Let's clear the smoke from the showroom floor. Whether you are building a Windows 11 local-AI workstation, rolling out developer boxes for an IT team, or comparing AMD's unified-memory mini PCs against Nvidia's DGX Spark and Apple's Mac Studio, the real story between 128 GB and 192 GB is not about raw speed. It is about capacity ceilings, firmware memory splits, memory-bandwidth physics, and how much you are willing to pay per gigabyte.
Decoding AMD's Naming Maze: Codename vs. Chip vs. Platform
Shoppers keep tripping over three distinct naming layers that AMD stacks on top of one another. Mixing them up is the easiest way to overpay for the wrong hardware:
- The Silicon Codename (
Strix Halovs.Gorgon Halo):
AMD's official product pages now list Strix Halo as the former codename for the Ryzen AI Max 300 Series, including the Ryzen AI Max+ 395, Max+ 392, Max+ 388, and Max 390. Engineers still use the codename freely—AMD's Ryzen AI Software 1.8.0 release notes reference Strix Halo alongside Phoenix, Hawk Point, and Krackan Point—so it has not vanished from technical documentation. However, the newer Ryzen AI Max+ PRO 495 belongs to the Ryzen AI Max PRO 400 Series, which AMD's documentation and driver reporting identify under the "Gorgon Halo" codename rather than Strix Halo. Treating the 495 as just another Strix Halo SKU misses a generational platform step. - The Processor SKU (
Ryzen AI Max,Max+, andMax PRO):
This is the chip name that actually determines CPU cores, integrated Radeon graphics, and maximum supported memory. One commercial wrinkle to watch: consumer retail listings promote the plain Ryzen AI Max+ 395, while enterprise workstations such as the HP Z2 Mini G1a ship with a Ryzen AI Max+ PRO 395 designation featuring manageability and ECC memory support. Always inspect the exact SKU and installed RAM in a vendor listing rather than assuming every "395" or "495" machine carries the same memory ceiling. - The Platform (
Ryzen AI Halo):
Ryzen AI Halo is not a processor. It is AMD's compact developer mini-PC platform built around Ryzen AI Max silicon, pre-validated for local AI workflows. The first-generation Ryzen AI Halo launched with the Ryzen AI Max+ 395 and 128 GB of unified memory, whereas the second-generation Halo swaps in the Ryzen AI Max+ PRO 495 and 192 GB of memory. Because AMD lists both configurations under the Ryzen AI Halo umbrella, checking the chip inside the box is mandatory before comparing quotes.
Ryzen AI Max+ 395 vs. Max+ PRO 495: Specs and What Actually Changed
When you line up the official specifications for the Ryzen AI Max+ 395, the lower-tier 392/388 parts introduced at CES 2026, and the flagship PRO 495, one pattern jumps off the page: CPU compute barely budged, while the memory subsystem took center stage.
| Specification | Ryzen AI Max+ 388 / 392 | Ryzen AI Max+ 395 | Ryzen AI Max+ PRO 495 |
|---|---|---|---|
| Series / Codename | Max+ 300 (Strix Halo) | Max+ 300 (Strix Halo) | Max PRO 400 (Gorgon Halo) |
| CPU Cores / Threads | 8C/16T (388) or 12C/24T (392) | 16 Zen 5 cores / 32 threads | 16 Zen 5 cores / 32 threads |
| Base / Max Boost Clock | Up to 5.0 GHz boost | 3.0 GHz base / Up to 5.1 GHz | 3.1 GHz base / Up to 5.2 GHz |
| Integrated Graphics | Radeon 8060S (40 CUs) | Radeon 8060S (40 CUs) | Radeon 8065S (40 CUs) |
| NPU Performance | Up to 50 TOPS | Up to 50 TOPS | Up to 55 TOPS |
| Peak FP16 Compute (Halo) | Not published | Up to 60 TFLOPS | Up to 61 TFLOPS |
| Max Unified Memory | Varies by OEM | Up to 128 GB | Up to 192 GB |
| Halo Memory Speed / Bandwidth | N/A | LPDDR5x-8000 / 256 GB/s | LPDDR5x-8533 / 273 GB/s |
| Max Windows GPU Allocation (VGM) | Varies by OEM | Up to 96 GB (on 128 GB Halo/HP) | Up to 160 GB (on 192 GB systems) |
| AMD Claimed Local LLM Ceiling | Small/mid-tier models | Up to 200B parameters | Up to 300B+ parameters (4-bit) |
| TDP Range | 45–120 W configurable | 55 W default (45–120 W) | 55 W default (45–120 W) |
| OS Support Listed | Windows 11 / Linux (OEM-dependent) | Windows 11, Ubuntu x86-64, RHEL x86-64 | Windows 11, Ubuntu x86-64, RHEL x86-64 |
Why 50% More Memory Does Not Mean 50% Faster Inference
Here is the trap buyers fall into when reading spec sheets: 192 GB vs. 128 GB is a 50% jump in capacity, not a 50% jump in tokens per second.
