AMD · Local AI machine
AMD Ryzen AI Halo (128 GB)
AMD machine: AMD Ryzen AI Max+ 395, 128 GB memory, 96 GB for the GPU, 256 GB/s memory speed, about 3.9 tok/s on a 70B model (estimate).
78.7Score
- #28 of 41 ranked
- #13 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 80.7Average of 2 sources: 81.1
Experts
30% weightcounts as 74.4Average of 2 sources: 71.6
Users
20% weightcounts as 80.0No sources found yet. No data yet. This part counts as 80.
Thin: only 2 reviews and buyer ratings, so they are pulled toward the average.
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
— No price yet.
No Austrian or EU listing found: Geizhals.at search shows no Ryzen AI Halo product. US price is 3,999 USD at Micro Center only, local pickup; heise estimates about 4,110 EUR including VAT. Stock reports conflict.
Prices checked 8 Oct 2026.
Specs
- Chip
- AMD Ryzen AI Max+ 395
- CPU
- 16 cores / 32 threads
- Memory
- 128 GB
- GPU memory
- 96 GBWindows/BIOS: AMD Variable Graphics Memory, up to 75% of RAM = 96 GB on 128 GB (AMD rule; the largest share the vendor tools allow). Signal65 set 96 GB on Windows for its tests; AMD says Variable Graphics Memory is preconfigured on Windows and Linux. Tuned Linux maxima need kernel parameters and are not used: up to 124 GiB (kyuz0 toolboxes on a Framework Desktop, Fedora, amdgpu.gttsize=126976, ttm.pages_limit=32505856, 4 GiB left to the OS), about 110 GB (runaihome), about 108 GB stable (Jeff Geerling).
- Memory speed
- 256 GB/s
- FP16 TFLOPS
- 60 TFLOPS
- NPU
- 50 TOPS
- OS
- Windows, Linux
- Software
- ROCm, Vulkan
- Ethernet
- 10 Gbit/s
- Thunderbolt / USB4
- USB4
- QSFP port
- No
- Idle power
- 12 W
- Load power
- 144 W
- Noise
- 45 dB(A)
- Volume
- 1 L
- Weight
- 1.2 kg
- SSD
- 2 TB
- Release
- 2026-07
Measured
- gpt-oss-120b
- 56.2 tok/s writing, 809 tok/s prompt · Vulkan and ROCm (runtime not named)
- Qwen 3.6 35B-A3B (MoE)
- 63.6 tok/s writing, 1273 tok/s prompt · Vulkan and ROCm (runtime not named)
- Qwen 3.6 27B (dense)
- 13 tok/s writing, 274 tok/s prompt · Vulkan and ROCm (runtime not named)
Notes
AMD's own developer box and the smallest Strix Halo box here (about 1 L), with 10 GbE and USB4. It matches the NVIDIA DGX Spark on gpt-oss-120b generation speed (56 vs 58 tok/s) but is 2 to 3 times slower at prompt processing. It is sold only at Micro Center in the US, so there is no Austrian price.
Idle power: ServeTheHome 11 to 13 W at the outlet (Debian), Signal65 31 W, HotHardware 33 to 35 W; ServeTheHome used. Load: ServeTheHome 98 to 144 W peaks, HotHardware 135 to 145 W typical and 150 W spikes, Signal65 196 W during gpt-oss-120b inference; ServeTheHome top value used. Noise: HotHardware 45 dBA at 12 in (30 cm) from the rear vents under a 90-minute stress test and calls it quiet; Tom's Hardware reports audible fans even at idle and a high-pitched whine under ComfyUI; Signal65 and LTT call it quiet or mild. USB ports: AMD says 3 USB-C plus 1 power USB-C, HotHardware says 4 USB-C (one for power), Phoronix says 4 USB 3.2 Type-C, ServeTheHome says 2 USB4 ports; speed not stated by AMD; a Level1Techs user ran USB4 networking. Weight is 'under 1.2 kg' from AMD. Stock: HotHardware and Micro Center list it; PCWorld said it could not be ordered (candidate research). HotHardware's score comes from the page's review markup. A speed run from another Strix Halo box was dropped. 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:phoronix.comservethehome.comsignal65.comheise.de