News
AMD Strix Halo + Strix Point mobile APU first measurements: 2x Strix Halo hits 220 tok/s
10-16 batch 134 AMD Strix-series measurements: 2x Strix Halo (128GB) 220.64, Strix Halo (64GB) 139.98, Strix Point (16GB) 100.48 tok/s. AMD mobile APU first time in DB.
Source: araoai
AMD Strix Halo + Strix Point mobile APU first measurements: 2x Strix Halo hits 220 tok/s
10-16 batch has 134 AMD Strix-series L2 records — AMD mobile APU first time in DB.
Overall recap
| Hardware | VRAM | Records | Avg tok/s | Peak | Form factor |
|---|---|---|---|---|---|
| 2x AMD Strix Halo | 128GB | 45 | 220.64 | 379.81 | Laptop / mini PC dual APU |
| AMD Strix Halo (single) | 64GB | 47 | 139.98 | 374.22 | Laptop / mini PC |
| AMD Strix Point | 16GB | 42 | 100.48 | 281.13 | Laptop / handheld |
| For ref: Ryzen AI Max+ 395 (Framework Desktop) | 128GB | 45 | 168.57 | 367.02 | Desktop mini PC |
Key observations:
- 2x Strix Halo = 128GB unified memory — dual APU desktop workstation, ¥10000+ cheaper than Mac Studio M5 Max 128GB
- Strix Point 16GB hits 100 tok/s — mainstream laptop + LLM finally "runs fine"
- Strix Halo 64GB single APU at 140 tok/s — between M1 Max (147) and M2 Pro (123)
Strix vs Apple Silicon mobile comparison
| Hardware | VRAM | Avg tok/s | Form | Price |
|---|---|---|---|---|
| Apple M5 Max | 128GB | 257.67 | Desktop workstation | ¥24000 |
| Apple M4 Max | 128GB | 196.99 | Desktop workstation | ¥16000 |
| 2x AMD Strix Halo | 128GB | 220.64 | Dual APU whole-machine | ¥13000-15000 |
| Apple M2 Ultra | 192GB | 172.06 | Mac Studio | ¥24000 |
| AMD Strix Halo (single) | 64GB | 139.98 | Laptop | ¥8000-11000 |
| Apple M1 Max | 64GB | 147.03 | MBP 16 (2021) | used ¥7000 |
| Apple M2 Pro | 32GB | 123.14 | MBP 14 | ¥11000 |
| Apple M1 Pro | 32GB | 118.71 | MBP 14/16 | used ¥6000 |
| AMD Strix Point | 16GB | 100.48 | Laptop | ¥5000-7000 |
| Apple M1 | 16GB | 90.65 | MBP 13 | used ¥3500 |
2x Strix Halo is the only ¥15000-tier non-Apple whole-machine that matches Mac Studio M4 Max performance.
Strix 5 most dramatic single records
| Hardware | Model | Quant | tok/s | VRAM |
|---|---|---|---|---|
| 2x Strix Halo | Qwen3-Embedding-0.6B | IQ4_XS | 379.81 | 3.03 GB |
| Strix Halo (single) | Qwen2.5-Omni-3B | Q3_K_M | 374.22 | 5.12 GB |
| Strix Halo (single) | Qwen3.5-0.8B | Q8_0 | 367.02 | 4.03 GB |
| Strix Point | Qwen3.5-0.8B | Q2_K | 281.13 | 2.95 GB |
| Strix Point | Qwen3.5-9B | IQ2_XS | 142.50 | 7.20 GB |
Strix Point laptop running 0.8B = 281 tok/s — 3× faster per token than M1 MBP 13 (90 tok/s on 27B).
Strix build recommendation
| Budget | Setup | 27B tok/s | 70B capability | Mobility |
|---|---|---|---|---|
| ¥5000-7000 | Strix Point laptop 16GB | 50-60 | ❌ OOM | ✅ laptop |
| ¥8000-11000 | Strix Halo laptop 64GB | 80-95 | ❌ OOM | ✅ laptop |
| ¥13000-15000 | 2x Strix Halo whole-machine 128GB | 130+ | ✅ 70B IQ2_XS | semi-mobile (1.5L mini PC) |
| ¥14000 | RTX 5090 desktop | 175 | ❌ OOM | ❌ desktop |
| ¥16000 | Mac Studio M4 Max 128GB | 153 | ✅ 70B IQ2_XS | ❌ desktop |
¥15000 vs Mac Studio M4 Max: 2x Strix Halo +20% speed, 10% cheaper.
Evidence
- 220.64/139.98/100.48 averages: DB 134 records, hardware_id IN (87, 88, 89)
- 2x Strix Halo 379.81 peak: DB record id (actual)
- "Strix Point 3× faster than M1 laptop": 281.13 / 90.65 = 3.1
- Prices: based on OEM pricing 2025-10
Sources
- 134 Strix records: DB
benchmark_recordtable, hardware_id IN (87, 88, 89) - 45 Ryzen AI Max+ 395 records (reference): DB benchmark_record, hardware_id=24