Strix Halo / AMD Ryzen AI Max+ 395 for Local AI (2026): 128GB Unified Memory in a Mini PC

Strix Halo — AMD's Ryzen AI Max+ 395 — is the most interesting local-AI hardware platform of 2025-2026. 128 GB of unified memory accessible by a 40-CU RDNA 3.5 iGPU, all in a 65-120 W mini-PC starting around $2,000. It runs Llama 3.1 70B in BF16 entirely on the iGPU. No consumer discrete NVIDIA card can do that. The Mac Studio M4 Max can, but costs twice as much.

This guide covers everything: the hardware spec, system options (Framework Desktop, Asus, HP), ROCm 6.3+ setup with gfx1151, BIOS memory allocation, real benchmarks vs Mac Studio M4 Max and dual-GPU NVIDIA, Ollama / vLLM / llama.cpp recipes, image and video generation, the XDNA 2 NPU situation, and the workloads where Strix Halo wins or loses.

Table of Contents

1. What Strix Halo Is
2. Why 128 GB Unified Memory Matters
3. Hardware Specs
4. Available Systems (Framework, Asus, HP)
5. vs Mac Studio M4 Max
6. vs Multi-GPU NVIDIA Builds
7. BIOS: Allocating Memory to iGPU
8. ROCm Setup for gfx1151
9. Ollama on Strix Halo
10. llama.cpp Native Build
11. vLLM-ROCm
12. PyTorch + Hugging Face
13. Image Generation Performance
14. Video Generation (Wan 2.2, Hunyuan)
15. The XDNA 2 NPU
16. Real Benchmarks
17. Power, Thermals, Acoustics
18. Use Cases Where Strix Halo Wins
19. Where Discrete GPUs Still Win
20. Buying Advice
21. Troubleshooting
22. FAQ

---

What Strix Halo Is {#what-it-is}

Strix Halo is the codename for the high-end variant of AMD's 2025 mobile/desktop AI platform branded as Ryzen AI Max+ 395 (and its lower-tier siblings 390, 385). Architecture:

• CPU: 16 Zen 5 cores / 32 threads (split across two CCDs)
• iGPU: 40 RDNA 3.5 compute units, codenamed gfx1151
• NPU: XDNA 2, ~50 TOPS INT8 (separate from CPU and iGPU)
• Memory: Up to 128 GB LPDDR5X-8000 unified (CPU + iGPU share it)
• Memory bandwidth: ~256 GB/s
• Process: TSMC N4P
• TDP: Configurable 55-120 W

Released early 2025. Successor (Medusa Halo) expected 2026-2027.

---

Why 128 GB Unified Memory Matters {#why-unified}

For local LLMs, the biggest constraint is "does the model fit on the GPU?" Discrete consumer GPUs cap at 24 GB (RTX 3090 / 4090 / 7900 XTX) or 32 GB (RTX 5090). At 24 GB you can run:

• 8B models in FP16 ✅
• 14B models in INT4-INT8 ✅
• 32B models in INT4 ✅
• 70B models only with offload (slow) ❌
• 70B models in FP16 ❌
• 405B models ❌

128 GB unified memory unlocks:

• 70B models in BF16 ✅ (no quantization quality loss)
• 100B-class models in INT4 ✅
• 200B+ MoE models in INT4 ✅
• Long contexts (131K+) on big models ✅

The trade-off: unified memory bandwidth (~256 GB/s) is much lower than discrete VRAM (~1,000+ GB/s), so per-token speed on small models is lower. But for big models, discrete GPUs can't run them at all without offloading.

---

Hardware Specs {#specs}

For local AI the 395 is the only variant worth considering — fewer iGPU CUs on the 390/385 means proportionally lower throughput. Always pair with the 128 GB memory option.

---

Available Systems (Framework, Asus, HP) {#systems}

Mid-2026 platforms shipping Strix Halo:

Best value: Framework Desktop with 128 GB. $1,999 mid-2026. Mini-ITX-style chassis, replaceable I/O, full Linux support, 65W typical / 120W boost. The closest to "open hardware" in this segment.

---

vs Mac Studio M4 Max {#vs-mac}

Mac wins on per-core efficiency, build quality, and macOS integration. Strix Halo wins on price (~50% less), Linux ecosystem, x86 software compatibility, and slightly higher 70B throughput. For pure $/perf on local LLMs in 2026, Strix Halo is hard to beat.

