AMD Strix Halo 128GB Local AI Desktop

Description

This is D-Central’s 128 GB AMD Strix Halo local-AI desktop path: a system built around the AMD Ryzen AI Max+ 395 with Radeon 8060S integrated graphics, sourced and first-boot configured in Quebec for $7,349 CAD before applicable tax and shipping. AMD designs the processor, integrated graphics and ROCm software ecosystem. The final computer comes from the OEM or system supplier named in the order confirmation. D-Central independently sources that system and performs the agreed setup; we do not manufacture AMD silicon, claim to be AMD, or rebrand an OEM chassis as our own.

What the $7,349 CAD product covers

• A current 128 GB system using the AMD Ryzen AI Max+ 395 and Radeon 8060S, with the exact brand, chassis, storage, ports, power adapter and included accessories confirmed before sourcing is committed.
• Your agreed Windows 11 or compatible Linux starting point, subject to the chosen OEM’s supported operating-system path and licensing.
• First-boot checks and an agreed local-inference setup using compatible versions of Ollama, llama.cpp or another specifically named runtime.
• A private browser interface when requested and compatible, plus a written record of the installed software versions and a model-fit discussion for your stated workload.
• Quebec configuration and handoff. Ongoing management, data migration, custom RAG ingestion, training, fine-tuning, identity integration and production SLA are separate work unless included in writing.

This page is not an inventory promise and does not silently commit you to a “similar” system. Strix Halo appears in different computers with different storage, cooling, firmware, networking and I/O. The product family and 128 GB capacity are fixed; the exact deliverable must be written into the order record. This listing is also not a representation that the buyer will receive AMD’s own Halo developer platform or a particular US-retail-exclusive SKU. If a specific Framework, Minisforum, HP or other chassis is mandatory, name it before ordering and wait for written confirmation.

Processor and memory facts from AMD

AMD’s official Ryzen AI Max+ 395 specifications identify a 16-core, 32-thread Zen 5 processor with boost frequency up to 5.1 GHz, x86-64 architecture and Radeon 8060S graphics with 40 graphics cores. AMD lists a 256-bit LPDDR5x-8000 interface and a maximum memory capacity of 128 GB. AMD gives a 55 W default TDP and a 45–120 W configurable range for the processor platform. Actual sustained power, acoustic behavior and performance depend on the selected system maker’s firmware and cooling design.

The 128 GB is shared system memory, not 128 GB of discrete GDDR attached to a replaceable graphics card. Firmware and operating-system policy decide how much can be made available to graphics workloads, while the OS and applications also require memory. D-Central will not advertise the full installed RAM as guaranteed usable “VRAM.” Before ordering, we identify the intended OS, the OEM firmware controls available on the quoted system and the model files you plan to run.

Why Strix Halo is interesting for private AI

A large shared memory pool can hold quantized models and context data that exceed the frame buffer of a conventional consumer GPU. The x86-64 CPU also preserves compatibility with a broad desktop and server software base. For a researcher, developer or small team that wants one quiet desk machine for local chat, document analysis, code assistance or offline experimentation, that balance can be more useful than buying several components and integrating a tower.

Capacity is not throughput. A model’s weights, quantization, KV cache, context length, batching, runtime and concurrent users all affect memory use and speed. A discrete GPU with high-bandwidth GDDR can be much faster when the entire workload fits. A DGX Spark may be the cleaner choice when CUDA and NVIDIA containers are requirements. Strix Halo makes the most sense when a large shared pool, x86, compact form factor and Windows-or-Linux choice are higher priorities. We do not publish invented tokens-per-second numbers or promise that “70B” means one consistent experience across different model families.

ROCm, Vulkan and runtime compatibility

Software support must be checked against current official matrices, not assumed from the AMD badge. AMD’s Radeon and Ryzen ROCm compatibility hub lists Ryzen AI Max 300-series paths for Linux and Windows. The ROCm 7.2 native Linux matrix identifies the Ryzen AI Max+ 395 as gfx1151 and validates a specific Ubuntu, ROCm, Python and PyTorch combination. That matrix also says only FP16 was officially validated for the listed APU path. These details change between releases, so D-Central confirms the target stack at order and setup time.

Ollama’s official GPU documentation lists the Ryzen AI Max+ 395 and describes ROCm requirements for Linux and Windows, with Vulkan as another supported route. llama.cpp build documentation includes HIP and Vulkan backends. These are projects maintained by their respective communities, not by D-Central. A runtime supporting an architecture does not guarantee that every model format, quantization or feature is accelerated correctly.

vLLM needs special care. Its official GPU installation documentation lists Ryzen AI MAX gfx1151 support with qualifying ROCm versions, but native vLLM is Linux-only; Windows users are directed to WSL or a community-maintained fork. If vLLM is mandatory, the order should specify Linux and the exact supported ROCm/vLLM combination. If Windows is mandatory, we select a Windows-supported Ollama, llama.cpp or other agreed path instead of promising identical behaviour.

Windows or Linux: decide from the workload

Choose Windows when the machine must also run Windows desktop software and the selected local-inference runtime has a documented Windows GPU path. Choose Linux when current ROCm, containers, Python environments or vLLM are central to the project. Dual-boot, virtualization and WSL introduce storage, driver and operational tradeoffs; none is included unless named in the order. AMD lists Windows 11, Ubuntu and RHEL among the processor’s operating-system ecosystem, but the selected OEM’s driver and firmware support controls the final system.

