Qualcomm Snapdragon X Dev Kit for Windows

[geerlingguy]

Basic information

• Board URL (official): https://www.arrow.com/en/products/c8380-12c-mp-32g/thundercomm
• Board purchased from: Arrow
• Board purchase date: 2024-07-16 (Arrived 2024-09-25)
• Board specs (as tested): X Elite - 12 core, 32GB RAM, 512 GB NVMe SSD
• Board price (as tested): $855.42

Linux/system information

# output of `screenfetch`
PASTE_HERE

# output of `uname -a`
PASTE_HERE

Benchmark results

CPU

• Geekbench 6: (TODO single / TODO multi - PASTE_URL)
• TODO Gflops (geerlingguy/top500-benchmark HPL result)

Power

• Idle power draw (at wall): TODO W
• Maximum simulated power draw (stress-ng --matrix 0): TODO W
• During Geekbench multicore benchmark: TODO W
• During top500 HPL benchmark: TODO W

Disk

MANUFACTURER_AND_MODEL_OF_DISK_HERE

Benchmark	Result
iozone 4K random read	TODO MB/s
iozone 4K random write	TODO MB/s
iozone 1M random read	TODO MB/s
iozone 1M random write	TODO MB/s
iozone 1M sequential read	TODO MB/s
iozone 1M sequential write	TODO MB/s

wget https://raw.githubusercontent.com/geerlingguy/pi-cluster/master/benchmarks/disk-benchmark.sh
chmod +x disk-benchmark.sh
sudo MOUNT_PATH=/ TEST_SIZE=1g ./disk-benchmark.sh

Run benchmark on any attached storage device (e.g. eMMC, microSD, NVMe, SATA) and add results under an additional heading.

Also consider running PiBenchmarks.com script.

Network

iperf3 results:

• iperf3 -c $SERVER_IP: TODO Mbps
• iperf3 -c $SERVER_IP --reverse: TODO Mbps
• iperf3 -c $SERVER_IP --bidir: TODO Mbps up, TODO Mbps down

(Be sure to test all interfaces, noting any that are non-functional.)

GPU

glmark2-es2 / glmark2-es2-wayland results:

1. Install glmark2-es2 with `sudo apt install -y glmark2-es2`
2. Run `glmark2-es2`
3. Replace this block of text with the results.

Note: This benchmark requires an active display on the device. Not all devices may be able to run glmark2-es2, so in that case, make a note and move on!

TODO: See this issue for discussion about a full suite of standardized GPU benchmarks.

Memory

tinymembench results:

Click to expand memory benchmark result

# Run the two commands below, then replace this code block with the full result.
git clone https://github.com/rojaster/tinymembench.git && cd tinymembench && make
./tinymembench

sbc-bench results

Run sbc-bench and paste a link to the results here:

wget https://raw.githubusercontent.com/ThomasKaiser/sbc-bench/master/sbc-bench.sh
sudo /bin/bash ./sbc-bench.sh -r

Phoronix Test Suite

Results from pi-general-benchmark.sh:

• pts/encode-mp3: TODO sec
• pts/x264 4K: TODO fps
• pts/x264 1080p: TODO fps
• pts/phpbench: TODO
• pts/build-linux-kernel (defconfig): TODO sec

Provisional Tests under Windows

• Idle power draw (Windows 11 Home, no apps running) - 11 W
• Sleep power draw - 4.4 W
• Power off power draw - 0-0.9 W
• Cinebench 2024 (Arm Windows) - 131 single (at 28 W) / 1227 multi (at 99 W)
• Geekbench 6.3.0 (Arm Windows) - 3020 single (at 28 W) / 15969 multi (at 80 W) (result)
• Heaven (https://benchmark.unigine.com/heaven) - TODO (at 40 W)

[geerlingguy]

A few notes as I'm unboxing:

• Box includes a power adapter and USB-C to HDMI adapter (I'm guessing the latter was added when they decided to drop the internal HDMI port).
• The bottom of the unit says "Evaluation only; not FCC approved for resale"
• According to devs on Discord, the unit comes with Windows 11 Home — and so far besides on-device partitions, there is no official way to download a recovery image or download other editions of Windows 11 built for Arm that work on it.
• The warranty states it is void if any parts are changed—and two of the phillips head screws in the base are covered by little 🚫 stickers
• Comparing the size, it's close to, but a little wider and less deep than a typical 1L 'square' mini PC. It's much larger (but shorter) than my GMKTek box (NUC-sized), and it's nearly the same width (just a hair wider) and a little deeper than the Windows Dev Kit (the one from 2023 / Project Volterra)

[MagicAndre1981]

there is no official way to download a recovery image or download other editions of Windows 11 built for Arm that work on it.

upgrading to a higher Windows SKU should work via entering a product key in settings app.