---

vs Multi-GPU NVIDIA Builds {#vs-nvidia}

A 2x RTX 4090 PCIe rig costs $3,000-3,500 (GPUs alone) plus system, total $4,500+. It delivers:

• Llama 3.1 8B FP16: 250+ tok/s (vs Strix Halo 48)
• Llama 3.1 70B AWQ: 38 tok/s (vs Strix Halo 32)
• Llama 3.1 70B BF16: doesn't fit in 48 GB
• SDXL 1024²: 4 sec (vs Strix Halo 14-18)
• Power: 600-900 W full-load (vs Strix Halo 90 W)

The dual-4090 wins on small-model speed and image gen. Strix Halo wins on 70B BF16 capability, power, noise, form factor, and price. For most "I want to run 70B locally without compromises" buyers, Strix Halo is the smarter choice.

---

BIOS: Allocating Memory to iGPU {#bios}

By default, Strix Halo systems allocate ~16-32 GB to the iGPU and leave the rest for system RAM. To run 70B models on the iGPU, you need to allocate ~96-110 GB to it.

Most BIOSes expose this under: Advanced → AMD CBS → NBIO Common Options → GFX Configuration → UMA Frame Buffer Size.

Some systems (Framework Desktop, ROG Flow) expose this in a simpler "AI Memory Reservation" menu. Reboot required after change.

---

ROCm Setup for gfx1151 {#rocm}

ROCm 6.3+ has official Strix Halo (gfx1151) support:

wget https://repo.radeon.com/amdgpu-install/6.3/ubuntu/jammy/amdgpu-install_6.3.60300-1_all.deb
sudo apt install ./amdgpu-install*.deb
sudo amdgpu-install --usecase=rocm,hiplibsdk -y
sudo usermod -aG render,video $USER
sudo reboot

Some workloads still need the gfx version override:

echo 'export HSA_OVERRIDE_GFX_VERSION=11.5.1' >> ~/.bashrc
echo 'export HCC_AMDGPU_TARGET=gfx1151' >> ~/.bashrc
source ~/.bashrc

Verify:

rocminfo | grep gfx
# Expected: gfx1151 (Strix Halo iGPU)
rocm-smi --showmeminfo vram
# Should show ~96-110 GB available depending on BIOS allocation

---

Ollama on Strix Halo {#ollama}

curl -fsSL https://ollama.com/install.sh | sh

Edit systemd to set the override:

sudo mkdir -p /etc/systemd/system/ollama.service.d
sudo tee /etc/systemd/system/ollama.service.d/override.conf <<EOF
[Service]
Environment="HSA_OVERRIDE_GFX_VERSION=11.5.1"
Environment="HCC_AMDGPU_TARGET=gfx1151"
EOF
sudo systemctl daemon-reload
sudo systemctl restart ollama

# Run a 70B model — fits entirely in unified memory
ollama run llama3.1:70b

For BF16 70B (large but possible on 128 GB):

ollama run llama3.1:70b-instruct-fp16

---

llama.cpp Native Build {#llamacpp}

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp

HIPCXX="$(hipconfig -l)/clang" \
HIP_PATH="$(hipconfig -R)" \
cmake -B build \
    -DGGML_HIP=ON \
    -DAMDGPU_TARGETS=gfx1151 \
    -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

./build/bin/llama-cli -m llama-3.1-70b-instruct-Q4_K_M.gguf -ngl 999 -fa

For long context: enable Q4 KV cache (--cache-type-k q4_0) — saves ~50% KV memory.

---

vLLM-ROCm {#vllm}

docker pull rocm/vllm:latest

docker run --device /dev/kfd --device /dev/dri \
    --group-add video --group-add render \
    --security-opt seccomp=unconfined \
    --shm-size 16G \
    -e HSA_OVERRIDE_GFX_VERSION=11.5.1 \
    -p 8000:8000 \
    rocm/vllm:latest \
    vllm serve casperhansen/llama-3.1-70b-instruct-awq \
    --quantization awq \
    --max-model-len 16384 \
    --gpu-memory-utilization 0.85

vLLM-ROCm on Strix Halo is functional but throughput-bound by the 256 GB/s memory bandwidth — single-stream decode is similar to llama.cpp / Ollama, but multi-user batching has less headroom than discrete GPUs.