For an offline deployment, dependencies, model files and updates require a planned transfer process. Some model repositories require license acceptance or authentication. D-Central can stage files you are legally entitled to use when that is part of the written scope, but the customer owns model selection, license compliance and data governance. “Local” does not automatically mean secure: user accounts, disk encryption, backups, firewall policy, browser access and physical control still need an owner.

What D-Central validates at handoff

Once the exact system arrives, we verify its identity and ordered memory/storage, perform the agreed first boot, record the OS and relevant graphics/runtime versions, and test that the named local interface launches. When appropriate, a small compatible model is used to confirm the software path. That is an installation check, not a benchmark, burn-in certification or promise about your production model. Your written handoff states what was installed and what was actually observed.

If you provide the exact model repository or file, quantization, context target and expected concurrent users, we can document a more useful fit assessment before the order. We will flag obvious conflicts such as CUDA-only software, a Linux-only serving requirement paired with a Windows-only order, unsupported model operators, or an unrealistic expectation that all 128 GB is free for graphics. Custom validation using your private data or application is separately scoped.

Choosing among Strix Halo, DGX Spark and a GPU tower

Strix Halo is the balanced x86 option. NVIDIA DGX Spark at $9,449 CAD is the integrated Arm64/CUDA option with NVIDIA’s DGX OS and 128 GB unified memory. A custom GPU inference rig is the quoted, repairable and expandable option when a particular NVIDIA or AMD discrete GPU and conventional tower architecture fit the workload. Our DGX Spark versus custom AI build guide explains the decision in more depth.

There is no universal winner. Select from required software first, model memory second, concurrency and performance third, and physical/power constraints fourth. A familiar operating system is not enough if the accelerator backend is unsupported. Likewise, theoretical capacity does not rescue a runtime that lacks kernels for the target model.

Compatibility boundary before you order

• Fixed: Ryzen AI Max+ 395, Radeon 8060S and 128 GB installed LPDDR5x system memory, at $7,349 CAD before applicable tax and shipping.
• Confirmed in writing: OEM and chassis, SSD capacity, ports, network interfaces, power adapter, included operating-system license, firmware options, requested OS and local-inference stack.
• Workload inputs required: exact model and quantization, context, concurrency, runtime, container or Python requirements, online/offline policy and any peripheral or display dependency.
• Not guaranteed: all RAM available to the GPU, CUDA or TensorRT compatibility, native Windows vLLM, a particular benchmark result, every ROCm release, every model format, component upgrades or a specific chassis that was not confirmed.
• Customer-owned: data backups, account security, model and software licenses, production monitoring, network policy and applications outside the written configuration scope.

Ordering and returns boundary

The price is $7,349 CAD before applicable taxes and shipping. No stock or delivery promise is made on this page. Before sourcing begins, the order record must identify the exact system, included accessories, setup scope, shipping cost and current procurement estimate. If the named chassis cannot be sourced on acceptable terms, D-Central will request approval for a clearly identified alternative or will not proceed; we will not make an undisclosed substitution.

This is configured, made-to-order hardware. D-Central’s Return and Refund Policy states that configured builds, special orders and made-to-order work are non-refundable once sourcing, assembly, configuration or service work has started, except where they arrive defective or dead on arrival. A defect or DOA must be reported within 30 days of delivery for review; this is not a general change-of-mind window. Contact support with the order number and evidence, and receive authorization before shipping anything back. The product must be complete with included accessories and free from damage caused by misuse, installation error, modification or improper handling. Shipping and the resolution process follow the approved written instructions. The posted policy and order confirmation control; this page does not create a separate warranty promise.

AMD Strix Halo 128 GB FAQ

Which brand and chassis will I receive?

The exact OEM, chassis, storage, ports and accessories are confirmed in the order record before D-Central commits sourcing. The processor family and 128 GB capacity are fixed, but this page does not promise an unconfirmed brand or AMD’s own Halo developer platform.

Is all 128 GB available as VRAM?

No. It is unified system memory shared with the OS, CPU and applications. Firmware and software decide the graphics allocation, and usable capacity is lower than installed capacity. We confirm the relevant firmware path for the selected system.

Can it run CUDA or TensorRT software?

Not as a native AMD accelerator path. CUDA and TensorRT are NVIDIA technologies. Choose software with documented ROCm, HIP, Vulkan or CPU support, or choose an NVIDIA system when CUDA is mandatory.

Should I choose Windows or Linux?

Windows is appropriate for Windows desktop requirements and a runtime with a supported Windows AMD path. Linux is the clearer choice for native vLLM and many current ROCm/container workflows. State the required applications before ordering.

Will it run my 70B model?

Possibly, depending on model architecture, quantization, context, runtime and desired speed, but the label “70B” is not enough to promise fit or performance. Send the exact files and operating target for assessment.

Does first-boot setup include my RAG application?

No, unless separately written into the order. The baseline covers the agreed runtime and interface. Data ingestion, retrieval design, identity, evaluation, backups and production operation are additional work.

What if the delivered system is defective?

Report the fault or DOA within 30 days of delivery with the order number and evidence. Wait for D-Central’s return instructions. Change of mind, an incorrect self-selected configuration or unsupported modifications are not covered by that defect-reporting window.

Sources and review: AMD Ryzen AI Max+ 395 specifications; AMD ROCm compatibility hub and native Linux matrix; Ollama, llama.cpp and vLLM official documentation; D-Central Return and Refund Policy. Reviewed 25 August 2026. Compatibility changes across releases, so the written order confirmation identifies the supported target stack.