Earlier this month, I received the Khadas Mind 2 (Intel Core Ultra 7 155H) mini PC with two accessories connecting through the Mind Link (PCIe x8) connector: The Mind xPlay Display and Keyboard combo, and the Mind Graphics 2 dock, adding a range of interfaces and an NVIDIA GeForce RTX 5060 Ti GPU with 16GB of VRAM.
After going through the specifications, an unboxing, and a partial teardown in the first part of the review, I tested the Mind xPlay using the Mind and Mind 2 mini PCs running Windows 11 and Ubuntu 24.04. I’ve now had time to test the Khadas Mind Graphics 2 dock with the Mind 2 mini PC running Windows 11, so I’ll report my experience with the NVIDIA GPU (3D graphics and AI), and test all features, including 2.5GbE networking, the built-in speakers, microphone array, and so on.
A few gremlins and system info
When I first connected the Mind Graphics 2 dock to the Mind 2 mini PC, the NVIDIA GeForce RTX 5060 Ti graphics card was detected, but was disabled with a message reading “the device has been stopped because it has reported problems”. HDMI output from the dock didn’t work at all. To fix that, I connected the Mind 2 to the display through the USB-C DP port, and I went to the Mind app on the mini PC, upgraded the firmware of the dock to version 1.9, and more importantly, downloaded and installed khadas-nv-43-fix.zip from https://dl.khadas.com/products/mind-graphics-2/drivers/.
That’s what the Device Manager window looks like without (left) and with (right) the Mind Graphics 2 dock.
The microphone and speakers from the “Khadas Mind Graphics Speaker” show up, as well as the NVIDIA GeForce RTX 5060 Ti, the 2.5Gbps Ethernet port, and a few extra USB controllers, including a Realtek USB 3.0 card reader.
GPU benchmarks using Khadas Mind 2 without and with the Mind Graphics 2 dock
Since the NVIDIA GeForce RTX 5060 Ti is the star of the show, I’ll get right into it with GPU benchmarks using the Mind 2 mini PC in standalone mode. mostly testing the 2.25 GHz 8x Xe core Intel Arc graphics inside the Intel Core Ultra 7 155H SoC, and with the dock leveraging the NVIDIA graphics card.
The following benchmarks will be used to evaluate both 3D graphics and AI performance:
- 3DMark – Fire Strike and Nomad Steel (higher-end test)
- Unigine – Heaven and Superposition (newer test, still 2017)
- Final Fantasy XV benchmarks to evaluate the performance of an actual (2024) game
- GeekBench AI 1.7.0
- MLPerf Client 1.6.1
I’ll first list the raw results with screenshots, before adding a comparison table to show both configurations side-by-side. Note that all tests were done in a room with an ambient temperature of 30-31°C.
Here are the results for the Mind 2 in standalone mode.
7,444 points in the 3DMark Fire Strike benchmark, with the software showing “Good” performance.
651 points in the more demanding Steel Nomad benchmark.
Unigine Heaven 4.0 was run at the standard 1920×1080 resolution, and the system could render the scene at 71.8 FPS on average, and a score of 1,809 points.
In the more recent Unigine Superposition benchmarks, the system achieved 3,817 points.
Final Fantasy XV benchmark was run at Standard Quality (1920×1080) resolution. The score was 3,487 points, and the Performance marked as “Standard”.
Time to switch to AI benchmarks, starting with Geekbench 1.7.0 using the OpenVino framework with GPU AI backed, and Intel Arc Graphics AI device.
The scores were 8,087 points for single precision, 12,987 points for half-precision, and 19,698 points for quantized. I used the 18 TOPS GPU, since the NPU (Intel AI Boost) was not an option.
The final test was MLPerf Client 1.6.1.
The NPU test all failed, but we still got the OpenVino GPU test for Llama 3.1 8B Instruct and Phi 3.5 Mini Instruct, which I’ll use for comparison.
Let’s now insert the Mind 2 on top of the Mind Graphics 2 dock and run all these tests again.
