AMD Strix Halo = Ryzen AI Max+ 395 (16 Zen 5 cores) fused with the Radeon 8060S iGPU (gfx1151, RDNA 4, 40 compute units) on quad-channel LPDDR5X-8000 — 256 GB/s of unified memory, up to 128GB. Our unit: a GMKtec EVO-X2 AI (~$1,999 class) with 128GB, running Ollama on ROCm. It is the machine behind every benchmark on this site.
| Class | Example (measured) | Speed | Verdict |
|---|---|---|---|
| Sparse MoE 30B-class | GLM 4.7 Flash (30B-A3B) | 48.0 tok/s | ✅ The sweet spot |
| Sparse MoE 35B-class | Ornith 1.5 35B-A3B | 52.4 tok/s | ✅ Fastest big model here |
| Small dense 7-14B | Ornith 9B, Llama 3.2 3B | 31.8-120 tok/s | ✅ Instant |
| Mid dense ~32B | Qwen3.8 27B | 10.7 tok/s | ⚠️ Usable, patience required |
| Large dense 70B | Llama 70B Q4 | 3-5 tok/s | ❌ Not a 70B platform |
| System | Price | Memory | 70B speed |
|---|---|---|---|
| Strix Halo (ours) | ~$1,999 | 128GB unified | 3-5 t/s |
| NVIDIA DGX Spark | $3,999 | 128GB unified | ~5 t/s |
| Mac Studio M4 Ultra | $7,000 | 192GB | 20-25 t/s (est.) |
| RTX 4090 (offload) | $1,600 | 24GB + CPU | 60-80 t/s (est., hybrid) |
| H100 80GB | $30,000+ | 80GB | 25 t/s (pure GPU) |
The punchline: at 2× the price, DGX Spark buys software maturity, not dense-model speed. And none of these beat the real answer on this platform — don't run dense 70B locally at all; run MoE.
Generation is bandwidth-bound: speed ≈ bandwidth ÷ bytes-moved-per-token. A 30B-A3B MoE moves ~3B parameters per token; a 27B dense moves 27B. Same 256 GB/s: 4.5× the speed for similar capability class. This is why our daily driver is GLM 4.7 Flash (48 tok/s) rather than any dense model.
VRAM carve: 32GB UMA — Q4 30B-class fits fully in VRAM (100% GPU residency measured) with system RAM to spare. Runtime: Ollama 0.33 on ROCm 7.2; the 0.18→0.33 upgrade alone lifted GLM 4.7 Flash from ~38 to 48 tok/s — re-benchmark after every runtime update. Context: prefill decays from 830 tok/s (840-token prompt) to 217 tok/s at 25k — keep chat loops under ~8k and use deep context only for retrieval. Concurrency: two sessions at 34.3 tok/s each — parallel agents are nearly free.
The two 128GB unified-memory boxes everyone compares. Our measured corner:
| Workload | Strix Halo (Ryzen AI Max+ 395) | DGX Spark |
|---|---|---|
| Dense 70B Q4 | 0.7 tok/s measured; honest answer is 3-5 tok/s class | ~5 tok/s (reported) |
| Sparse MoE 30B-class | 48-52 tok/s sustained (our table above/on this page) | usable, at ~2x the price |
| Price | mini-PC street price | roughly 2x |
Bottom line from running one daily: for 70B dense both are slow; for MoE and mid-size models both are usable, and Strix Halo does it at half the cost. Full methodology and configuration notes below.
Buy Strix Halo if you want one quiet desk-side box for MoE chat/agents + image generation + everyday computing at ~$2k. Skip it if your workload is dense 70B+ — that's Mac Studio Ultra, multi-GPU, or API territory.
The measured LLM comparison Local Model Atlas AI-OS