AMD · Local AI machine
GMKtec EVO-X2 (64 GB)
AMD machine: AMD Ryzen AI Max+ 395, 64 GB memory, 48 GB for the GPU, 256 GB/s memory speed, about 3.9 tok/s on a 70B model (estimate).
79.9Score
- #23 of 41 ranked
- #8 of 18 AMD machines
Why this score
The score adds up these parts. Each part has a weight. A part counts as the average of its sources, pulled toward 80 when there are few sources.
AI power
50% weightcounts as 77.5Average of 2 sources: 76.3
Experts
30% weightcounts as 84.1Average of 2 sources: 86.1
Users
20% weightcounts as 79.6Average of 1 source: 79.2
Shown, not counted
These sit inside an average we already count, or come from the maker's own shop.
Good: 3 reviews and buyer ratings.
80 is the field average. A part counts as 80 + (average − 80) × N ÷ (N + 1), where N is its number of sources. One source counts half of the gap from 80. No sources count as 80.
Price
€1949 at galaxus.at · Geizhals.at listing, not checked live
EVO-X2 64 GB / 1 TB, Windows 11 Pro; lowest of 3 Austrian-market offers on Geizhals (others EUR 2,217.48 getgoods.com, EUR 2,252.50 Proshop.at)
Prices checked 8 Oct 2026.
Specs
- Chip
- AMD Ryzen AI Max+ 395
- CPU
- 16 cores / 32 threads
- Memory
- 64 GB
- GPU memory
- 48 GBWindows/BIOS: AMD Variable Graphics Memory, up to 75% of RAM = 48 GB on 64 GB (AMD rule; the largest share the vendor tools allow). Tom's Hardware's EVO-X2 review says 64 GB models allocate up to 32 GB in the BIOS. Tuned Linux GTT needs kernel parameters and is not used: about 56 GB (runaihome, amdgpu.gttsize=57344).
- Memory speed
- 256 GB/s
- FP16 TFLOPS
- 60 TFLOPS
- NPU
- 50 TOPS
- OS
- Linux, Windows
- Software
- ROCm, Vulkan
- Ethernet
- 2.5 Gbit/s
- Thunderbolt / USB4
- USB4
- QSFP port
- No
- Idle power
- 8.2 W
- Load power
- 214.9 W
- Noise
- 47.5 dB(A)
- Volume
- 2.8 L
- Weight
- 1.6 kg
- SSD
- 1 TB
- Release
- 2025-05
Measured
- gpt-oss-120b, MXFP4
- 46.1 tok/s writing, 818.1 tok/s prompt · llama.cpp ROCm (llama-bench, fa=1, ub 2048)
- Qwen3.5-122B-A10B, Q4_K_M (71 GiB)
- 23.1 tok/s writing, 315.7 tok/s prompt · llama.cpp Vulkan (llama-bench)
- Qwen3.6-35B-A3B, UD-Q4_K_M
- 61.5 tok/s writing, 1050.8 tok/s prompt · llama.cpp Vulkan/RADV (llama-bench)
Notes
Early Strix Halo mini PC and the cheapest 64 GB Strix Halo box in this index (EUR 1,949), in a 2.76 L chassis. It is also the loudest box Notebookcheck tested (47.5 dB(A) at 15 cm), with up to 215 W peak draw and only 2.5 GbE.
Noise: Notebookcheck 47.5 dB(A) at 15 cm (about 41 dB(A) in quiet mode), ServeTheHome 41-43 dBA during Llama 3.3 70B inference (studio noise floor 34 dBA, distance not stated); the table uses Notebookcheck. Power: ServeTheHome 147-160 W during Llama 3.3 70B Q6_K inference and 8-14 W idle, Notebookcheck 8.2 W idle and 214.9 W stress maximum; the table uses Notebookcheck. Maker page calls the port 'Gigabit LAN (2,5G)' while Geizhals and Tom's Hardware list 2.5 GbE. Geizhals also lists a 96 GB variant (no offers) that is not in candidates.json. The maker page (de.gmktec.com) shows EUR 1,999.99 for the base configuration, EUR 50 above galaxus.at, and mentions a EUR 50 summer code. Tom's Hardware's 80/100 is Geizhals' normalised value, read via Geizhals; the review page showed no star count. The Amazon.de rating was read from a search-result excerpt. GPU memory follows AMD's Variable Graphics Memory rule for every box with this chip (75% of RAM: 48 GB on 64 GB, 96 GB on 128 GB); tuned Linux figures are in the GPU memory note. FP16 TFLOPS is AMD's figure for the chip.
More reading:servethehome.comtomshardware.comcomputerbase.denotebookcheck.com3dtested.com