Khadas Mind 2 AI Maker Kit Review – Part 3: Ubuntu 24.10 and the importance of power limits

I’ve already reviewed the Khadas Mind 2 AI Maker Kit with Windows 11 Home, and today, I’ll report my experience with Linux on the Intel Core Ultra 7 258V mini PC using Ubuntu 24.10 operating system.

I would usually review systems with Ubuntu 24.04 LTS OS, but considering the Intel Core Ultra 7 258V SoC is so new, I installed Ubuntu 24.10 when I tested whether disabling VT-d (IOMMU) would improve Intel Arc GPU performance (it does to some extent), and it turns out it was a good decision because Ubuntu 24.04 requires lots of fixes and workarounds to work the Khadas Mind 2 AI Maker Kit, at least until Ubuntu 24.04.2 is released later this month with a more recent kernel.

Khadas Mind 2 AI Maker Kit Ubuntu 24.10 Review

Khadas Mind 2 AI Maker Kit – Ubuntu 24.10 system information

My Ubuntu 24.10 installed has both Linux 6.11 (default) and Linux 6.13 kernels, but I did most of the tests using Linux 6.11 since it’s what comes with Ubuntu 24.10. The About window confirms we have a KHADAS Mind 2 AI Maker Kit running Ubuntu 24.10 on an octa-core Intel Core Ultra 7 258V processor with 32GB RAM and 1.0 TB of storage.

Khadas Mind 2 AI Maker Kit Ubuntu 24.10 About

We can get a few more details in a terminal window:


And even more with inxi utility:


Everything is there, except WiFi which just shows as “Device-1: Intel driver: N/A”. That’s why I’ll be using a 100Mbps Ethernet USB adapter before fixing WiFI on the Intel Core Ultra 7 258V mini PC.

Ubuntu 24.04 benchmarks on Khadas Mind 2 AI Maker Kit

As usual, I’ll start with Thomas Kaiser’s sbc-bench.sh script:


The Khadas Mind 2 AI Maker Kit may not be able to extra the full performance of the Intel Core Ultra 7 258V under load, as while the CPU temperature never gets over 69.0°C (in the single-core OpenSSL benchmark), the CPU frequency drops sharply in multi-core benchmarks (4700 MHz to 2200 MHz) and the temperature is only up to 59.0°C in that test pointing so suboptimal settings.

The script tells us to check the thermal limits, so let’s do that:


The PL1 and PL2 power limits are set to 200W!? (PBP) and 37W (MTP), while PL4 (Peak power) is not listed at all… For reference, we had PL1, PL2, and PL4 set to 30W, 37W, and 95W respectively in Windows 11 Home. So something is odd here. I rebooted to Linux 6.13-rc1 to check again, but the results are the same. Back to Linux 6.11, I decided to change the power limits for Zone 1 (PSYS) to match the limits on Windows 11:


Let’s run sbc-bench.sh again:


The 7-zip score is now much better at 31,480 MIPS on average, so there’s some work to be done on the BIOS or Linux kernel side to fix the default values… As a side note, the Orion O6 motherboard with a 12-core Armv9 processor achieved a similar score (31,060 MIPS) in the same 7-zip benchmarks. Arm has come a long way.

Let’s carry on testing with Geekbench 6.3.0 single-core and multi-core benchmarks.

GeekBench 6.3.0 Linux KHADAS Mind 2 AI Maker Kit

That would be 2829 points for the single-core score and 7014 points for the multi-core score before changing the power limits. But what about after?

Geekbench 6.3.0 Linux Mind 2 AI Maker Kit PL1 30W PL2 37W

Similar single-core score (2882) and a much better multi-core score (9618).

I’ve also run Geekbench AI 1.2.0. to check whether the Intel AI Boost accelerator is supported, but by default, the benchmark relies on the TensorFlow Lite framework and runs on the CPU…

GeekBench AI 1.2.0 Linux Intel Core Ultra 7 258V CPU

You’ll find the full results on the Geekbench website, even though they are not super interesting since it’s CPU only, and we’d like it to run on the GPU and/or NPU.  So let’s check the options available in Linux:


So while we could run Geekbench AI on the CPU, GPU, and NPU on Windows 11, the program only supports CPU processing on Linux, and it does not matter whether I try with Linux 6.13… So I tried to install the Intel AI Boost drivers following the instructions on GitHub:


I did get a few more kernel log messages related to intel_vpu after that:


The NPU device is also there:


But I still don’t get an NPU option in Geekbench AI:


I’ll stop for now, and I may revisit this once I test AI on the Intel Core Ultra 7 258V in more detail.