---

PyTorch + Hugging Face {#pytorch}

pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm6.3

Run any HF Transformers model:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-70B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

The 128 GB unified memory means BF16 70B loads without sharding. Inference at 14 tok/s, training is impractical (compute-bound, not memory-bound).

---

Image Generation Performance {#image-gen}

Strix Halo is roughly 3-4x slower than RTX 4090 for image gen — compute-bound, not memory-bound. For LLM-primary use cases this is acceptable; for image-gen primary, get a discrete GPU.

---

Video Generation (Wan 2.2, Hunyuan) {#video-gen}

Slow but functional for occasional generation. The 128 GB memory means you can fit unquantized video models that don't fit on a 4090, but compute time makes routine use impractical.

---

The XDNA 2 NPU {#npu}

The 50-TOPS XDNA 2 NPU is largely unused for general LLM inference as of mid-2026. AMD's Lemonade SDK and Ryzen AI Software stack target it for specific accelerated paths — Llama 3.1 8B, Phi 3.5 Mini, etc. — but the iGPU is faster for most workloads.

The NPU shines for:

• Always-on background AI (Windows Studio Effects, Recall)
• Battery-sensitive laptop workloads
• Specific quantized models AMD ships optimized kernels for

For mainstream LLM inference, ignore the NPU and use the iGPU. Future ROCm releases may expand NPU coverage.

---

Real Benchmarks {#benchmarks}

Framework Desktop 128 GB, Ubuntu 24.04, ROCm 6.3, BIOS UMA = 96 GB.

For Mac M4 Max comparison see Apple M4 for AI Guide. For dual-GPU comparison see Multi-GPU Ollama Setup.

---

Power, Thermals, Acoustics {#power}

Configurable cTDP from 45 W to 120 W. Sweet spots:

For 24/7 inference servers, 65-90 W is the sweet spot — small efficiency gains plateau above 90 W. Most Strix Halo systems ship with adjustable power profiles in BIOS or vendor utility.

---

Use Cases Where Strix Halo Wins {#use-cases}

1. 70B local inference on a single device — no consumer NVIDIA card matches.
2. Large model fine-tuning data prep — load full BF16 model for inspection / activation logging.
3. Multi-tenant home server — 128 GB lets you run multiple medium models simultaneously.
4. Quiet / efficient desktop — 65-90 W for usable 70B inference.
5. Privacy / air-gapped 70B — single mini-PC, easy to lock down. See Air-Gapped AI Deployment.
6. Long-context RAG / agents — can fit full 131K context on 70B with Q4 KV cache.
7. Mac Studio alternative for PC users — same use case, half the price.

---

Where Discrete GPUs Still Win {#discrete-wins}

1. Small model speed — RTX 4090 hits 127 tok/s on 8B vs 48 on Strix Halo.
2. Image generation — 3-4x faster on RTX 4090.
3. Multi-user concurrent serving — vLLM PagedAttention scales better with VRAM bandwidth.
4. Fine-tuning — discrete GPU compute throughput dominates.
5. TensorRT-LLM — NVIDIA only.
6. FP8 hardware — RDNA 3.5 lacks FP8.
7. Latency-sensitive paths — discrete VRAM bandwidth wins on first-token-time.

---

Buying Advice {#buying}

Buy Strix Halo if:

• You want to run 70B models locally without offload pain.
• You value quiet, low-power, mini-PC form factor.
• You're comfortable with Linux + ROCm or Windows + WSL2.
• $2,000-3,000 budget; Mac Studio M4 Max is too expensive.
• LLMs are the primary workload; image gen is occasional.

Buy a discrete GPU instead if:

• You primarily run 7B-32B models — RTX 4090 / 7900 XTX much faster.
• You do heavy image / video generation.
• You need TensorRT-LLM or ExLlamaV2.
• You serve many concurrent users.

Buy Mac Studio M4 Max instead if:

• You're already in the Apple ecosystem.
• You need MLX-specific frameworks.
• Budget allows the ~2x premium.

---

Troubleshooting {#troubleshooting}

---

FAQ {#faq}

See answers to common Strix Halo questions below.

---

Sources: AMD Ryzen AI Max product page | Framework Desktop announcement | ROCm 6.3 release notes | Internal benchmarks Framework Desktop 128 GB.

Related guides:

• AMD ROCm Setup for Local LLMs
• Radeon RX 7900 XTX for Local AI
• Apple M4 for AI Guide
• Mac Studio vs PC AI Build
• Best Mini PC for Ollama
• Homelab AI Server Build