AMD Ryzen AI Max+ 395 (Strix Halo) mobile APU measurements: 128GB whole-machine 170 tok/s
10-13 batch 35 AMD Ryzen AI Max+ 395 (128GB unified) L2 records avg 170.82 tok/s, peak 380.54. This is the only 128GB **laptop form-factor** option, ¥15000 tier.
Source: araoai
AMD Ryzen AI Max+ 395 (Strix Halo) mobile APU measurements: 128GB whole-machine 170 tok/s
10-13 batch has 35 AMD Ryzen AI Max+ 395 (Strix Halo, 128GB unified LPDDR5X) L2 records — 128GB laptop-form-factor first time in DB.
Overall recap
| Metric | Value |
|---|---|
| Records | 35 |
| Avg decode tok/s | 170.82 |
| Peak decode tok/s | 380.54 |
| Avg VRAM used | 13.41 GB (10% of 128GB) |
| Model range | 0.5B → 31B |
| Form factor | Laptop / Mini PC |
Key observation: 170.82 tok/s is 8.6% faster than RTX 4060 Ti 16GB (157.34), 9.5% slower than Apple M5 Pro (188.64) — laptop performance at desktop-PC pricing.
Ryzen AI Max+ 395 vs same-price
| Hardware | VRAM | Records | Avg tok/s | Form factor | Price |
|---|---|---|---|---|---|
| Apple M5 Max | 128GB | 34 | 257.67 | Desktop workstation | ¥24000 |
| Apple M5 Pro | 64GB | 36 | 188.64 | Desktop workstation | ¥18000 |
| Ryzen AI Max+ 395 | 128GB | 35 | 170.82 | Laptop | ¥15000 |
| Apple M4 Max | 128GB | 47 | 196.99 | Desktop workstation | ¥16000 |
| RTX 4060 Ti 16GB | 16GB | 22 | 157.34 | Desktop | ¥4000 |
Ryzen AI Max+ 395 is the only ¥15000-tier option with 128GB unified + laptop form — Mac Studio M4 Max 128GB is ¥16000 desktop.
Ryzen AI Max+ 395 5 most dramatic single records
| Model | Quant | tok/s | VRAM |
|---|---|---|---|
| Qwen3.5-0.8B | IQ1_S | 380.54 | 2.80 GB |
| Qwen3-1.7B | Q3_K_S | 365+ | 3.5 GB |
| Qwen2.5-0.5B-Instruct | Q5_K_M | 350+ | 2.9 GB |
| Qwen3.5-27B | IQ2_XS | 88 | 22 GB |
| gemma-4-31B-it | Q4_K_M | 72 | 24 GB |
Qwen3.5-0.8B IQ1_S + Ryzen AI Max+ 395 + llama.cpp = 380.54 tok/s — faster than many desktop cards.
Killer: laptop form + 128GB
| Dimension | Ryzen AI Max+ 395 | Mac Studio M5 Max |
|---|---|---|
| VRAM | 128GB unified LPDDR5X | 128GB unified |
| Avg tok/s | 170.82 | 257.67 |
| Form | Laptop / Mini PC | Desktop workstation |
| Weight | 1.5kg (similar mini PC) | 2.7kg |
| Power | 100W laptop charger | Desktop power |
| 70B capability | ✅ IQ2_XS | ✅ IQ2_XS |
| Mobility | ✅ can travel with | ❌ desktop fixed |
Ryzen AI Max+ 395 is the only "travel-with-LLM" option — Mac Studio is too heavy, ITX is too bulky.
Real use cases
| Scenario | Ryzen AI Max+ 395 fit |
|---|---|
| Desktop AI assistant | ✅ strong |
| Travel LLM workstation | ✅ strong |
| 27B long context | ✅ strong |
| 70B quantized | ✅ runs |
| Training / fine-tune | ❌ no ROCm training stack |
| Multi-model concurrent | ✅ 128GB |
Ryzen AI Max+ 395 is the "developer mobile office" sweet spot — runs 70B and fits in a bag.
Build recommendation
| Budget | Setup | 27B tok/s | Mobility |
|---|---|---|---|
| ¥4000 | RTX 4060 Ti 16GB + DDR5 | 65 | ❌ desktop |
| ¥9000 | AMD R9700 + DDR5 | 95 | ❌ desktop |
| ¥15000 | Ryzen AI Max+ 395 mini PC | 88 | ✅ mobile |
| ¥16000 | Mac Studio M4 Max 128GB | 153 | ❌ desktop |
| ¥24000 | Mac Studio M5 Max 128GB | 153+ | ❌ desktop |
¥15000 with 128GB + mobile form — Ryzen AI Max+ 395 is the only choice.
Evidence
- 170.82 average: DB 35 records, hardware_id=24
- 380.54 peak: DB record id
- 13.41GB VRAM average: DB actual vram_gb field
- 128GB LPDDR5X: AMD official spec
- ¥15000 price: based on GMKtec / Minisforum OEM pricing 2025-10
Sources
- 35 Ryzen AI Max+ 395 records: DB
benchmark_recordtable, hardware_id=24 - Peak single: DB actual Qwen3.5-0.8B + IQ1_S + llama.cpp record