AMD Ryzen AI Max PRO 400 Hits 192GB Memory for Local 300B-Parameter LLMs

Photo by Olivier Collet on Unsplash

AMD says the Ryzen AI Max 400 series is the first x86 client processor capable of running a 300-billion-parameter large language model in local memory, a claim it staked publicly two days before Computex 2026. The chip, codenamed Gorgon Halo, tops out at 192GB of unified memory with up to 160GB allocatable as VRAM, leaving 32GB for the operating system and background processes, according to AMD's announcement blog.

That ceiling is the entire story. Everything else about Gorgon Halo is a minor revision of the Strix Halo platform that shipped in early 2025.

What Changed and What Didn't in Gorgon Halo

The silicon is largely the same. Tom's Hardware characterizes Gorgon Halo as a minor refresh of Strix Halo, with identical Zen 5 CPU cores, RDNA 3.5 graphics, and the XDNA 2 NPU architecture, with the memory ceiling as the headline upgrade. The flagship Ryzen AI Max+ PRO 495 carries 16 Zen 5 cores, 40 RDNA 3.5 compute units, a 5.2 GHz CPU boost, and a 3.0 GHz GPU clock, representing a 100 MHz bump over the Max+ 395, Tom's Hardware reports.

The NPU gets a modest nudge. VideoCardz confirms the XDNA 2 block is now rated at 55 TOPS, up from 50 TOPS on the Ryzen AI Max+ 395, while the memory subsystem remains a 256-bit LPDDR5X-8000 interface.

Three PRO SKUs have been announced so far. VideoCardz lists the Ryzen AI Max+ PRO 495 with 16 cores and 40 CUs, the Max PRO 490 with 12 cores and 32 CUs, and the Max PRO 485 with partial details. Non-PRO consumer variants have not been confirmed. Tom's Hardware notes AMD is announcing the PRO 400 Series first, with non-Pro versions listed as coming soon.

Why the 192GB Unified Memory AI PC Matters

The 50 percent memory increase is doing work that the CPU and GPU updates cannot. Prior Strix Halo capped at 128GB of unified memory, and Gorgon Halo raises that ceiling to 192GB. For inference workloads, memory capacity is the hard constraint, not compute. A 300B-parameter model in 4-bit quantization needs roughly 150GB of memory just to load weights, before any KV cache or context window overhead.

That is the math behind AMD's positioning. AMD claims the Ryzen AI Max 400 is the first x86 client processor able to run a 300B+ parameter LLM, a claim Tom's Hardware notes wins in a category of one because Intel makes no comparable large SoC and Apple uses ARM. The competitive set is Apple Silicon and Nvidia's DGX Spark, not anything in AMD's own x86 lineup.

There is a buried trade-off here that the spec sheet doesn't advertise. LPDDR5X-8000 on a 256-bit bus delivers roughly 256 GB/s of bandwidth. A discrete RTX 5090 ships with 1,792 GB/s. Gorgon Halo can hold a 300B model in memory, but token generation on that model will be bandwidth-bound at a fraction of what a comparably-sized discrete GPU could push if it had the VRAM. The platform's value is fitting the model at all, not running it quickly.

That trade-off matters most for interactive workloads. Batch generation, embedding pipelines, and overnight document processing tolerate slow token rates. A coding assistant generating completions in real time does not. Buyers expecting cloud-grade responsiveness from a local 300B model will be disappointed regardless of how much memory the chip ships with.

The $3,999 Ryzen AI Halo Developer Box Has a Catch

AMD is shipping its own reference system. The Ryzen AI Halo developer box launches at $3,999 with pre-orders opening in June 2026, sold exclusively at Micro Center, according to AMD's announcement. The first wave does not use the new chip.

Read that sentence again. AMD confirms that the first wave of Ryzen AI Halo boxes ships with the prior-generation Ryzen AI Max+ 395 Strix Halo at 128GB, not the new Gorgon Halo chip. Buyers paying $3,999 in June get the 2025 silicon and the 128GB memory ceiling. HotHardware reports that a version of Ryzen AI Halo based on the new Ryzen AI Max+ PRO 495 is confirmed but has no ship date beyond at some point.

For the 128GB box, AMD states it can run models up to 200 billion parameters under company-tested conditions. Hitting the advertised 300B+ ceiling requires waiting for the Gorgon Halo variant, which has no confirmed release window.

HotHardware describes the Ryzen AI Halo as AMD's first-party answer to Nvidia's DGX Spark, under 1 liter in size, with Wi-Fi 7 and a 2TB PCIe 4.0 SSD. The operating system story is the cleanest differentiator: Ryzen AI Halo supports both Windows and Linux, while Nvidia's box is Linux-only.