We’ve already tested the GPU performance with Unigine Heaven Benchmark 4.0 and the Intel Core Ultra 7 258V mini PC last time around when we disabled VT-d and tried Linux 6.11 and Linux 6.13 kernel. Results varied between 42 FPS and 48.2 FPS with the best result achieved with VT-d/IOMMU disabled and Linux 6.13-RC1, but note that disabling VT-d/IOMMU has some security implications. I’ll use the score for Linux 6.11 and VT-d enabled since those are the default values on Ubuntu 24.10: 1,057 points / 42 FPS on average at the standard 1920×1080 resolution. For reference, the score was 2045 points / 81.2 FPS on Windows 11 with Direct3D11.

Intel Arc GPU Unigine Heaven Benchmark 4.0 VT d enabled

The test above was made with the default power limits. So I repeated the benchmark with PSYS PL1 = 30 W, PL2 = 37W, Linux 6.11, and IOMMU/VT-d enabled.

Unigine Heaven Benchmark 4.0 Linux Intel Core Ultra 7 258V adjusted PL1 PL2

We’re getting closer to the optimal performance with 67.4 FPS and a score of 1698 points.

I also tested several 4K or 8K YouTube videos in Firefox and the results were rather disappointing (with default power limits). Note that I had enabled WiFi 6 at that point (See below how I did that) since the 100 Mbps Ethernet USB dongle might have been a bottleneck otherwise.

Khadas Mind 2 AI Maker Kit Youtube 8K 60 Firefox Ubuntu
8K 60 FPS – Firefox (AV1 codec)

I was a little too overconfident since 8K 60 FPS YouTube videos would play perfectly in Windows 11, and I started directly with an 8K 60 FPS video. But I had to quickly stop because the video was unwatchable and “Stats for Nerds” reported 1062 frames dropped out of 1589.

Khadas Mind 2 AI Maker Kit Youtube 4K 60 Firefox Ubuntu
4K 60 FPS – Firefox (VP9 codec)

4K 60 FPS plays better, but it’s not perfect with 723 frames dropped out of 27640 when playing the video for a little over 6 minutes.

Khadas Mind 2 AI Maker Kit Youtube 4K 30 Firefox Ubuntu
4K 30 FPS – Firefox (VP9 codec)

No problem for 4K 30 FPS.

Khadas Mind 2 AI Maker Kit Youtube 8K 30 Firefox Ubuntu
8K 30 FPS – Firefox (AV1 codec)

But when I switched to 8K 30 FPS it was also unwatchable. It looks like the system can’t handle 8K AV1 video decoding with the GPU.

I installed Google Chrome and tried again the 8K 60 FPS video, but the video was still unwatchable and I could see the loading logo show up often despite the “buffer health” being just fine.

Khadas Mind 2 AI Maker Kit Youtube 8K 60 Chrome Ubuntu
8K 60 FPS – Chrome (AV1 codec)

What happens when I change the PSYS power limits to 30W and 37W?

Khadas Mind 2 AI Maker Kit Ubuntu 24.10 YouTube adjusted power limits

Still not watchable with some stuttering from time to time, but it’s a big improvement with much fewer frames dropped, and I can imagine that with further optimizations in the future the video may play as smoothly as in Windows 11…

I finally ran Speedometer 2.0 in Firefox to test web browser performance. That’s mostly a single-core benchmark.

Khadas Mind 2 AI Maker Kit Speedometer 2.0 Firefox Linux

That would be 295 runs per minute, a pretty good/expected score for a high-end mini PC.

Since Speedometer 2.0 is deprecated, I’ve also run Speedometer 3.0 to get data for future reviews.

Khadas Mind 2 AI Maker Kit Speedometer 3.0 Firefox Linux

The Khadas Mind 2 AI Maker Kit achieved 18.4 points in Speedometer 3.0 in Firefox.