[geerlingguy]

Some hardware details:

• Ethernet port is on a separate daughtercard marked "Running HDMI Board" with an unpopulated HDMI port, and a display connector that is not connected to anything.
• Ethernet provided by RTL8125BG (2.5 Gbps Realtek NIC) on mPCIe card
• WiFi is a "T99H432.10" module, presumably Qualcomm though I can't find documentation on it. I have found one laptop, the MSI Alpha 17 C7V, which has the same model number (but .03 instead of .10), and it is listed as "WiFi 7". The WiFi module is E-key, and has a little heatsink mounted to it (with thermal pads on top and bottom).
• There's an unpopulated PCIe slot (no slot is installed, but the pads are on the board. I would really like that to be populated. Looks like full size x16, not sure if electrically connected as x16, x8, x4, x1, or not at all...
• There's a 2280 size NVMe SSD, a Foresee XP2200F512G, made by Longsys. It is mounted with a full-length heatsink (with thermal pads on top and bottom).
• There's a tiny daughtercard called "Running ECB" - does it have something to do with cryptography? Platform security? Not sure, it's mounted above a NanoSIM and microSD slot. The plug it is plugged into is labeled "EC_I/F"
• That's right, there's a hidden NanoSIM slot (with nothing in it), next to the microSD card slot that is exposed through the front of the case.
• The cooler is a massive copper heatsink with heat pipes and a massive cooling stack for the SoC itself. There are thermal pads on all the little power circuits, and a large amount of thermal compound on all four RAM chips and the SoC itself.
• The fan is APEXE model B82DBHA2467, with a 4-pin standard fan header on the motherboard (easy to remove, separate from the copper heatsink). It is connected to the heatsink with screws, and there is a taped on duct to port the exhaust out the side of the Dev Kit.
• There are a couple CSI ports, many unpopulated headers, an unpopulated eMMC spot near the RAM, and there's a micro USB debug port labeled DBG_UART

[geerlingguy]

Blog post with many photos and more detailed notes after zooming way way in: Qualcomm Snapdragon Dev Kit for Windows Teardown (2024).

[utp216]

• Box includes a power adapter and USB-C to HDMI adapter (I'm guessing the latter was added when they decided to drop the internal HDMI port).

I wonder if you solder the appropriate HDMI port to the PCB if it would function. Maybe they didn't want to pay for HDMI licensing?

[apache405]

The ECB is an embedded controller board. ITE is one of the more popular EC chips out there.

[jglathe]

Arrow has just canceled my order for this one - I ordered on the same day 🤨. Thought they would ship none, then, but... whatever. The hardware seems nice enough, though. Do you want to run Linux on the thing? That was my intent. You also can run the Windows Dev Kit 2023 with Linux.

[SpieringsAE]

• Box includes a power adapter and USB-C to HDMI adapter (I'm guessing the latter was added when they decided to drop the internal HDMI port).

I wonder if you solder the appropriate HDMI port to the PCB if it would function. Maybe they didn't want to pay for HDMI licensing?

doesn't the DP to HDMI converter get around that? as in the licensing cost is already part of the converter chip?

But yeah I wondered the same thing, maybe even make a different daughterboard that just exposes DP directly.

Same with the unpopulated PCIE slot, wonder how far populating that gets, on windows it might be locked out of the ACPI tables, but on Linux you might be able to just add it to the devicetree blob

[talynone]

Richard Campbell with hardware design experience speculated HDMI FCC approval was the issue, more info:

https://www.youtube.com/watch?v=NlEV38cO8Gs

[talynone]

Getting the Windows ARM ISO is doable but requires a bit more effort right now:

https://www.makeuseof.com/download-install-windows-11-arm-iso/

Of course make sure you have a backup or source for all the system specific drivers before wiping the existing installation.