The price-to-silicon mismatch is awkward. A developer spending $3,999 in June 2026 is paying flagship money for a 2025 chip that has been on the market for over a year, with the upgrade path gated by an unannounced ship date. AMD has not said whether early buyers get any path to the newer silicon.

AMD Ryzen AI Max PRO 400 Release Date and OEM Timing

OEM systems arrive later. AMD's announcement confirms ASUS, HP, and Lenovo will introduce systems with Ryzen AI Max PRO 400 starting Q3 2026, which puts laptops and mini-workstations into the holiday buying window. That is the realistic timeline for anyone who wants the new chip rather than the refurbished Strix Halo platform Micro Center is selling in June.

The predecessor's distribution history is the warning shot. HotHardware notes that Strix Halo had a niche rollout, available only in the Framework Desktop, ROG Flow Z13, and GMKtec EVO-X2, suggesting Gorgon Halo may see similarly limited early availability. Three OEMs naming the platform in May does not guarantee broad SKU coverage by Q3.

Memory supply is the other risk. Tom's Hardware raises a supply concern that global DRAM shortages are pushing prices up, calling consistent 192GB shipping a minor miracle given Apple's own struggles with high-memory configurations. A 192GB LPDDR5X module loadout per system is expensive even before margin. Expect the top SKU to be allocated, not stocked.

Gorgon Halo Specs and Software Support

The software stack is the part AMD got right. AMD confirms support for PyTorch, vLLM, llama.cpp, Ollama, ComfyUI, LM Studio, and AMD ROCm, with both Windows and Linux supported. That covers the actual tools developers use for local inference rather than a marketing list of frameworks nobody runs.

Here is the quick-reference breakdown of the flagship part, drawn from VideoCardz and Tom's Hardware:

• CPU: 16 Zen 5 cores / 32 threads, up to 5.2 GHz boost
• GPU: 40 RDNA 3.5 compute units, up to 3.0 GHz
• NPU: XDNA 2, 55 TOPS
• Memory: Up to 192GB LPDDR5X-8000 on a 256-bit bus, 160GB allocatable as VRAM
• Platform positioning: Local LLM inference up to 300B+ parameters (AMD claim)

The positioning AMD is leaning on is broader than benchmarks. AMD frames the platform as enabling Agent Computers, systems that understand prompts, plan actions, and execute tasks locally. Whether that framing sticks depends on whether developers ship agent applications that actually need a local 300B model rather than a 70B model with cloud fallback. Most current agent frameworks do not.

How Gorgon Halo Stacks Up Against Apple and Nvidia

The direct comparison AMD wants is against Apple Silicon. HotHardware notes AMD positions against Apple's Mac Mini M4 Pro, citing that Apple's lower memory ceiling limits local AI workload scale. Mac Studio with the M3 Ultra reaches 512GB in unified memory configurations, which puts Apple ahead on raw capacity at the high end. AMD's pitch is x86 compatibility, ROCm tooling, and dual-OS support, not memory leadership against Apple's biggest workstation.

Against Nvidia's DGX Spark, the comparison is more favorable. Spark targets the same local-inference developer segment but ships as a Linux-only appliance with a Grace ARM CPU and Blackwell GPU. The Ryzen AI Halo's x86 compatibility and Windows support is the practical advantage for shops with mixed Linux and Windows tooling.

The Intel comparison does not exist yet. Panther Lake and the rest of Intel's AI PC roadmap target the thin-and-light segment, not 192GB workstations. For broader context on that lineup, see our coverage of Intel Panther Lake and the Core Ultra Series 3 and the wider AI PC Hardware category.

What to Watch Between Now and Q3 2026

Three specific dates and decisions will determine whether Gorgon Halo lands or stays niche. The first is the June pre-order opening at Micro Center, which will indicate whether AMD allocated meaningful volume of the 128GB Strix Halo boxes or treated it as a developer-seeding gesture. The second is Computex itself, where OEM partners will name specific SKUs and pricing tiers rather than vague Q3 commitments. The third is the ship date for the PRO 495 version of the Ryzen AI Halo box, which AMD has not announced.

If the Gorgon Halo variant of the Ryzen AI Halo slips past Q4 2026, the 300B-parameter pitch becomes a 2027 story, and Apple's M5 Ultra or Nvidia's next Spark refresh will be in the picture. For more on the broader chip-economics question this raises, see our analysis of Cerebras' wafer-scale approach to inference and other coverage in the AI Infrastructure category.