Ubuntu 26.04 vs Windows 11 on ACEMAGIC K5: Dual Boot, AI Benchmarks and 8K Playback

All figures come from CNX's single review unit, the 12GB RAM / 512GB SSD model. The results are not general claims about every K5 or every Ubuntu setup.

What was tested​

CNX installed Ubuntu 26.04.1 LTS alongside Windows 11. Canonical published 26.04.1 as the first point release of the 26.04 LTS series, bundling the bug and security fixes collected since the April release. Phoronix notes it does not include a new hardware-enablement stack. That arrives with 26.04.2. The kernel in the review was 7.0.0-42, which matches the Linux 7.0 base of this LTS. Ubuntu 26.04 LTS shipped on 23 April with GNOME 50 and kernel 7.0.

The dual-boot install, and one odd error​

The reviewer shrank the Windows C: partition by roughly half, then installed Ubuntu into the freed space.

The first attempt failed with the message "No journal files were found." CNX did not think it was BIOS-related, but the install worked after Secure Boot was set to Disabled in the firmware. The reviewer says the problem was "somehow resolved," so Secure Boot is not confirmed as the cause. This is one machine's observation, not a documented fix.

If you hit the same error on a similar mini PC, the order of steps would be:

1. Shrink the Windows partition from within Windows first.
2. Boot the Ubuntu USB installer.
3. If the installer fails, enter the BIOS and set Secure Boot to Disabled.
4. Retry the install.

Disabling Secure Boot has a side effect for anyone who also uses Windows 11. Windows keeps booting, but the Secure Boot state is a security feature some organisations check. Re-enable it afterwards if your setup allows. CNX does not say whether it tried that.

After the install, Ubuntu's About screen listed the model as "Default String." That is a firmware metadata quirk, not a functional problem. The rest of the hardware was recognised:

• Firmware: AMI UEFI, version P3WL21001, dated 26 June 2026.
• Graphics: Intel's xe kernel driver with Mesa 26.0.8.
• Wi-Fi: Realtek RTL8852BE (rtw89_8852be).
• Ethernet: Realtek RTL8125 2.5GbE (r8169).
• Bluetooth: Realtek radio (btusb), reported as version 5.2.
• SSD: a "CF620" 512GB NVMe drive.

CPU and platform benchmarks​

The Core 3 304 is a five-core part with one performance core and four efficiency cores. In sbc-bench the P-core measured about 4.28GHz against a 4.5GHz maximum. The E-cores measured about 3.28GHz against 3.3GHz. Those readings did not change after a heat-up test that took the CPU from 55°C to 80°C.

Other CPU results:

• 7-Zip: 17,430 MIPS on average multi-threaded, and 5,051 single-threaded.
• Geekbench 7.1.0: 2,294 single-core, 6,299 multi-core.
• Power limits: 22W long-term, 35W short-term, 60W peak, the same as in CNX's Windows review.
• Unigine Heaven 1080p: 419 points, 16.6 FPS average. On Windows the same test gave 473 points and 18.8 FPS.
• Speedometer: 564 on version 2.0 and 31.9 on version 3.1, both in Firefox.

The AI and GPU compute gap​

This is the biggest practical difference from Windows. Geekbench AI 1.7.0 initially saw only CPU backends on Ubuntu.

CNX added Intel's Kobuk graphics PPA and installed the OpenCL, Level Zero, clinfo and GSC packages. It also added the user to the render group and rebooted. After that the iGPU showed up as an OpenVINO backend.

The CPU run through OpenVINO scored 903 (single precision), 632 (half precision) and 2,562 (quantized). The GPU run then failed during image classification with a CL_OUT_OF_RESOURCES error. CNX also installed an NPU driver, but Geekbench AI did not detect the NPU.

On Windows 11, CNX's GPU run scored 2,752 / 9,722 / 11,787 and its NPU run scored 850 / 16,657 / 26,212. The Windows machine could use the GPU and NPU in the same benchmark.

Caveats:

• The failure describes this software stack: Geekbench AI, OpenVINO and the Kobuk PPA packages.
• It does not show that the GPU or NPU is unusable in other Linux applications.
• CNX notes the same error on a Lunar Lake LattePanda kit. The GPU test did work on the GEEKOM IT13 Max.