30,659 points on Fire Strike with “Excellent” performance.
3,487 points on Steel Nomad with “Good” performance.
The Unigine Heaven Benchmark 4.0 is a bit old for a powerful NVIDIA GPU, and the system delivered 365.8 FPS on average with a score of 9,214 points.
The kit achieved 24,895 points in the Unigine Superposition benchmark.
In the Final Fantasy XV benchmark, the score was 15,542 points, and the performance was rated as “Extremely High” in Standard Quality.
I also maxed out the benchmark with a custom configuration enabling all options to high/highest to stress the system a bit more. The score was 6,028 points, and the performance was considered “High”.
Geekbench AI was configured with the ONNX framework, DirectML backend, and NVIDIA GeForce RTX 5060 Ti device.
The scores were 20,587 points for single precision, 36,368 points for half-precision, and 16,329 points (lower than on the iGPU) for quantized.
All six Base tests from MLPerf Client could be completed. I’ll go with the CUDA results for comparison.
Time for the comparison table. A ratio of 1x means the performance is identical, over 1x the NVIDIA GPU is faster, and under 1x the Intel Arc iGPU is faster.
| Benchmarks | Mind 2 | with Mind Graphics 2 | Ratio |
|---|---|---|---|
| 3DMark Fire Strike | 7,444 | 30,659 | 4.12x |
| 3DMark Steel Nomad | 651 | 3,487 | 5.36x |
| Unigine Heaven | 1,809 | 9,214 | 5.09x |
| Unigine Superposition | 3,817 | 24,895 | 6.52x |
| Final Fantasy XV | 3,487 | 15,542 | 4.46x |
| GeekBench AI | |||
| - Simple precision | 8087 | 20,587 | 2.55x |
| - Half precision | 12,987 | 36,368 | 2.80x |
| - Quantized | 19,698 | 16,329 | 0.83x |
| MLPerf Client | |||
| Llama 3.1 8B Instruct (TPS) | 14.6 | 73.9 | 5.06x |
| Phi 3.1 Mini Instruct (TPS) | 25.5 | 116.4 | 4.56x |
For 3D graphics, the Mind Graphics 2 delivers 4x to 6x+ higher performance compared to the Intel Core Ultra 7 155H SoC with Intel Arc graphics. AI models like Llama 3.1 8B Instruct and Phi 3.1 Mini Instruct, tested with MLPerf, are also much faster thanks to the NVIDIA GPU, which yielded about 5x higher performance in terms of tokens per second. GeekBench AI is an outlier with around 2.5x to 3x higher performance for single and half precision, but somehow the quantized score (INT8) is lower than running on the Intel GPU with OpenVino. It looks like the culprit may be DirectML’s NVIDIA INT8 support rather than anything related to the hardware.
The good news is that I have no stability problem running tests with the Mind 2 connected to the Mind Graphics 2 dock, once the initial driver issue was resolved.
Khadas Graphics 2 features – 2.5GbE, Audio, Video output….
Test all features
- HDMI 1 – 4K 60Hz video OK, audio OK; note I had to manually enable the NVIDIA High Definition Audio driver in Device Manager for this to work (only the built-in speakers would show otherwise)
- HDMI 2 – 4K 60Hz video OK, audio OK
- DisplayPort – 4K 60Hz video OK, audio OK
- Quad display – OK. 2x 4K monitors connected to HDMI 2.1 and DP + a Full HD transflective LCD connected to HDMI 2.1 on the Khadas Mind Graphics 2, and one Full HD portable monitor to the USB-C DP port on the Mind 2 mini PC
- Microphone array – OK
- Speakers – OK (stereo), very good quality, still works when the display(s) is/are turned off
- 3.5mm audio jack – OK, tested with USB-powered speakers
- Card reader – OK. Tested with a Raspberry Pi microSD card (via full SD card adapter)
- 2.5GbE port
- OK for unidirectional tests. iperf 3.19: 2.35 Gbps Rx; 2.35 Gbps Tx
- Failed for full-duplex/bidirectional. iperf 3.19: 2.36 Gbps Tx/196 Mbps Rx (potentially not ideal for a BitTorrent node…)
- USB
- Front USB-C (10Gbps) port – OK, tested with HWiNFO 64 and CrystalDiskMark (1039 MB/s sequential read)
- Rear USB-C (40 Gbps) port – Neither an ORICO NVMe enclosure nor a Beelink Expand M USB-C dock is recognized. This may be a limitation of the Mind Graphics 2 dock when connected through the Mind Link. The USB-C port doesn’t support 40 Gbps/Thunderbolt, and should fallback to 10Gbbps USB, but it doesn’t seem to work in that case. This port should be most useful when connected to a laptop or other mini PC as a dock/eGPU.
