HyperAIHyperAI

Command Palette

Search for a command to run...

2 months ago
AMD
xAI
Anthropic

AMD Ryzen AI Max 400 'Gorgon Halo' packs 192GB unified memory

AMD has officially introduced the Ryzen AI Max 400 series, codenamed Gorgon Halo, a refreshed lineup of large System-on-Chips (SoCs) built on Zen 5 CPU architecture, RDNA 3.5 graphics, and XDNA 2 Neural Processing Units. While this range is a minor evolutionary update to the previous Strix Halo chips, it introduces a significant hardware advantage: support for up to 192GB of unified memory. This specification positions the Gorgon Halo as the first x86 client processor capable of running large language models with over 300 billion parameters locally. The lineup currently comprises three Pro models targeting the commercial market: the Ryzen AI Max+ Pro 495, the Pro 490, and the Pro 485. The flagship Pro 495 features 16 CPU cores, 32 threads, and a boost clock of 5.2 GHz, paired with a 40-core Radeon 8065S GPU. The Pro 490 and Pro 485 variants offer 12 and 8 cores respectively, both utilizing 32-core Radeon 8050S graphics. Despite the different core counts, all three models support the same massive 192GB memory limit. Of this total, 160GB can be allocated as VRAM for AI workloads, while the remaining 32GB is reserved for the system. AMD has indicated that a consumer version of Gorgon Halo may follow, but currently, the focus is on enterprise-grade security and manageability. In terms of immediate availability, the Gorgon Halo chips themselves will be featured in systems expected to launch starting in the third quarter of 2026. However, AMD is already selling its Ryzen AI Halo desktop box, a compact device measuring 5.9 by 5.9 by 1.7 inches, which utilizes the predecessor Ryzen AI Max+ 395 processor. Pre-orders for the box open in June with a starting price of $3,999 for a configuration featuring the Max+ 395, 128GB of unified memory, and 2TB of storage. AMD positions the Ryzen AI Halo box as a cost-effective alternative to cloud computing and competitor hardware like Nvidia's DGX Spark. While the Nvidia box costs $4,700 and includes a GB10 chip, the AMD solution runs on Windows and Linux, whereas the DGX Spark is Linux-only. Benchmarks suggest the Ryzen AI Halo delivers up to 14% higher tokens per second than the DGX Spark when running the GLM 4.7 Flash 30B model. Furthermore, AMD claims a single unit can save users up to $750 monthly compared to cloud services, achieving a return on investment within six months assuming high daily token usage typical of AI agents. The Ryzen AI Halo box is equipped with Wi-Fi 7, Bluetooth 5.4, and a 10Gbps Ethernet port, along with multiple USB-C ports and HDMI 2.1b output. With a rated thermal design power of 120W, the device packs substantial computational power into a small form factor. Although the full Gorgon Halo ecosystem rollout remains unclear beyond the Q3 2026 system announcement window, the new memory capabilities and competitive pricing of the existing box highlight AMD's strategic push into the localized AI inference market.

Related Links