Descriptions:
Sam Witteveen reviews the AMD Ryzen AI Halo, a new workstation built around the Ryzen AI Max Plus 395 chip with 128 GB of unified LPDDR5X memory — a significant leap from his earlier look at the Radeon AI Pro R9700’s 32 GB discrete GPU. The unified memory architecture lets the CPU and GPU share one large memory pool, with users able to allocate up to 75% or more to GPU use. That makes it possible to run models like GPT OSS 120B and large Qwen 3 variants entirely locally, without the severe token-speed collapse that occurs when layers spill from discrete VRAM into system RAM.
The video walks through the out-of-box Linux experience, the AMD Ryzen AI Developer Center dashboard, and performance across multiple workloads. Witteveen tests large language models at different quantization levels, an agent setup using the Hermes agent framework, and generative media pipelines built in ComfyUI — including image and video models such as LTX and Wan. He also notes the onboard NPU (~50 TOPS) and expects LLM stacks to begin leveraging it more actively. A forthcoming Pro495 variant supporting up to 192 GB is mentioned as a future candidate.
The sponsored-but-demo-driven format makes this a practical reference for developers weighing local AI hardware options, especially those wanting to run 70B+ parameter models or build self-hosted image and video generation pipelines without ongoing cloud costs.
📺 Source: Sam Witteveen · Published July 21, 2026
🏷️ Format: Review
