K HDR HEVC copy of Dune.
Finally, I enabled hardware transcoding for all of the video formats. And now, let’s look at the performance With VPP tone mapping enabled, we get around 82 FPS while transcoding this particular scene. Which is the highest score so far, even better than the dedicated Intel Arc GPu And get this, if we use software tone mapping instead of VPP, we get as much as 102 FPS while transcoding the movie. That’s really impressive! I’ve also ran the QuickSync benchmarking script by Alex Kretzschmar, and here’s what the results look like We got 188 FPS in the 1080p, H264 test 64 FPS in the 4K H264 test 101 FPS in the 1080p H265 test And finally, 34 FPS in the 4K H265 test. And here’s how that compares to some other Intel GPUs.
As you can see, the 7th gen Intel Xe graphics chip blows pretty much any other integrated Intel GPU I’ve tested so far out of the water. And that’s with in-band ECC enabled. Speaking of which, enabling the in-band ECC on this board was not easy. Initially, I couldn’t even find the BIOS option. So, I decided to update the firmware. LattePanda’s documentation is very brief on that front, and the only thing they offer is a link to a ZIP archive on GitHub.
And after unpacking it and seeing an EFI folder, I guess that you’re just supposed to dump everything onto a flash drive and boot from it? So that’s what I did, and... It worked! After the update was finished, I rebooted the machine, and voilà! In-band ECC. Sadly, even after that, I couldn’t see any kind of indication that the ECC support was present in the OS. So I started digging, and I found this little guide from LattePanda that Patrick from STH mentioned in his review of the board.
According to this document, you actually need to patch the Linux kernel in order to add in-band ECC support for our particular CPU, since this functionality is not supported for Raptor Lake CPUs yet. At least at the moment of making this video, with the kernel version 6.5.3 So, I used a little guide from Proxmox forums to download the kernel sources, replaced the value in the EDAC driver, and typed `make` to compile the kernel.
And, after installing the new kernel, and rebooting the machine, we finally have ECC! Needless to say, having to patch the kernel isn’t ideal, so let’s just hope that in-band ECC support gets added for Raptor Lake CPUs soon. As I already mentioned in the beginning of the video, enabling in-band ECC does have an impact on performance, especially when it comes to GPU tasks GPUs love fast memory, and that’s why you see super fast GDDR6 and HBM chips on dedicated GPUs. Integrated GPUs on the other hand have to use the relatively slow system memory, and the in-band ECC functionality makes it even slower. By disabling the in-band ECC in the BIOS, we do actually see some improvement in terms of performance.
In 1080p H264 test, we now get 196 FPS In 4K H264 test, the board scores 69 FPS In 1080p H265 test, we get 111 FPS and finally, in 4K H265 test, the Intel Xe graphics gets 36 FPS. Now, the difference is not super dramatic, but it is definitely there. One more thing I wanted to test on this board was Thunderbolt networking.
I usually have a 10 gig SFP+ connection between my main home server and my Macbook But this little computer doesn’t have a full-size PCIe slot for an SFP+ card. What it does have though, is a second Thunderbolt 4 port. Which, in theory, is even faster than 10GbE, and doesn’t require any adapters. So, I connected my Macbook to the LattePanda Sigma, ran `ip a` , and here you can see a new networking device, called `thunderbolt0`.
Then I used this forum post to create a quick and dirty udev rule, so that every time I connect the two computers together, Proxmox would assign a static IP address to the networking interface. Finally, I also set a static IP on my Mac, and that’s pretty much it. Now let’s see how fast it can go! As you can see, the iperf3 uni-directional test gives us 18 gigabits. not too bad! A bi-directional test gives us around 13 gigabits, which is still faster than 10 gig! The only problem with Thunderbolt networking is that passive Thunderbolt cables don’t get very long. 2 meters is pretty much the limit, and for anything longer than that you’ll need an optical cable ...which cost an absolute shitton of money. Like at this point you can buy four Thunderbolt docks and just daisy chain them until you get to the right length.
