AMD’s IFA 2026 Mini PCs Bring 192GB Local AI to the Desktop

IFA 2026 showed AMD partners pursuing two sharply different kinds of compact desktop. At one end are Gorgon Point systems such as the GEEKOM A9 Max and Acemagic’s AMD F2A: conventional mini PCs with strong integrated graphics, NPUs and familiar upgrade paths. At the other are Gorgon Halo machines from Acemagic and Minisforum, which use up to 192GB of unified memory to make much larger local AI workloads feasible. These are plugged-in desktops and NAS appliances, so laptop measures such as battery life, display quality and charging are not relevant.

The short version: mainstream mini PC or local-AI workstation?

• Choose Gorgon Point for a compact productivity PC, light creation, development and moderate local AI. Its strengths are smaller size, conventional DDR5 expansion and much lower pricing.
• Choose Gorgon Halo when memory capacity is the constraint: large quantized language models, longer contexts, document collections and local AI experimentation. Its trade-off is soldered memory and workstation-class pricing.

GEEKOM A9 Max: the upgradeable Gorgon Point option

The GEEKOM A9 Max is a compact mini PC based on AMD’s Ryzen AI 9 HX 470. That chip has 12 cores and 24 threads—four Zen 5 and eight Zen 5c cores—plus Radeon 890M integrated graphics with 16 RDNA 3.5 compute units and a 55-TOPS XDNA 2 NPU. 1823

Its appeal is conventional expansion rather than exotic unified memory. Tested configurations pair 32GB DDR5-5600 with a 2TB PCIe 4.0 M.2 2280 SSD and an open M.2 2230 slot; reviews list two SO-DIMM slots and up to 128GB of memory. 18 GEEKOM’s current product page advertises dual SO-DIMM and dual M.2 support, with capacity claims up to 256GB DDR5 and 8TB of storage, so buyers should confirm the supported maximum for the exact regional SKU before purchasing. 19

Connectivity is unusually generous for the size: two 40Gbps USB4 ports, two HDMI 2.1 outputs, dual 2.5GbE, Wi-Fi 7, Bluetooth 5.4 and an SD 4.0 reader are listed by GEEKOM. 19 The chassis measures 5.31 × 5.20 × 1.81 inches. 19

GEEKOM lists the A9 Max at $1,299–$1,599, while a reviewed 32GB/2TB configuration was priced at $1,599. 1819 That makes it a practical small desktop for multi-display office work, home-lab duties, local development, lighter content work and modest on-device AI—not a substitute for a 192GB local-model workstation.

Acemagic’s split showcase: F2A, F7A and F9A

F2A: Intel Panther Lake or AMD Gorgon Point

Acemagic displayed the F2A in Intel and AMD versions, making it one of the clearer direct platform comparisons at IFA.

Model	Processor and graphics	AI claim	Memory status
F2A Intel	Intel Core Ultra X7 358H, 16 cores/16 threads, up to 4.8GHz; Arc B390 integrated graphics	Up to 180 TOPS combined platform AI	Acemagic reporting says 32GB onboard LPDDR5X; other showcased configurations cited up to 64GB, so final retail SKUs should be checked. 211
F2A AMD	Ryzen AI 9 HX 470, 12 cores/24 threads, up to 5.2GHz; Radeon 890M	55 TOPS NPU; up to 86 TOPS for the complete platform	32GB onboard LPDDR5X in the IFA reporting. 2

There is no credible performance verdict yet. Intel’s 180-TOPS and AMD’s 86-TOPS figures are combined CPU/GPU/NPU claims, while AMD’s widely cited 55-TOPS figure refers specifically to its NPU. They are not equivalent measurements of token generation, prompt processing, gaming, or any other single workload. 2 Independent testing will need to establish how the systems behave with particular models, precision formats, drivers, power limits and inference backends.

F7A: a less-defined compact model

Acemagic also showed the F7A as part of its Panther Lake and Gorgon Point mini-PC lineup. The available IFA reporting does not establish final worldwide configurations, detailed I/O, price or release timing, so it is best viewed as an early product-family reveal rather than a finished buying option. 2

F9A: a roughly 2-liter Gorgon Halo workstation

The Acemagic F9A is the more consequential reveal. It is a roughly 2-liter AI/gaming mini workstation built around AMD’s Ryzen AI Max+ PRO 495: 16 Zen 5 cores, 32 threads, boost clocks up to 5.2GHz, and Radeon 8065S graphics with 40 RDNA 3.5 compute units at up to 3.0GHz. 134

The defining specification is up to 192GB of LPDDR5X-8533 unified memory over a 256-bit interface. Reported connectivity includes dual USB4, dual 2.5GbE, HDMI 2.1, DisplayPort 2.1, Wi-Fi 7 and PCIe 4.0 x4 OCuLink for an external GPU. Storage reporting points to an M.2 2280 PCIe 4.0 x4 slot, while other coverage describes dual-SSD capability; final configuration details should therefore be confirmed at launch. 91113

Acemagic’s show-floor price indication was around €8,000, and final retail dates, regions and SKUs were still unsettled in the available reporting. 1315

Minisforum’s Gorgon Halo systems: workstation and AI-agent NAS

MS-S1 MAX-P495

Minisforum’s MS-S1 MAX-P495 is the most thoroughly specified Gorgon Halo desktop shown at IFA. It uses the same Ryzen AI Max+ PRO 495 and supports up to 192GB LPDDR5X unified memory, of which up to 160GB can be allocated as graphics memory. Minisforum rates the system at up to 55 TOPS on the NPU and lists 120W sustained and 160W peak operating modes. 50