[anonymix007]

T99H432.10

I believe the actual model is NCM865A. NCM865 is used in some WiFi 7 products and AFAIK the only difference between them is the Bluetooth connection interface (USB vs UART).

[SpieringsAE]

Richard Campbell with hardware design experience speculated HDMI FCC approval was the issue, more info:

https://www.youtube.com/watch?v=NlEV38cO8Gs

tbh, why even hdmi at all, DP is right there, on the board, why put extra chips on a board to make the display interface worse (arguably).

[theholypumpkin]

I'm very excited about the Linux benchmarks!!!

I'm waiting for Tuxedo Computers with their Snapdragon X Model to be released maybe at the end of 2024 and Wondering what performance I can expect in comparison to my Late 2013 MacBook Pro.

Now with kernel 6.11 out, the Snapdragon X chips should at work, but looking at the release notes looks like it's still not fully baked. Probably by 6.12.

[lukasMega]

what is the power consumption in common use cases (idle in headless mode, idle with desktop, load, ...) Thank you!

[talynone]

There's not many disk imaging/cloning apps that work on Windows ARM right now, this procedure seemed to have worked for others (read the comments for gotchyas):

https://old.reddit.com/r/Surface/comments/1dniufo/ssd_clonemigrate_on_arm/

[talynone]

Can't seem to find an Windows ARM native build of iperf3, but x86/x64 version should work fully without issue in emulation.

In general the following is a great resource for finding native Windows ARM apps (well maintained/updated):

https://armrepo.ver.lt/

[driver1998]

There's not many disk imaging/cloning apps that work on Windows ARM right now

WinPE on ARM64 supports x64 emulation, and some disk partitioning/imaging software works fine under emulation. I tried DiskGenius and it works fine. dd for Windows might work as well.

[jglathe]

MiniTool ShadowMaker works on arm for an image. restore... better an x86 box, but works reliably.

[geerlingguy]

Regarding the HDMI port, Richard Campbell speculates that port could be at least a significant reason for the shipping delay, if they weren't able to get it cleared for compliance (maybe also the reason the FCC clearance didn't happen).

[geerlingguy]

Regarding Windows 11 image creation—I have created aarch64 images using UUPdump.net in the past, and that works fine, it's just... not official in any way, shape, or form.

What I would really like to see is Microsoft providing an arm64 ISO like they do for x86 on their Windows 11 Download page.

[talynone]

Little off topic, but a bigger reason some smaller open source projects don't have WIndows ARM64 builds even more than the availability of the Dev Kit is because GitHub (which is owned by Microsoft) Windows ARM64 runners are currently only available for enterprise customers. GitHub has said it should open it up to everyone soon but if Microsoft wants faster adoption they really should light a fire on that.

[geerlingguy]

Notes from my first boot:

• The Ethernet LEDs don't seem to light. Don't see an issue with the cabling, it's like the NIC EEPROM doesn't have the LEDs lighting (connected in Windows fine at 2.5 Gbps).
• Windows 11 Home is silly but at least the account setup bypass still works.
• After setting up an account, I think there were around 6 or 7 total reboots for Windows Updates, plus a firmware update that made it so my DVI-input monitor no longer worked with the Dev Kit (had to switch to a newer HDMI monitor).
• The fan is a lot louder than I expected. It ramps down to almost silent at idle (thankfully), but the fan ramps up aggressively any time I do anything (even opening an app, or for background activity). The lowest speed is still louder than my Mac Studio maxed out.

[geerlingguy]

Single core gets up to around 30W, multi-core Cinebench 2024 gets up to 100+ W, and the fans ramp up to about 58 dBa measured 1' away. Definitely a loud little machine.

Want to get Linux going so I can do some real testing ;)

[SpieringsAE]

Do you have a link to the geekbench test results? Would love to see results of the independent tests

[geerlingguy]

I'm updating all the final results in the OP at the top of the issue, and just added a link to the Geekbench result there.

[talynone]

Heaven has no native ARM64 tests, so that may not give the best results.

Solar Bay, Wild Life: and Wild Life: Extreme in 3DMark have Windows ARM64 native versions.