In local LLM inference, performance splits into two phases:
- Prompt Processing (Time to First Token / Prefill): Bound primarily by raw GPU compute (TFLOPS). Moving from 60 FP16 TFLOPS on the 395 Halo to 61 TFLOPS on the PRO 495 Halo is less than a 2% difference on paper.
- Token Generation (Decode Speed): Bound by how fast the GPU can stream model weights out of memory. On the Halo platform, memory moves from LPDDR5x-8000 (256 GB/s) on the 395 to LPDDR5x-8533 (273 GB/s) on the PRO 495.
Notice the math: 273 GB/s versus 256 GB/s is roughly a 6.6% to 7% increase in memory bandwidth. If you run a 32B or 70B model that already fits comfortably inside 128 GB, the PRO 495 will feel marginally snappier—not transformative. Where the 192 GB machine earns its keep is running models that outright crash or spill to sluggish disk swap on 128 GB.
Inside the 128 GB vs. 192 GB Memory Architecture (and Windows VGM Caveats)
Unlike a traditional desktop PC where your CPU owns DDR5 system RAM and a discrete RTX or Radeon card guards its own separate bank of GDDR6/GDDR7 VRAM, Ryzen AI Max uses a single unified pool of LPDDR5x memory shared across CPU, GPU, and NPU.
How much of that pool the integrated Radeon GPU can actually use for AI weights depends heavily on your operating system and OEM firmware:
- Windows 11 and Variable Graphics Memory (VGM):
Windows expects graphics memory to look like dedicated VRAM for many DirectX, Vulkan, and ROCm workloads. AMD's Variable Graphics Memory (VGM) setting carves out a dedicated block of system RAM exclusively for the GPU (leaving the remainder for Windows, background apps, and the CPU).- On a 128 GB workstation like the HP Z2 Mini G1a or 395 Halo, VGM lets you dedicate up to 96 GB to the GPU (with up to 112 GB total GPU-addressable memory when counting shared memory). However, compact laptops and tablets vary by OEM BIOS—some 128 GB portable designs cap dedicated graphics memory at 64 GB.
- On a 192 GB PRO 495 system, VGM scales up to 160 GB of dedicated graphics memory, leaving 32 GB for Windows 11 and system overhead. Crucially, AMD's Ryzen AI Halo User Guide notes that Windows 11 users who push VGM above 128 GB on the 192 GB platform must follow additional configuration steps rather than assuming a single click covers every workflow.
- On Halo hardware, VGM is exposed directly inside AMD Software: Adrenalin Edition (under Performance / Tuning) and inside the AMD Ryzen AI Developer Center (Settings → Graphics Performance Settings), which splits controls into Dedicated Video Memory and Shared Video Memory (adjustable from 10% to 90%, defaulting to 75%, with changes applied after a reboot).
- Linux Memory Allocation:
Linux skips the rigid Windows VGM carve-out. Instead, the kernel driver maps memory dynamically on demand from the shared pool. By default, the Linux kernel's TTM limit caps GPU-mappable memory at roughly half of system RAM, though developers routinely raise that ceiling using AMD'samd-ttmutility and a reboot (for instance, freeing up 100 GB+ on a 128 GB machine or 160 GB+ on a 192 GB machine).