This is not a low-power mini PC simply because it is physically compact. It is designed for sustained local AI and professional workloads, with a cooling and power design intended to support the APU under heavier desktop loads. 50

The expansion story is stronger than on most mini PCs: reporting describes dual M.2 storage slots, including PCIe 4.0 x4 and x1 connections, plus a full-length PCIe x16 physical slot operating at PCIe 4.0 x4. Reported I/O includes dual 10GbE, USB4 and USB4 v2, Wi-Fi 7 and HDMI 2.1. 6061

IFA price indications varied. One report cited roughly €5,000–€6,000, while others referenced pricing around €7,000 or noted that the final price was not settled. Treat all of these as preliminary rather than a final MSRP. 35156

N5 MAX-P495

The N5 MAX-P495 applies the same processor and up-to-192GB unified-memory concept to an AI-agent NAS. Minisforum says the N5 and MS-S1 use Ryzen AI Max+ PRO 495 with Radeon 8065S graphics, up to 131 TOPS of combined AI performance and up to 160GB of graphics-memory allocation. 62

That combination is notable because it can place models, embeddings, documents and shared storage on the same local device. It is a potentially useful layout for a small team or home lab that wants local retrieval-augmented generation, document search or agent workflows alongside network storage. However, detailed bay configuration, expansion, final pricing and availability should remain provisional until Minisforum publishes final specifications. 5457

Gorgon Point is a refined Strix Point, not a clean-sheet leap

AMD’s Ryzen AI 400 “Gorgon Point” family retains the basic Zen 5/Zen 5c CPU mix, RDNA 3.5 graphics and XDNA 2 NPU architecture associated with Ryzen AI 300/Strix Point. The changes are evolutionary: higher clocks, faster memory support on top models and faster NPU bins. 394043

For example, the Ryzen AI 9 HX 470 in the A9 Max and AMD F2A has 12 cores/24 threads, a 5.2GHz boost clock, 16 Radeon 890M compute units and a 55-TOPS NPU. The higher HX 475 reaches 60 NPU TOPS; it is the chip with the biggest NPU uplift, not the HX 470. 3640

In short, Gorgon Point is a sensible refresh for mainstream AI PCs. It improves the platform, but it does not change the category in the way Gorgon Halo’s memory design does.

Why 192GB unified memory matters for local AI

Gorgon Halo’s differentiator is not its NPU rating alone. The Ryzen AI Max+ PRO 495 combines a 16-core CPU, a 40-CU Radeon 8065S GPU and a unified LPDDR5X-8533 memory pool of up to 192GB. 350

A large shared pool gives the CPU, GPU and NPU access to the same general memory space. In a local AI workflow, that can make room for model weights alongside context, KV cache, embeddings, source documents, application memory and operating-system overhead. The practical benefits are primarily about fitting larger jobs locally:

• larger quantized local language models than conventional mini PCs can comfortably host;
• longer contexts and larger KV caches for multi-document analysis;
• offline retrieval-augmented generation, where documents and embedding indexes remain on local storage;
• coding assistants, extraction and summarization pipelines, and multi-step agents that need persistent context;
• image-generation or GPU-accelerated workflows that benefit from more GPU-addressable memory.

AMD partners quote up to 160GB allocatable as graphics memory, but that does not create 160GB of dedicated VRAM in addition to 192GB of system memory. It is one shared pool, so the operating system and CPU workloads still require capacity. 50

The other necessary caveat is performance: a machine that can load a large model is not automatically fast at running it. Tokens per second, prompt ingestion, maximum usable context, thermals, power draw, concurrent users, quantization and software support all matter. Claimed combined figures such as 131 TOPS are not a substitute for model-specific benchmarks. 5055

The benefits—and cost—of soldered LPDDR5X

Soldered LPDDR5X enables compact systems to offer 192GB capacity and a wide 256-bit memory interface, while keeping a shared CPU/GPU memory design. 3 That is the technical foundation for the F9A and MS-S1 MAX-P495’s local-AI proposition.

The compromise is permanence. Unlike the A9 Max’s SO-DIMMs, this memory is not a practical user upgrade or replacement path. Buyers must choose their capacity at purchase, and the 192GB configuration is a major contributor to workstation-level pricing. With early indications of €5,000–€6,000 for the Minisforum system and around €8,000 for Acemagic’s F9A, these are specialized machines rather than value-oriented mini PCs. 315

What AMD’s IFA lineup means

The most important IFA 2026 message was not another race for headline TOPS. It was a segmentation of local AI hardware:

• GEEKOM A9 Max: a comparatively affordable, compact and upgradeable Gorgon Point desktop.
• Acemagic F2A: a forthcoming Intel-versus-AMD mini-PC choice whose performance claims cannot yet be compared directly.
• Acemagic F9A and Minisforum MS-S1 MAX-P495: premium Gorgon Halo workstations intended to push local model capacity to 192GB.
• Minisforum N5 MAX-P495: the same high-memory concept adapted to local AI plus shared storage.

For most buyers, the A9 Max class is the more rational general-purpose PC. The 192GB Gorgon Halo systems matter when a workload genuinely cannot fit comfortably in a mainstream machine’s memory—and when the buyer accepts that early price, availability and performance claims still need verification through independent testing.