[geerlingguy]

@talynone - I understand that — just wanting to get a picture of how well x86 stuff is emulated and runs on this platform ;)

[talynone]

Recommend making sure you have the OpenCL compatibility pack from the Microsoft Store installed .

https://apps.microsoft.com/detail/9nqpsl29bfff?hl=en-US&gl=US

You can also now download/install latest ARM64 Vulkan runtimes and SDKs from:

https://vulkan.lunarg.com/sdk/home#windows

[geerlingguy]

Was able to install the Windows Subsystem for Linux with Ubuntu 24.04.

1. Run PowerShell as Administrator
2. Run wsl --install
3. After it completes, reboot
4. Run wsl --list --online to show available Distros, select one to install
5. (I ran wsl.exe --install -d Ubuntu-24.04)
6. WSL dumps you into the Ubuntu environment once it's done

To quit, type exit. To get back in, type wsl.

[jjlauer]

I realize most of the discussion is around getting linux to run, but I did have a question regarding the stock Windows install. I upgraded the Snapdragon X Dev Kit to Windows Pro by adding a license key in order to get Hyper-V running. I am able to create Windows arm64 VMs that run great (by simply using a Windows .arm64 .iso), but Linux VMs on Hyper-V are awful. Here are my questions:

1. Do Linux VMs run better on non Dev Kit hardware such as the various laptops that also ship with the Snapdragon X SOCs?
2. I can only get something like Ubuntu 24.04/24.10 arm64 .isos to install on Hyper-V if i limit the vCPUs to 1. If I try for anything more than 1 vCPU, the install/boot process hangs immediately after grub. If I stick to 1 vCPU, it does install successfully and boots, but the performance is terrible. Is there a trick or a better distro that runs well on Hyper-V? It's confusing since WSL runs perfect. My guess is the WSL kernel has drivers/enhancements vs. stock Ubuntu arm64 that allow it to run great on this SOC.

[jglathe]

No idea why Hyper-V has an issue with non-WSL linux. I've heard this from several people, it's the same issue on the laptops.

[Mis012]

If you mean graphical performance, then I would assume you will get sw rendering by default. Speaking of which, running a Linux VM on Windows will presumably use the windows drivers in any case, and I'd be very surprised if the windows drivers had better performance than freedreno/turnip from what I've heard.

[jjlauer]

@Mis012 No, I should have clarified that its not graphical related. I'm just installing server version of Ubuntu. The CPU performance of just the 1 vCPU is slow/awful. Other odd things as well like compiling code, network performance, pings -- they all work poorly too. If you run Geekbench 6, they are orders of magnitude worse than both the host Windows OS or even a guest Windows OS instance in Hyper-V. None of this happens on WSL -- which runs as well as the host Windows OS for all those things mentioned above.

[ankushnarula]

I haven't tried this yet, but this project claims to have generated a nixOS derivation from inside WSL2 which is then used to generate a flawless bootable ISO installer of nixOS for Hyper-V. The README makes some interesting claims:

A NixOS derivate that generates a bootable iso for Snapdragon X Elite's Hyper V.

No existing linux .iso I am aware of work properly on Hyper V. A standard NixOS image may boot, but runs very slowly, taking hours to boot the kernel.

I noticed WSL2, which runs on Hyper V is quite fast. So I created an image with the Microsoft WSL 2 Kernel. I tried using Microsoft's kernel config, as well as letting Nix generate the config. Both work.

[jjlauer]

@ankushnarula I just tried the NixOS iso that was a release on that repo and it worked! Installed it with 4 vcpus, installer came up quickly. So far it feels snappy. Seems like other distros could be built using the same technique, by simply using the WSL2 kernel.

[jjlauer]

[@ankushnarula]() I installed stock Ubuntu 24.04 arm iso (using 1 vcpu), compiled the WSL kernel, and voila Ubuntu is running really well in early tests. I was able to also use up all the vCPUs and assign all 12 of them and ubuntu booted and seems to be running well. Perhaps folks on this thread that are closer to building kernels can figure out if its simply the config & drivers that are the difference or if the WSL kernel changes something else important. Either way, seems like a custom distro would be an elegant solution to getting linux VMs to run well on these qualcomm socs.

[talynone]

Rumored Qualcomm mini pc coming from Geekom:

https://www.windowscentral.com/hardware/computers-desktops/a-snapdragon-x-elite-mini-pc-is-on-the-way-but-its-not-from-qualcomm

[jglathe]

Update on booting the SnapDragon DevKit with display: first signs of life.