Model Sizing Reality Check: Why "300B Parameters" Needs Context
AMD advertises local support for up to 200B-parameter models on the 128 GB Halo and 300B-parameter and larger models at 4-bit quantization on the 192 GB PRO 495. Take those figures as theoretical weight ceilings at modest context lengths, not blank checks for 128K-token sessions.
Model weights are only part of your memory bill. On top of the GGUF or quantized weight footprint, you must budget for:
- The KV (key-value) cache, which balloons rapidly as prompt and context lengths climb toward 32K, 64K, or 128K tokens.
- Runtime buffers, compute graphs, and OS overhead.
- Concurrent workflows—if you run a local IDE assistant, a vector database, and an LLM simultaneously, they all sip from the same unified bucket.
For a concrete reference point, AMD's LM Studio guidance for GPT-OSS-120B (MXFP4) lists a 63.39 GB model download and recommends a 128 GB system, while AMD's developer community documented quantizing a 35B mixture-of-experts (MoE) model to W4A16 on a 128 GB Max+ 395 using llama.cpp, vLLM, and Lemonade. If your daily driver models live in the 7B to 70B or 120B MoE range, 128 GB fits them with room to breathe. Once you step up to 150B–300B+ MoE models (such as full-precision or higher-bit Qwen, DeepSeek, or GLM variants approaching 150–160 GB just for weights), 128 GB hits a brick wall and 192 GB becomes essential.
Ryzen AI Halo, Project Zenith, and the Software Stack
Hardware capacity is half the battle; on Windows and Linux, AMD has spent late 2026 racing to close the software convenience gap that long favored Nvidia's CUDA.
Ryzen AI Halo Meets Microsoft Project Zenith on Windows 11
On October 12, 2026, AMD unveiled the second-generation Ryzen AI Halo powered by the PRO 495, followed on November 9, 2026, by commercial availability of the PRO 495 Halo with Microsoft Project Zenith preinstalled (available through Micro Center online and retail locations, alongside Linux-preloaded PRO 495 Halo listings such as Micro Center SKU CPL-MFGB0810001).
Project Zenith—first announced by Microsoft and AMD at IFA 2026 in early September—is not a separate Windows edition. It is a Microsoft-curated, "ready-to-code" Windows 11 baseline targeted at developer-class PCs with at least 64 GB of unified memory and 250+ GB/s of memory bandwidth. Out of the box, Project Zenith ships with:
- Core developer tools preinstalled and configured: Visual Studio Code and Windows Terminal pinned to the taskbar, Git, GitHub CLI, GitHub Copilot CLI, Azure CLI, Python, Node.js,.NET, PowerShell 7, and Windows Subsystem for Linux (WSL).
- AMD Ryzen AI Developer Center: A unified Windows and Linux dashboard linking to curated AMD AI Playbooks (ComfyUI, vLLM, Lemonade Server, LM Studio, n8n + gpt-oss-120B workflows), firmware/BIOS updates, VGM sliders, and a Best Known Configuration (BKC) sync button. That BKC tool is a lifesaver for IT teams and developers: if a PyTorch or driver update breaks a local workflow, one click rolls the entire validated stack back to a known-good combination.
- AMD Sync: A built-in companion utility offering streamlined SSH access, remote terminals, code editing, and live telemetry from another Windows or Linux machine.
Existing Halo owners do not need to buy a new box to mirror Project Zenith's setup—AMD notes the software baseline can be pulled using Microsoft's Windows Developer Config (WinGet) via GitHub, and Project Zenith is slated to expand to additional OEM Ryzen AI Max devices.
