AMD Ryzen AI Max PRO 400: 192GB RAM for 300B Local Models

AMD opened IFA 2026 in Berlin with a chip that forces a straight answer about where on-device AI actually lives. The Ryzen AI Max PRO 400 series, codenamed Gorgon Halo, places 192GB of unified memory on a single x86 die and claims it can run 300 billion parameter models without the cloud. That headline is real silicon, but it hides a supply chain problem that decides whether you will ever actually buy one.

The launch is not a compute redesign. Gorgon Halo reuses the Zen 5 CPU, RDNA 3.5 GPU, and XDNA 2 NPU from the previous Strix Halo generation and changes the memory system, per AMD's IFA 2026 announcement at amd.com. That distinction is the whole story, and it matters more than most coverage has let on.

Why unified memory changes the arithmetic of local inference

Running a large language model locally is almost entirely a memory problem, not a compute problem. During inference the GPU reads weights over and over to generate one token at a time. It does relatively little matrix math compared to training. The binding constraint is how much memory the GPU can see and how fast it can read.

A typical laptop GPU exposes 8 to 24GB of VRAM. A 70 billion parameter model at FP16 precision needs roughly 140GB, which no portable card can hold. At 4-bit quantization the same model drops to about 38 to 40GB, which fits a workstation card but still beats most portable hardware.

Gorgon Halo solves the capacity problem by having the CPU and GPU share one 192GB pool of LPDDR5X-8533 memory on the same die. Up to 160GB can be handed to the integrated GPU as VRAM. Because there is no PCIe bus between CPU and GPU, the system avoids the transfer penalty that makes discrete laptop GPUs so limiting. The LPDDR5X-8533 interface runs across four channels at a 256-bit width, delivering about 273 GB/s of bandwidth.

The math for AMD's claim is transparent. At 4-bit, a 300 billion parameter model needs roughly 150GB of weights. That leaves about 10GB in the 160GB VRAM allocation for the key-value cache that grows with context length. It is a tight fit, but it works on paper.

The three SKU lineup

AMD announced three processors, and only the top one gets a silicon bump.

Ryzen AI Max+ PRO 495: 16 Zen 5 cores (32 threads), 3.1GHz base to 5.2GHz boost, Radeon 8065S GPU with 40 compute units at 3.0GHz, 55 TOPS NPU, 80MB cache, 45 to 120 watts TDP.
Ryzen AI Max PRO 490: 12 cores, 3.2GHz base to 5.0GHz boost, Radeon 8050S GPU with 32 compute units, 50 TOPS NPU, 76MB cache.
Ryzen AI Max PRO 485: 8 cores, 3.6GHz base to 5.0GHz boost, Radeon 8050S GPU with 32 compute units, 50 TOPS NPU, 40MB cache.

All three support 192GB of unified memory and the 160GB VRAM allocation. The GPU and NPU on the 490 and 485 are unchanged from the equivalent Strix Halo parts. AMD standardized the PRO feature set across the series, adding remote management similar to Intel vPro. Only PRO, enterprise versions exist at launch, so consumer buyers have no confirmation yet.

The memory bandwidth ceiling that no one can hide

273 GB/s is meaningfully lower than the 1 to 2 TB/s of HBM3 bandwidth in AMD's own Instinct MI300X data center accelerators, and nowhere near the 5.3 TB/s some HBM3e configurations reach. That is why Gorgon Halo is an inference platform, not a training platform. LPDDR5X can stream large quantized weights at a usable token rate, but it cannot keep pace with training-class memory.

This is the honest gap that card-based coverage tends to skip. AMD has demonstrated the model will fit and run, but independent inference benchmarks at production scale had not published when this article was written. The claim is about capacity, not speed.

The NVIDIA comparison, honestly

AMD launched into a market NVIDIA has already occupied. NVIDIA's DGX Spark ships with a 20-core ARM CPU and Blackwell GPU in the GB10 superchip, 128GB of unified memory at 273 GB/s, and 1 petaFLOP of FP4 sparse compute. NVIDIA says it can run up to 200 billion parameter models, and the Founders Edition starts at $4,699. An RTX Spark variant targets Windows laptops and compact desktops in fall 2026.

On raw capacity AMD leads: 192GB versus 128GB for both NVIDIA systems, and the 160GB VRAM allocation also exceeds NVIDIA. On training throughput NVIDIA keeps an edge, because Blackwell has dedicated hardware for FP4 and FP8 sparsity that RDNA 3.5 lacks. The CUDA ecosystem also carries a decade of developer tooling ahead of it.

The x86 foundation has its own weight here. Windows legacy software, enterprise applications, and the large library of x86-compiled inference frameworks run natively on Gorgon Halo. NVIDIA's ARM systems sometimes need an emulation layer for certain workloads.

Why the DRAM shortage matters more than the spec sheet

Here is the part AMD never puts on a slide. A global DRAM shortage has intensified through 2026, driven by AI infrastructure build-out and a shift by Samsung and SK Hynix toward HBM production for data center GPUs. LPDDR contract prices rose roughly 146 percent year-over-year through mid-2026 according to Supplyframe Commodity IQ data, and analysts at Nanya expect the shortage to run through at least the first half of 2027, as detailed in Tech Times' coverage.

The comparison to Apple is instructive. Apple removed the 512GB and 128GB Mac Studio configurations because of memory module supply constraints. It was the same SK Hynix 24GB LPDDR5X packages that AMD's 192GB Gorgon Halo configurations require.

So the 192GB specification is real at the silicon level. Whether 192GB systems ship at volume, at accessible prices, and within the Q3 2026 window depends on supply chain dynamics that remain unsettled. Buyers should watch OEM pricing before assuming the headline translates to a purchasable product.

What this means for builders and data scientists

The practical question is whether a 300 billion parameter model at 4-bit quantization on local LPDDR5X beats a smaller local model or a larger cloud model. The answer depends entirely on the use case.

For regulated industries where data must not leave the premises, healthcare systems, legal firms, and financial institutions, a local 300B model can replace cloud API dependencies entirely. For agentic workflows that run multiple models concurrently and query each other, the 192GB pool provides headroom no discrete laptop GPU can match. For developers iterating on large models without per-token cloud costs, local inference at this scale cuts operating costs.

The caveat is that 4-bit quantization introduces measurable quality loss on reasoning-intensive tasks compared to FP16. AMD claims the platform can run the weights, not that quantized inference equals full precision on a data center cluster. Teams where reasoning accuracy matters most should weigh the quantization tradeoff carefully.

Availability and outlook

AMD confirmed commercial systems from HP and Lenovo, targeting Q3 2026, before the end of September. Form factors span commercial AI PCs, mobile workstations, and compact small-form-factor desktops. HP's President of Personal Systems Ketan Patel said the company is designing systems for "the next generation of agentic AI-driven workflows." Lenovo's Luca Rossi framed it as moving from the cloud to where work actually happens.

The developer-focused Ryzen AI Halo desktop launched in June using the prior-gen Ryzen AI Max+ 395 with 128GB, sold exclusively through Micro Center. Updated PRO 400 units are expected later this year from more OEM partners, but AMD confirmed no consumer version announcement yet. Developers who want this platform today must go through enterprise channels.

The 400 series is a memory-system refresh that reuses proven compute silicon, which is both its strength and its limitation. It proves x86 can host genuinely large local models, and it puts AMD and NVIDIA on a clear collision course for compact AI hardware. Whether that proof reaches actual buyers before the DRAM shortage tightens further is the question to watch.