AMD Ryzen AI Max+ 395 (128GB unified, Strix Halo)
Compare AMD Ryzen AI Max+ 395 (128GB unified, Strix Halo) against another machine in the Arena
Reference: Zen 5 on Wikipedia
- Class
- APU
- Memory
- 128 GB LPDDR5X-8000
- Runs (Q4_K_M)
- —
- Bandwidth
- 256 GB/s
- TDP
- 120 W
- Released
- 2025-03-17
- Price (US)
- $1999.99 new as of 2026-08
Sourcing and disambiguation notes
40 CU RDNA 3.5 iGPU (Radeon 8060S) + 16-core Zen 5 CPU + XDNA2 NPU (50 TOPS int8, AMD-quoted, not independently verified against an LLM workload). AMD-quoted memory bandwidth is 256 GB/s theoretical; community measurement puts real achieved bandwidth at ~215 GB/s. The GPU cannot address all 128GB: the BIOS/UMA setting caps dedicated VRAM at 96GB per node, and the community benchmarks below allocate GPU memory via GTT/kernel parameters instead, reaching ~110-120GB usable. ROCm for gfx1151 is available via TheRock nightly builds and, since ROCm 6.4.4/7.0, official releases; Vulkan is the lower-friction default and is generally faster for short-context decode, while ROCm pulls ahead at long context with Flash Attention. ROCm 7.0.1 showed a severe prefill regression against 6.4.4 on this chip, so the backend VERSION, not just the backend, materially changes the results below. The recorded price is GMKtec's EVO-X2 sale price for this 128GB/2TB configuration; GMKtec lists it at $2,199.99 and other retailers have shown $1,700-$2,300. The Framework Desktop, the machine behind the original $1,999 launch figure, is reported to have risen to roughly $3,449, which could not be confirmed against Framework's own order page.
Specification sources
- https://www.amd.com/en/products/processors/desktops/ryzen/ryzen-ai-halo/ryzen-ai-max-plus-395.html — vendor, 2025-03-17
- https://www.amd.com/en/blogs/2025/amd-ryzen-ai-max-395-processor-breakthrough-ai-.html — vendor, 2025-03-17
- https://community.frame.work/t/amd-strix-halo-ryzen-ai-max-395-gpu-llm-performance-tests/72521 — forum, 2025-07-22
Measurements
Decode is token generation — the speed you feel while an answer streams. Prefill is prompt processing — the wait before it starts. Why bandwidth predicts decode speed.
No records yet for this hardware. Know of a published benchmark on it? Submit the link.
What verified, single-source and estimated mean, and the same rows with every filter and sort in the benchmarks explorer.
Speed over time
What it can run
Fit is arithmetic, not a measurement — how it is computed. Against 128 GB; models that fit are listed largest first.
No modeled quant fits in 128 GB.