Khadas Mind 2 AI Maker Kit Ubuntu benchmarks comparison against other mini PCs

Let’s compare the Ubuntu 24.10 benchmark results of the Khadas Mind 2 AI Maker Kit against the Raptor Lake-powered Khadas Mind Premium mini PC (Intel Core i7-1360p) and other high-end mini PCs namely the GEEKOM GT1 Mega (Intel Core Ultra 9 185H),  GEEKOM A8 (AMD Ryzen 9 8945HS), and GEEKOM Mini IT13 (Intel Core i9-13900H) which were tested with Ubuntu 22.04, 24.04, or 24.10.

Let’s list the basic specifications of the five systems under test first.

Khadas Mind Maker KitKhadas Mind PremiumGEEKOM GT1 MegaGEEKOM A8GEEKOM IT13
SoCIntel Core Ultra 7 285VIntel Core i7-1360PIntel Core Ultra 9 185HAMD Ryzen 9 8945HSIntel Core i9-13900H
CPU8-core (4P+4E) Lunar Lake processor up to 4.8 GHz (P-Cores) and 3.7 GHz (E-Cores)12-core (4P+8E) Raptor Lake processor up to 5.0 GHz (P-Cores), up to 3.7 GHz (E-Cores)16-core/22-thread (6P+8E+2LP) Meteor Lake processor up to 5.1 GHz (P-cores), up to 3.8 GHz (E-cores), up to 2.5 GHz (LP-cores) 8-core/16-thread processor up to 5.2 GHz 14-core/20-thread (6P+8E) Raptor Lake processor up to 5.40 GHz (P-cores), up to 4.10 GHz (E-Cores)
GPU8 Xe cores Intel Arc 140V GPU96 EU Intel Iris Xe Graphics up to 1.5 GHz8 Xe cores Intel Arc GraphicsAMD Radeon 780M Graphics96 EU Intel Iris Xe Graphics
NPUIntel AI Boost (47 TOPS)N/AIntel AI Boost (34 TOPS)Ryzen AI (16 TOPS)N/A
Memory32GB LPDDR5-853332GB LPDDR5-520032GB DDR5-560032GB DDR5-560032GB DDR4-3200
Storage1TB NVMe SSD1TB NVMe SSD2TB NVMe SSD2TB NVMe SSD2TB NVMe SSD
Default OSWindows 11 Home Windows 11 HomeWindows 11 ProWindows 11 ProWindows 11 Pro

Here are the benchmark results.

Khadas Mind 2 AI Maker KitKhadas Mind PremiumGEEKOM GT1 MegaGEEKOM A8GEEKOM Mini IT13
sbc-bench.sh
- memcpy25,504.6 MB/s (P-Core)25,389.5 MB/s (P-core)21,364.6 MB/s (P-core)20,318.5 MB/s24,014.4 MB/s (P-core)
- memset65157.0 MB/s (P-Core)24,731.8MB/s (P-core)36,928.3 MB/s (P-Core)62,156.7 MB/s26,647.9 MB/s (P-Core)
- 7-zip (average)19,980
31,480 (adjusted PL)
44,43067,96068,79056,540
- 7-zip (top result)22,093
32,768 (adjusted PL)
50,39671,62369,29760,981
- OpenSSL AES-256 16K1,665,362.60k (P-Core)1,771,334.31k (P-Core)1,698,239.83k (P-Core)1,422,136.66k1,844,401.49k (P-Core)
Geekbench 6 Single2,82920932,6052,6612,745
Geekbench 6 Multi7,014
9,618 (adjusted PL)
8,89113,72813,27511,974
Unigine Heaven score1,057
1,698 (adjusted PL)
1,3491,9561,9721,333
Speedometer 2.0 (Firefox)295242
278298273

