What Strix Halo is not for
The AMD Strix Halo / AI Max+ 395 platform is suitable for many local AI workflows, but it is not a universal AI machine. This page lists the main limitations so you can make an informed decision.
Not a training platform
Strix Halo is designed for inference, retrieval-augmented generation, agents, and multimodal workflows. It is not optimized for large-scale model training or fine-tuning that requires sustained high FLOPS over many hours.
If your workflow involves pre-training models, full fine-tuning of large models, or distributed training runs, a multi-GPU NVIDIA workstation or cloud cluster is typically more appropriate.
Not a CUDA machine
Strix Halo uses AMD integrated graphics, not NVIDIA discrete GPUs. This means CUDA and TensorRT are not available. Workflows built exclusively on CUDA may need to be ported to ROCm, DirectML, or CPU inference paths.
Many open-source tools support AMD GPUs through ROCm or Vulkan, but compatibility varies by project and version. The exact software stack is confirmed during the quote process.
Not the fastest image-generation box
Image generation with Flux, SDXL, and similar models is possible, but raw pixels-per-second throughput is lower than on an NVIDIA RTX 4090 or 5090 workstation. Strix Halo trades peak throughput for large unified memory capacity.
If your primary metric is maximum image-generation speed and you do not need 128 GB of unified memory, a dedicated NVIDIA GPU workstation is likely a better fit.
Other important boundaries
Memory is shared
CPU, GPU, and NPU share 128 GB. Running multiple large models concurrently may exhaust available memory.
No user-upgradeable RAM
LPDDR5X is soldered. You cannot add more RAM after purchase.
Software stack required
This is not a plug-and-play appliance. Model setup, quantization, and tool configuration are required.
Regional availability varies
Supply batches, certification, and shipping options vary by destination and partner brand.