The display is connected with the type-c to HDMI adapter shipped with it, and it works, if... the system comes up.

Requirements:

• device tree with USB4 retimers defined and wired in
• needs the adsp firmware loaded for dp altmode operation
• sufficiently new dp altmode support
• device tree should also have the USB type-a ports enabled with its repeaters.

This sounds easy enough, but it will fail (it did for me) when you try to boot from type-c, like we did before 'cause the type-a ports weren't operational either. This is because the adsp firmware, despite being available in the initramfs (only), will be loaded after mounting the rootfs (which is on type-c), causing a VBUS interruption, leading to re-detection of the type-c drive, rendering the already mounted rootfs unusable. The way to avoid this is to use type-a USB stick/drive or internal SSD. My current devkit dt supports type-a, this is how it came up.
A (sort of) major hurdle is that you need the adsp firmware accessible for dp altmode to work and get you a display. This won't be on the image to boot it. The dedicated wdk23 images had a trick to boot first, fetch the firmwares from the local SSD, and reboot to proceed with installation. I guess we need something similar here.
For those who want to try themselves, this is the branch and the built kernel package.

[jglathe]

Finally GUI on the DevKit. You need Ubuntu 24.10 and its newer gnome/mesa to get it up.

Graphics:
  Device-1: adreno-43050c01 driver: adreno v: N/A
  Device-2: x1e80100-mdss driver: msm_mdss v: N/A
  Device-3: x1e80100-dpu driver: msm_dpu v: N/A
  Display: server: X.org v: 1.21.1.13 with: Xwayland v: 24.1.2 driver: X: loaded: modesetting
    unloaded: fbdev dri: msm
    gpu: adreno,hdmi_msm,mdp4,msm,msm-dp-display,msm-mdss,msm_dpu,msm_dsi,msm_dsi_phy,msm_hdmi_phy,msm_mdp
    tty: 132x50 resolution: 1920x1080
  API: EGL v: 1.5 drivers: msm,swrast platforms: gbm,surfaceless,device
  API: OpenGL v: 4.6 compat-v: 4.5 vendor: mesa v: 24.2.3-1ubuntu1 note: console (EGL sourced)
    renderer: Adreno X1-85, llvmpipe (LLVM 19.1.0 128 bits)

Also, this is still an endeavour. If the display comes up depends on (unknown). Re-plugging the type-c to HDMI adapter should bring it up, tough. Different issue if you have 2 screens (with a similar adapter) - will not work reliably, or at all. Either screen alone works.
As for replicating this... I'm working on an image to be used which should give you a pre-installed desktop with ubiquity setup. It might take a few more days, I'm afraid.

The actual image to boot, fetch the firmwares and reboot is ready. I'm currently doing an in-place upgrade to 24.10 in the chrooted environment, if it works well this image can be published. Otherwise, I need to assemble one from the RPi Desktop Image for Ubuntu 24.10.

All in all, it is as I expected: Having only DP Altmode as display leaves you with a lot of dependencies to meet to make it work. Making this work from a generic install image is nearly impossible without breaking some rules.

[Mis012]

I'm surprised nobody tried to plug in an AMD GPU yet, the x16 slot may be missing but m.2 to x4 adapters exist

[jglathe]

Still busy with bringup, tbh. Also, the integrated GPU is not bad, there are new performance patches on the mailing list.

[Mis012]

I suppose even the laptops have m.2 slots, but with @geerlingguy iirc saying at some point that people claimed to run GPUs with no driver changes on HDKs I'd have expected someone here to try to connect a GPU as soon as it's booting Linux.

[geerlingguy]

@Mis012 - I don't recall anyone getting an eGPU (e.g. AMD, Nvidia, or Intel) to work on any Qualcomm hardware... I've gotten them working on Raspberry Pi, Ampere Altra/Altra Max, AmpereOne, and to some extent on a few other Arm chips. But not Qualcomm (at least not personally).

[Civil]

I'm surprised nobody tried to plug in an AMD GPU yet, the x16 slot may be missing but m.2 to x4 adapters exist

I right now don't have time to play with the devkit, so I haven't tried latest kernel. I also don't want to lose NVMe drive to be honest, but devkit also have thunderbolt, which should be possible to use for external GPU at some point (but I'm not even sure how Linux would work with GPU on Thunderbolt).