Four Inference Paths: Vulkan, ROCm, FastFlow HIP EP, and the XDNA 2 NPU
On Ryzen AI Max hardware, developers now choose among four distinct inference runtimes:
llama.cppvia Vulkan: The universal, low-friction baseline across Windows and Linux. Easy to build, broadly compatible, and reliable as a fallback.- ROCm / HIP (
gfx1151target): AMD's compute stack powering PyTorch andvLLM, heavily documented for Linux (Ubuntu and RHEL x86-64) and expanding on Windows 11. - FastFlowLM (FastFlow) v1.0.13 + HIP EP (Early Access): Released November 13 and detailed by AMD on November 19, 2026, FastFlow adds a new GPU backend (
oga-hipep, built on ONNX Runtime GenAI and HIP EP v0.5.3) alongside its existing NPU runtime inside a ~17 MB package. Using three commands (flm pull,flm run,flm serve), it exposes an OpenAI-compatible API supporting 20+ model architectures and context lengths up to 128K tokens. In AMD's own measurements on a Ryzen AI Max system at 16K context, HIP EP achieved ~43% faster time-to-first-token and ~7% higher decode throughput thanllama.cppon Vulkan (recording speeds such as 26.10 tok/s on Qwen2.5 14B Instruct, 67.04 tok/s on Qwen3.5 35B-A3B, and 85.96 tok/s on GPT-OSS 20B at 2,048-token prompts). Note AMD's caveat: the GPU path is explicitly early access, and those numbers are vendor HIP EP benchmarks rather than independent third-party reviews. - Unsloth Desktop + XDNA 2 NPU: Announced November 19, 2026, AMD and Unsloth made XDNA 2 the first NPU supported by Unsloth Desktop on both Windows 11 and Linux. Using Lemonade and FastFlowLM under the hood with
Q4NX4-bit models, it serves an OpenAI-compatible local API for over 100 agentic tools—including Cline, Open WebUI, GitHub Copilot, Hermes Agent, and OpenClaw. Unsloth demonstrated a Qwen3.6 35B A3B MoE model running on the NPU at 17.2 tok/s on a 64 GB system while claiming 2x to 4x lower energy per generated token than traditional GPUs—ideal for persistent background agents on battery or low-noise desktops.
On the driver side, Windows users should verify they are running AMD Software: Adrenalin Edition 26.10.1 WHQL (or newer), which formally covers both the Ryzen AI Max 300 and 400 Series and adds FSR Upscaling 4.1 support.
Pricing Reality and OEM Systems: What 192 GB Actually Costs You
AMD does not publish official consumer list prices for the Ryzen AI Max+ 395, Max+ PRO 495, or the Ryzen AI Halo developer platform. Instead, OEMs and boutique mini-PC makers set street prices—and in late 2026's tight LPDDR5x memory market, the jump from 128 GB to 192 GB carries a steep sticker shock.
Look at how real OEM systems line up across both tiers:
- 128 GB Tier (Broad OEM Choice & Better $/GB):
- First-Gen AMD Ryzen AI Halo (Max+ 395, 128 GB, 2 TB SSD): Launched at $3,999 in Windows 11 Pro or Linux configurations.
- HP Z2 Mini G1a (Max+ PRO 395, 128 GB ECC, Windows 11 Pro): Enterprise compact workstation priced around $3,700–$4,600 depending on storage and retailer promotions, featuring up to 96 GB VGM and dual NVMe bays.
- Lenovo ThinkCentre X Ultra (Max+ PRO 495, 64 GB / 128 GB): A fascinating hybrid—Lenovo pairs the newer PRO 495 silicon with up to 128 GB of LPDDR5x-8533 (not 192 GB) in a 1.6-liter chassis starting at $3,699, featuring dual 10 GbE ports designed to cluster up to four units together.
- Thin-and-light laptops & tablets: Devices like the ASUS ROG Flow Z13 (128 GB, typically capped at 64 GB graphics allocation) bring Strix Halo silicon into portable form factors, though sustained 120 W workloads depend heavily on chassis thermals.
- 192 GB Tier (Workstation & Boutique Mini-PC Territory):
- TOPC TXA2 (PRO 495, 192 GB, 2 TB SSD): Listed around $6,499.
- GMKtec EVO-X5 Pro (PRO 495, 192 GB): Launched around $6,599.
- Framework Desktop (PRO 495, 192 GB): Opened pre-orders at $6,799.
- HP ZBook Ultra G3a (PRO 495, 192 GB laptop): Debuted at $7,449.
- Minisforum MS-S1 MAX-P495 (PRO 495, 192 GB): Positioned between roughly $7,400 (limited pre-order pricing) and $9,200+.