- USB 3.2 Gen2 (10 Gbps) port #1 (left) – OK, tested with HWiNFO 64 and CrystalDiskMark (1039 MB/s)
- USB 3.2 Gen2 (10 Gbps) port #2 (right) – OK, tested with HWiNFO 64 and CrystalDiskMark (1038 MB/s)
- Mind unlock button – OK. This button is required to remove the Mind 2 from the dock. A short press should bring a pop-up window, but once it didn’t work, and I had to press it for 8 seconds to force an unlock (video output is turned off when triggered).
- Volume buttons – OK
- Fingerprint reader – OK (automatic sign-in). The Mute function also works fine.
Conclusion
If you can afford it, I find the Khadas Mind ecosystem to be pretty neat. The Khadas Mind 2 mini PC itself is already powerful, but you can convert it into a laptop with the Mind xPlay (13-inch display and keyboard) or a full gaming platform or AI workstation with the Mind Graphics 2 eGPU dock with a (non-replaceable) NVIDIA GeForce RTX 5060 Ti 16GB graphics card.
I had some frustrating times, notably with the BIOS update and manually installing the patch for the NVIDIA drivers, but once I got everything set up, I was very happy with the performance of the GPU, which is about 4 to 6 times faster for 3D graphics, and 2.5 to 5x faster for AI workloads than the integrated Intel Arc graphics in the Core Ultra 5 155H Meteor Lake SoC. The combo basically provides the performance of a high-end tower PC in a much smaller form factor.
Other features mostly worked great, too. This includes the built-in stereo speakers (good quality too), the dual-mic array, the 3.5mm audio jack, USB 3.2 Gen2 (10 Gbps) ports, and the fingerprint scanner. The 2.5GbE port also works fine in normal conditions, except for saturated bi-directional transfers, where the performance collapses in one direction using iperf 3.19. The USB-C (40 Gbps) port on the rear panel was not usable for me when testing it with two Thunderbolt mass storage devices. That’s because the port is limited to 10 Gbps USB 3.2 when it’s attached to the Mind 2 through the Mind Link. USB fallback should have worked in theory, but it didn’t here. This USB-C appears to be especially useful when using the Mind Graphics 2 as an eGPU dock with another machine, but we don’t own any hardware with a Thunderbolt port right now, so I had to skip that part.
I’m not quite done yet, as the next step will be to test the Sermoon S1 3D scanner with the Mind 2 + Mind Graphics 2 dock, since the Intel Raptor Lake laptop I used last time around was clearly not powerful enough for the task. I’m not sure I’ll test the Mind Graphics 2 with Ubuntu 26.04, as we received more review samples than anticipated, and I’m probably busy until mid July, but let me know if you’d like a Linux test anyway.
I’d like to thank Khadas for sending the Mind 2, Mind xPlay, and Mind Graphics 2 dock for review. You’ll find the mini PC for about $1,099.00 on AliExpress and the Khadas store in its Intel Core Ultra 7 155H/32GB/1TB configuration, while the Mind xPlay kit goes for $399 on AliExpress/Khadas store, and the Mind Graphics 2 dock sells for $1349 (AliExpress or Khadas Store). This means a complete system, as reviewed above (without the xPlay), costs $2448 plus shipping and taxes.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links.