The good news is that the Khadas Mind 2 AI Maker Kit has the highest memory bandwidth of all mini PC we’ve tested, and its single-core performance is excellent as demonstrated with Geekbench 6 single-core and Speedometer 2.0 benchmarks. But GPU performance is much lower than in Windows 11 and compared to the other high-end mini PCs we’ve tested in Linux. Multi-core performance is also quite lower both because it only has eight cores, but also due to thermal throttling with default power limits. It was only when I changed the power limits to the same as in Windows 11 (PL1 = 30W, PL2 = 37W) that the scores (shown as adjusted PL in the table above) were closer to expectations. Khadas engineers will have some work to do once they are back from Chinese New Year’s holidays…

Storage and USB testing

Let’s check the SSD performance with iozone3:


Sequential read and write speeds are about 4615 MB/s and 4797 MB/s respectively, both excellent as expected.  For reference, the computer achieved 5267 MB/s sequential reads and 4939.70 MB/s sequential writes with CrystralDiskMark on Windows 11.

I also tested the USB ports to find out if we could get the advertised speed using lsusb (USB 3.2) or boltctl (USB4/Thunderbolt) utility as well as iozone3 to measure the speed using an ORICO M234C3-U4 M.2 NVMe SSD enclosure.

First USB Type-A (#1) port from the left:


Left USB Type-C port:


Here’s a summary for all four ports from left to right:

  • USB-C #1 – USB4 – 40 Gbps – Read speed: 2942 MB/s; write speed: 2171 MB/s
  • USB-C #2 – USB4 – 40 Gbps – Read speed: 2968 MB/s; write speed: 2186 MB/s
  • USB-A #1 – USB 3.2 – 10 Gbps – Read speed: 993 MB/s; write speed: 984 MB/s
  • USB-A #2 – USB 3.2 – 10 Gbps – Read speed: 994 MB/s; write speed: 992 MB/s

All good. No problem here.

WiFi 6 performance

Since the Khadas Mind 2 AI Maker Kit does not come with Ethernet, we’ve only tested WiFi 6 performance using Xiaomi Mi AX6000 router and UP Xtreme i11 Edge mini PC on the other end connected through 2.5GbE. There’s a small problem though the AX211 WiFi is detected as a BE201 module and unsurpringly loading the firmware does not work meaning we don’t have WiFi:


Khadas provides a solution in the link we included in the introduction. In Linux 6.13-RC1, I only had to install the new firmware:


But with Linux 6.11, I had to install the new firmware as above and compile the backported wifi driver:


The firmware is now loaded and iwlwifi driver is listed in inxi:


I could connect to 5 GHz WiFi 6…

Khadas Mind 2 AI Maker Kit Ubuntu Linux WiFi 6

…before running iperf3:

  • Download

  • Upload:


1.35 Gbps download speed and 1.57 Gbps upload speed are pretty good, albeit not quite as fast as in Windows (1.78 Gbps / 1.93 Gbps) in the same test environment.

Khadas Mind 2 AI Maker Kit stress test in Ubuntu 24.10

We’ve seen the CPU temperature never exceeds 60°C with the default power limits in Linux, so it’s pretty clear thermal throttling would not happen under those conditions. So I run a stress test on all eight cores of the Intel Core Ultra 7258V processor after setting PL1 to 30W and PL2 to 37W in PSYS Zone (1). I monitored CPU temperature and frequency with Psensor and the sbc-bench.sh script.

Khadas Mind 2 AI Maker Kit Stress Test Ubuntu 24.10

The temperature maxes out at 70°C and the frequency stabilizes at 3300 MHz on the P-cores and 3701 MHz on the E-cores. There’s still some legroom at least for CPU-only workloads even after adjusting the power limits.

Fan noise

Khadas Mind AI Maker Kit is cooled with a fan that’s inaudible at idle and light workloads but becomes noisier under stress, although I don’t find it annoying. I measured the fan noise with a sound level meter placed at around 5 centimeters from the top of the device:

  • Idle – 37.9 – 38.3 dBA (low-speed fan)
  • YouTube 8K 60FPS video in Firefox (volume off, adjusted power limits) –  46.9 – 47.8 dBA
  • Stress test on all 8 cores (adjusted power limited) –  48.8 – 49.4 dBA

For reference, the meter measures around 37.2 dBA in a quiet room.