Do the quick cost-per-gigabyte math (always including tax and shipping). At ~$3,700–$4,000, a 128 GB machine works out to roughly $29–$31 per GB of unified memory. At ~$6,500–$6,800, a 192 GB machine jumps to roughly $34–$35 per GB—an extra $2,500 to $3,000+ upfront just for the final 64 GB of capacity.
How Ryzen AI Max Compares Against Nvidia DGX Spark and Apple Mac Studio
Because independent matched benchmarks across all platforms remain sparse, weigh the architectural trade-offs carefully:
- Vs. Nvidia DGX Spark / RTX Spark: Nvidia owns the CUDA software ecosystem, meaning custom Python libraries, training scripts, and enterprise tutorials work out of the box with minimal porting friction. AMD counters with standard x86-64 Windows 11 and Linux compatibility, a wider pool of competing OEMs, and—on the PRO 495—up to 192 GB of unified memory versus 128 GB on standard Spark boxes.
- Vs. Apple Mac Studio (M5 Ultra): Apple offers whisper-quiet acoustics, polished macOS integration, and substantially higher memory bandwidth on Ultra-tier silicon. AMD counters with Windows 11, Ubuntu, and RHEL support, native x86 tooling, and gaming/CAD compatibility through Radeon graphics.
Practical Pre-Purchase & Migration Checklist for Windows and Linux Users
Before clicking "buy" on any Ryzen AI Max system—and during your return window—protect yourself with a disciplined verification routine:
- Confirm the exact memory capacity and Windows VGM ceiling in writing. A 128 GB mini PC with 96 GB VGM behaves very differently from a 128 GB tablet capped at 64 GB, and a PRO 495 mini PC like Lenovo's ThinkCentre X Ultra may ship with 128 GB rather than 192 GB.
- Audit your CUDA dependencies. If your workflow relies on
llama.cpp, Ollama, LM Studio, orvLLM, migration to Vulkan, ROCm, or FastFlow is mostly a rebuild and model-file copy. If your pipeline imports custom CUDA-only kernels, budget engineering hours for porting before retiring your Nvidia rig. - Check warranty and service terms. Because unified LPDDR5x memory is soldered, a RAM fault means replacing the entire mainboard. Enterprise buyers get stronger on-site warranties from HP or Lenovo than from small-batch mini-PC brands.
- Benchmark your exact workload during the return window. Don't trust synthetic peak TFLOPS. Load your target quantized GGUF model, verify the OS sees the allocated memory (
rocminfoandfree -gon Linux, or Task Manager / Adrenalin Performance on Windows), warm up with a 2,000-token prompt, and measure both Time to First Token and decode tokens/sec at 4K and 16K context lengths across Vulkan, ROCm, and FastFlow HIP EP.
Verdict: Should You Buy 128 GB or 192 GB?
- Buy the Ryzen AI Max+ 395 (or a 128 GB PRO 495 system) if your models sit at 120B parameters or below. For 7B–70B dense models, 35B MoE coding assistants, creative 3D/CAD rendering, and everyday local AI agents, 128 GB hits the sweet spot on price-per-gigabyte and OEM warranty coverage. Paying $2,500+ extra for 192 GB buys headroom that will sit idle—and only ~7% more memory bandwidth.
- Pay up for the 192 GB Ryzen AI Max+ PRO 495 (or PRO 495 Halo) only when capacity is non-negotiable. If you actively deploy 150B–300B+ quantized MoE models locally, juggle multiple heavy models concurrently, or need long-context sessions that blow past 96 GB of VRAM, buying 192 GB on day one is far cheaper than replacing a soldered 128 GB machine six months later.
- Approach the Max+ 388 and 392 cautiously. With 8 and 12 CPU cores respectively, they make sense for lightweight local NPU/GPU tasks only when an OEM pairs them with generous memory at a steep discount below 395 pricing.
In the unified-memory AI era, buy for the model size you cannot fit anywhere else, verify the graphics memory split before checkout, and never mistake extra gigabytes for extra speed.
References
- Ryzen AI Max 192GB vs 128GB: Bigger Wins for LLMs - shattered.io shattered.io · 2026-11-21T20:19:49+00:00
- minisforum.com minisforum.com
- AMD Ryzen™ AI Halo User Guide | AMD AI Playbooks developer.amd.com