If you plan to run local AI workloads on this class of hardware, Windows currently has the more complete path in CNX's testing.

Video playback: 4K fine, 8K not on Linux​

Firefox on Ubuntu handled YouTube 4K30 with 2 dropped frames out of 10,065. At 4K60 it dropped 208 of 18,118 frames over five minutes.

8K was a different story. 8K30 dropped 1,559 of 4,617 frames. 8K60 dropped 1,514 of 1,576. Chrome did not help. CNX's Windows 11 test played up to 8K60 smoothly. CNX says the Beelink EQi (also a Core 3 304) struggled with 8K on Ubuntu too, so this looks platform-wide rather than K5-specific. CNX does not diagnose a cause, so treat it as a limit of that Linux playback setup.

Since the box outputs at most 4K, few people will miss 8K.

Storage, USB and networking​

The NVMe SSD measured about 1,708 MB/s read and 1,599 MB/s write in Iozone. CNX's Windows CrystalDiskMark figures were 1,780 MB/s read and 1,676 MB/s write.

USB results with an NVMe enclosure:

• Front ports (USB-C and two USB-A): about 1,000–1,012 MB/s read.
• Rear USB-A ports: about 440 MB/s.
• The USB-C port: advertised at 20Gbps, but it negotiated 10Gbps because the reviewer had no 20Gbps hardware to test it.

CNX's Windows review noted that the ports are not marked by speed. The practical rule is to put fast drives on the front panel.

Networking over 60-second iperf3 runs:

• 2.5GbE: about 2.35Gbit/s in each direction, including bidirectional.
• Wi-Fi 6 (5GHz, Xiaomi AX6000 router): 824Mbit/s down and 809Mbit/s up.
• Bluetooth: file transfer to and from a Redmi Note 14 Pro 5G worked.

Thermals, noise and power​

Under a five-core stress load the CPU jumped from about 44–45°C idle to 90–91°C. Clocks peaked near 4.2GHz (P-core) and 3.3GHz (E-cores). They then settled around 3.5–3.6GHz and 2.7GHz. CNX says power limits were hit before the CPU reached its thermal junction limit, so it saw no thermal throttling on Ubuntu.

That differs from the Windows review. There, HWiNFO reported thermal throttling on the efficiency cores in 3DMark and on all five cores in Cinebench R23. CNX also noted a 28–30°C test room, so a cooler room might change the result. The sbc-bench clock readings and the stress-test frequencies are separate observations, so don't merge them.

Noise, measured 5cm above the case:

• Idle: 37.0–37.3 dBA.
• Stress: 46.3–47.0 dBA.
• Quiet-room baseline: 36.3–36.7 dBA.

CNX says the K5 is louder under load than the Beelink EQi Wildcat Lake. These are the reviewer's own readings, not standardised acoustic tests.

Power draw with the supplied 19V/3.42A adapter:

• Off: 0.7W.
• Idle: 4.3–4.5W.
• 4K60 YouTube: 14–16.5W.
• Stress: 35–37.6W at first, then 27.3–28W.

Via USB-C PD, readings were slightly lower: 3.4–4.2W idle and 25.2–25.6W sustained under stress. CNX's Windows review noted that these readings exclude charging-station losses.

How it compares​

The K5 tracks the Beelink closely. Its memory bandwidth is lower, which CNX attributes to the slower RAM. Its Speedometer scores are much higher. CNX guesses that software changes since the July Beelink test explain this, but that is conjecture. The Ryzen box wins on multi-core and graphics, while both Intel boxes beat it on single-core in these runs.

CNX lists the 12GB/512GB model at $399.99 on Amazon, with a 12GB/256GB version at $379.99. Those are CNX's reported prices, not independently checked ones.

Verdict​

CNX found Ubuntu works well for general use on the K5: solid single-core speed, full-speed 2.5GbE, decent Wi-Fi 6 and smooth 4K. The caveats are the install hiccup, the GPU/NPU compute gaps in AI benchmarks, and poor 8K playback. For Windows 11 users curious about Linux, the review suggests dual-booting is feasible. Anyone leaning on on-device AI should check driver support first.