[Mis012]

@geerlingguy well, I was also surprised when I (thought?) I heard that, since idk that the BAR space used to be particularly large, but it would be third hand information in any case since I'm pretty sure you don't have any HDK personally.

[Mis012]

re: thunderbolt/usb4, in general it works fine on Linux afaik but I don't see any qcom-specific drivers anywhere, not even downstream... so I wouldn't expect it to work here yet.

For testing could run off of USB, I wouldn't necessarily expect anyone to find it so amazing to use that they'd be willing to daily drive a GPU connected to m.2 anyway but assuming the test result is "I need this permanently", the option to make the x16 slot work is still there (arguably would be easier with schematics/boardfiles for the cdp, but if you care enough can probably bribe someone in china since I think qcom gives those to all partners)

[Civil]

I don't think schematics is required to put back x16 slot. If it lqck passive components - they should be standardized and listed in specs for pcie (that is arguably easier to get than schematics of crd), but problem is that it would risk the devkit.

And same with populating back HDMI - it risks damaging PCB that would be hard to replace.

For pcie - if it's just the slot missing, I can probably put it back myself, but my skills are not good enough nto put back passive components.

[Mis012]

passives would just be caps presumably, I'd be more worried that there is supposed to be a redriver :P

repopulating HDMI is imho completely unnecessary, just need to get an appropriate breakout board for whatever ffc pin amount / spacing that connector uses and wire it up to a DisplayPort.. ..port :P Then you can use any ready-made active converter if you still want HDMI

[jglathe]

For the record, I have the dev kit booted from nvme with Ubuntu 24.10. Connected to a 2k display @75 with a type-c to DP cable. Seems pretty stable... DP Altmode is getting better with every rc, so it might be an alternative to hardware mods. Booting from type-c gets you in a timing range where the previous hacks to avoid issues with the rootfs on type-c no longer work. Therefore it's either USB-A (working fine) or internal nvme. For DP Altmode to work you need the adsp firmware loaded. Otherwise no display, which makes it a little hard for an installer. I guess I will finish the setup of the new image by the weekend.

[Mis012]

I wouldn't call it hw modding, the HDMI board is not connected anyway :P
something like this would be needed: https://hackaday.io/project/174315-direct-digital-negative-uv-printer-mkiii/log/183264-connecting-lcd-to-displayport

[Civil]

the HDMI board is not connected anyway :P

AFAIR it is still connected, at least I saw second cable that looked like eDP one going to it (that is in addition to network cable from the NIC)

[jglathe]

It is usable now. Kernel is 6.13-rc6, reasonably recent.
Bringing up the SnapDragon Dev Kit for Windows with Linux ‐ with working display

[jglathe]

The sdhc slot works too, btw. It is sort of hard-wired to be the first boot option, which is useful. It only does boot from it with an ESP and some efi bootable executable. What doesn't work yet (doesn't work from USB-A either) is to boot Linux on the nvme when grub is on sdhc or USB-A.
There seems to be a limitation on the type of sd cards to use: Only V60 or V90 UHS-II seems to work (at least on Linux). I guess there is some fine tuning to be done, data rates suggest it is using UHS-I mode, but UHS-I cards are not working.

[jglathe]

Interesting development: Chatting with someone who has a populated PCIE Slot on this, and also HDMI. jglathe/linux_ms_dev_kit#30 I'm trying to enable the software side (dtb), not looking too bad.

[Civil]

Nice! Now I have more reasons to risk soldering pcie slot myself :) I'm not that confident in my skills but if it's only slot missing - I can try.

[allokey]

Nice! Now I have more reasons to risk soldering pcie slot myself :) I'm not that confident in my skills but if it's only slot missing - I can try.

Let's give it a try, but be careful with the soldering temperature. That is a PCIe GEN4 x8 interface.

[BatMahn]

Perhaps not the best place to ask, but does anyone happen to have an image of the OEM SSD they wouldn't mind sharing?

I erased Windows and tossed that Ubuntu image on. Worked well for a few weeks, but now it won't boot anymore.

[jglathe]

@BatMahn send me a DM on matrix or so