Mind 2 AI Maker Kit power consumption in Ubuntu 24.10

After making sure the internal battery is fully charged, I measured the power consumption with a wall power meter:

  • Power off – 1.6 Watt
  • Idle – 3.9 – 4.2 Watts
  • Video playback – 32.6 – 35.9 Watts (Youtube 8K 60FPS in Firefox – AV1 codec)
  • CPU stress test (stress -c 8)
    • First ~30 seconds – 40.8 – 40.9 Watts
    • Longer run – 33.5 – 33.6 Watts

During the measurements, the mini PC was connected to WiFi 6, an HDMI display with its own power adapter, and a USB RF dongle for a wireless keyboard and mouse combo.

Conclusion

If the Khadas Mind 2 AI Maker Kit was a mini PC for end users, I would have no hesitation telling people to stay away if they wanted to run Ubuntu 24.10 or other Linux distributions on the mini PC. That’s because it’s a complete mess and workarounds are needed for it to work properly. But considering it’s made for developers, once you install the WiFi driver and firmware, Intel AI Boost NPU drivers, and adjust the power limits, it works relatively well albeit I was unable to reach Windows 11’s performance levels in tasks like 8K 60 FPS YouTube video playback and 3D graphics, meaning more work is needed. Geekbench AI is unable to use the GPU or NPU on Linux, while it can on Windows 11.

On a positive note, the processor offers the fastest single-core performance of the many mini PCs we’ve tested, the WD SSD is super fast, and so is the AX211 WiFI 6 module once you’ve installed the drivers. All features work as expected after tweaks, except the battery which I have trouble understanding. Sometimes I have no problems waking up the mini PC, and other times I have to press the power button to shut it off before starting it again. Its eight cores may be limiting depending on what you want to do, but its up to 115 TOPS of AI performance makes it a good machine for AI development, although I have not tested the OpenVino SDK just yet.

I’d like to thank Khadas for sending the Mind 2 AI Maker Kit for review, also referred to as the Mind Maker Kit. The company sells the AI developer kit on its website for $1,199 plus shipping. I plan to test AI features on Windows and Ubuntu in the final part of the review once I have some answer from Khadas since some Copilot+ features are not enabled for some reason…

Share this:

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress

Radxa Orion O6 Armv9 mini-ITX motherboard
Subscribe
Notify of
guest
The comment form collects your name, email and content to allow us keep track of the comments placed on the website. Please read and accept our website Terms and Privacy Policy to post a comment.
5 Comments
oldest
newest
back2future
back2future
15 days ago

[ was interested what are specification details for the battery: “Mind Maker Kit is equipped with a built-in 5.55Wh battery to ensure an uninterrupted workflow. When the power is disconnected, the device instantly enters sleep mode, preserving your progress without any interruptions. With up to 25 hours of standby time and a 2-second wake-up feature, you can seamlessly resume your desktop projects exactly where you left off. This ensures smooth transitions between tasks and locations, enhancing your productivity and development efficiency.” and astonished, that the expansion interface (“Mind Link Expansion Interface”) is a PCIe 5.0 x8 (256GT/s) “Additionally, the open-source… Read more »

Vall
Vall
13 days ago

> Idle – 3.9 – 4.2 Watts

That’s *outstandingly low* power consumption for an Intel CPU. Lower than any I’ve seen recently!

Has Intel finally got its game together in this regard with Meteor Lake?

Thanks Jean-Franc foi one more in-depth review. 👍

tkaiser
tkaiser
13 days ago

> Has Intel finally got its game together in this regard with Meteor Lake?

Huh? If a system is carefully designed and software/settings fit then you find many Intel powered systems that idle at or below 4W. And that does not only apply to ‘slow’ machines based on Jasper or Alder Lake-N but you’ll find also reports for e.g. i3-13100 idling that low.

BTW: when measuring at wall replacing a crappy power brick with something better can easily half idle consumption. Check for example Jeff testing a Twin Lake box idling north of 10W with vendor supplied power brick.

tkaiser
tkaiser
13 days ago

Found it: 2.5W idle (vendor provided power brick and all losses included) after adjusting settings to fit our needs: https://www.cnx-software.com/2021/10/15/2-5gbe-mini-pc-asus-pn41/#comment-585825

Boardcon CM3588 Rockchip RK3588 System-on-Module designed for AI and IoT applications