So for me personally, I guess I’m gonna stick with SFP+ for now. Finally, let’s talk about the power consumption This board is actually very power efficient, especially considering how much I/O it packs I’ve seen it draw as little as 3.6W from the wall at idle, but it mostly stayed around 4 to 5W. When it comes to C-States, we see the board going as low as C8 in Proxmox with no VMs running. While compiling the Linux kernel, the power consumption went up to 65W.
And in terms of the noise, here's what the board sounds at idle vs. full load. So, 3.6W from the wall is very impressive However, once we plug in the Thunderbolt DAS, that’s where the fun ends. All of a sudden, the board is stuck at C2, with all of the PCIe devices showing 100% usage. The power draw from the wall jumps from just 3.6W to 15W, and that’s with all 4 drives spun down.
And this is gonna sound weird, but you can actually… hear the difference. Now, just to show you how much of that overhead comes from Thunderbolt, I’ve built this contraption. Same ASMedia ASM1166 controller, same four hard drives, but this time, no Thunderbolt. Only pain, suffering, M.2 and a jumped PicoPSU with a SATA splitter.
You can even see my hands shaking in horror. After powering up the board, we see that all four of our hard drives are detected, and all of our PCIe devices now have ASPM enabled. Looking at powertop, we see C8, which is great. But the best part, by far, is power consumption. 8.3 watts from the wall, at idle. 7W lower than the Thunderbolt setup! And despite how unsexy this setup is, it’s a hell of a lot more reliable in software, too So yeah, If I were a bit more tech savvy, I’d definitely develop some kind of a drive enclosure based on M.2, instead of Thunderbolt.
Maybe an idea for the next video? So after conducting all kinds of fun, dangerous and impractical experiments on this thing, let’s come down to earth and talk about a more pragmatic subject Price. The 16 GB model of LattePanda Sigma starts at $579, that’s with no storage and without the WiFi module. And the 32 GB model starts at $629, for the no SSD no WiFI variant.
But that’s before tax. The German reseller Reichelt.de sells the board for a whopping 729€ for the 16 gig version. I used an online import tax calculator to see how much it would cost to import the board myself... And yeah, that pretty much tracks. 729€ a lot of money.
And sure, the LattePanda Sigma definitely packs a punch when it comes to performance, and comes with a lot of I/O – Thunderbolt, triple M.2, GPIO, and so on. But let’s take a look at the competition. An Intel NUC 13 Pro with two Thunderbolt ports, two M.2 slots,
and the same CPU would set you back 460€. Add two 8 gig sticks of DDR4 RAM for around 40€, and you’ll still have 229€ to spend on a 4TB SSD of your choice. And if you’re willing to sacrifice Thunderbolt, Aliexpress and Amazon offer a ton of noname MiniPCs from Topton, Kingnovy, and other brands. For instance, you can get this little PC from Topton for as little as 296€ including tax. No Thunderbolt, but it does have two 2.5 GbE ethernet ports, unlike the Intel NUC. And Ryzen-based mini PCs have been killing it lately! Like this Minisforum UM760, which can be had for 419€ after tax.
It comes with two USB 4 ports, upgradable DDR5, and Ryzen 7640HS Which actually outperforms Intel i5-1340p And when it comes to Latte Panda Sigma… Well, i’m not gonna sugarcoat it, guys. I don’t think that the built-in microcontroller and the GPIO are worth the 729€ price tag. And in some cases, you’re actually getting less for your money. All of those other miniPCs, which are a lot cheaper, come with upgradable memory and an actual case. I get that LattePanda was going for a “maker” aesthetic, but having an actual outer shell as an accessory would definitely be nice. Because as of now, your only option is going DIY.
To sum it up, I definitely think that LattePanda Sigma is a cool product. However, it’s definitely being held back by trying to be everything at once A mini PC, a maker board, a single board computer, et cetera. And the price tag means that despite its name, Sigma is just not gonna be able to compete with other MiniPC offerings.
In my humble opinion, you’re honestly better off buying a cheaper alternative and spending 20 bucks on an Arduino. Or a pizza. So that’s gonna be it for this video, I hope you guys enjoyed it, And as usual, I would like to thank my Patrons
2023-12-07 13:55
