From Strix Halo to Gorgon Halo: What’s New in Ryzen AI Max 400
AMD’s Ryzen AI Max PRO 400 series, codenamed Gorgon Halo, is a mid-cycle refresh of its Strix Halo-based Ryzen AI Max 300 line. Architecturally, the new chips stick with Zen 5 CPU cores, RDNA 3.5 integrated graphics, and an XDNA 2 neural processing unit, so raw compute gains are modest. The flagship Ryzen AI Max+ PRO 495 gets a small clock bump to a 5.2GHz CPU boost and a rebranded Radeon 8065S GPU with 40 compute units, while the NPU tops out at 55 TOPS. Lower-tier PRO 490 and PRO 485 parts retain 5.0GHz boosts and 32 compute units. The real story is not the cores but the platform design: a 256-bit memory bus and LPDDR5X support tuned specifically for AI PCs, mini PCs, and compact workstations that need serious on-device AI capability without discrete GPUs.
Why 192GB Unified Memory Changes Local AI Computing
The standout feature of Ryzen AI Max 400 is its support for up to 192GB of LPDDR5X-8533 unified memory, a 50% jump over the previous 128GB ceiling. This upgrade leverages new 24GB LPDDR5X chips and pushes bandwidth to around 273GB/s. For local AI computing, that capacity matters more than a small frequency bump. AI workloads—especially large language models and multi-agent systems—are often memory-bound. AMD says Gorgon Halo can allocate up to 160GB of that pool as VRAM, leaving 32GB for CPU tasks, enabling on-device inference of models that previously demanded cloud GPUs or multi-card desktop rigs. In practical terms, developers and power users can run larger context windows, more concurrent sessions, or multiple models side by side on a single AI PC, reducing offloading and making local-first AI workflows realistically viable.
Ryzen AI Halo: A Developer Platform for Local-First AI
To turn these silicon capabilities into real-world tools, AMD is building out its Ryzen AI Halo developer platform. The current Halo system is based on the Ryzen AI Max+ 395, offering up to 128GB of unified memory, Radeon 8060S graphics with 40 compute units, and an XDNA 2 NPU rated at 50 TOPS. The platform is geared toward developers building generative and agentic AI applications that run locally, with support for ROCm and mainstream AI frameworks. AMD plans to transition Halo to the Ryzen AI Max PRO 400 series in the third quarter, lifting the platform ceiling to 192GB and NPU throughput up to 55 TOPS. That shift should widen the types of models and complex multi-step agents that can be prototyped, fine-tuned, and deployed entirely on x86 client systems, shrinking reliance on remote infrastructure during development.
AI PC Buyers vs Cloud: Who Really Needs 192GB Memory Chips?
Despite the headline-grabbing 192GB memory figure, Ryzen AI Max 400 is not for every AI PC buyer. The benefits are clearest for a niche but growing audience: researchers, small teams, and advanced enthusiasts running sizable LLMs or multimodal models locally. For them, the ability to keep large models and long contexts on-device can reduce latency, protect data privacy, and cut ongoing cloud usage. For mainstream users, however, the modest performance uplift over Ryzen AI Max 300 and the likely premium of Halo-class systems mean cloud-based AI may remain more practical. Availability is another constraint: AMD expects OEM systems from brands like ASUS, HP, and Lenovo in the third quarter, while Halo developer platforms are only just entering preorder. Until such hardware is widely stocked, cloud GPUs and hybrid setups will continue to dominate heavy AI workloads.