Intelligent Living frames Strix Halo as a small x86 “AI supercomputer” able to run 70-billion-parameter and larger local models without a discrete GPU. AMD’s own specifications support the underlying hardware premise. The Ryzen AI Max+ 395 combines 16 Zen 5 CPU cores, Radeon 8060S integrated graphics with 40 compute units, an XDNA 2 NPU, and up to 128GB of LPDDR5X-8000 memory on a 256-bit interface. AMD also officially supports Windows 11, Ubuntu, and RHEL on the processor.
The important correction is that 128GB of system memory is not automatically 128GB of usable GPU memory in every workload. On Windows, AMD’s Variable Graphics Memory feature officially supports assigning up to 96GB to graphics. That is a major allocation by PC standards, but it still leaves a meaningful distinction between AMD’s shared-memory design and DGX Spark’s 128GB coherent memory pool, which Nvidia presents as directly available to its Grace Blackwell system.
For Windows users, that difference affects model selection, context size, and whether a workload can remain fully accelerated instead of spilling into ordinary system-memory paths.
Ryzen AI Max+ 395 changes the VRAM equation
Strix Halo is unusual because it brings workstation-scale memory capacity to an integrated GPU. A typical gaming desktop separates system RAM from GPU VRAM: a machine may have 64GB of DDR5 and an RTX 5090 with 32GB of GDDR7, but a model that needs 45GB of accelerator memory cannot simply use the spare system RAM without a performance penalty.
Ryzen AI Max systems work differently. The CPU and Radeon 8060S GPU use the same LPDDR5X pool, avoiding the copy between separate host and GPU memory pools. AMD rates the Max+ 395 platform for 256GB/s of memory bandwidth, enough to make quantized local inference practical on models that would not fit within the VRAM of mainstream discrete GPUs.
AMD has specifically promoted the processor’s ability to use up to 96GB as graphics memory in a 128GB configuration. That makes a 70B-class model at 4-bit quantization a realistic target, with capacity left over for the operating system, context window, retrieval data, or another local service. AMD’s more recent ROCm guidance also documents local inference on 122B-class weights using a 128GB Ryzen AI Max+ system.
Capacity alone does not equal speed. LPDDR5X bandwidth is far below the bandwidth available from high-end GDDR7 cards, and it is a fraction of the bandwidth of datacenter accelerators using HBM. Local LLM decoding can be bandwidth-bound, meaning a model may fit perfectly yet still generate fewer tokens per second than it would on an expensive Nvidia GPU.
That makes Strix Halo a capacity-first machine. It is compelling for a developer who needs a large model on a desk, does not need high concurrent throughput, and wants a normal Windows or Linux PC alongside the AI workload. It is a weaker answer for an organization trying to serve many users, process long prompts rapidly, or run CUDA-specific production tooling.
DGX Spark still has the specialized AI advantage
Nvidia’s DGX Spark uses the GB10 Grace Blackwell Superchip with 128GB of coherent LPDDR5X memory, a 20-core Arm CPU complex, Blackwell GPU hardware, and a ConnectX-7 network interface capable of 200Gbps networking. Nvidia rates the platform at up to 1 petaflop of sparse FP4 tensor performance and 273GB/s of memory bandwidth.
Those specifications are not directly comparable with AMD’s 126 “platform TOPS” figure. AMD’s total combines compute capabilities from different components, including the NPU, while Nvidia’s FP4 figure describes a tensor workload format associated with the Blackwell GPU. Buyers should not use either headline number as a proxy for LLM tokens per second, image-generation speed, or PyTorch compatibility.
The more consequential divide is software. DGX Spark is purpose-built around Nvidia’s CUDA software environment, DGX OS, NGC containers, and high-speed interconnect. Tom’s Hardware found the system capable and well rounded for local AI, while also noting its appeal is tied to the established CUDA stack rather than its utility as a general desktop PC.
AMD has made tangible progress with ROCm on Ryzen AI Max hardware, including official Windows guidance for ComfyUI and local image generation. However, a Windows or Linux installation using Radeon 8060S still requires users to confirm application support, backend compatibility, model format, and driver versions. A package that assumes CUDA may work through an alternative backend, require a ROCm-compatible build, or not work acceptably at all.
For an IT team, that means DGX Spark costs more but offers a more prescriptive software appliance. Strix Halo is less constrained and more familiar to administer as an x86 endpoint, but it demands more validation before deployment.
The “half-price” claim has become a moving target
The submitted report relies heavily on a low-price comparison, citing roughly $2,000 systems against a DGX Spark price approaching $4,699. The low-end figure is real as a list-price possibility for some Ryzen AI Max+ 395 mini PCs, but it is not a dependable comparison for a buyer trying to acquire a 128GB system today.
For example, Minisforum’s US store lists its 128GB MS-S1 Max at $3,639, discounted from $4,549. The same company has also shown lower pricing on a regional storefront, illustrating the problem: availability, region, storage configuration, warranty coverage, and whether a configuration is actually in stock can alter the apparent bargain substantially.
The DGX Spark’s price has risen, too. Nvidia increased the Founders Edition price from $3,999 to $4,699 amid memory-market pressure, according to reporting by Tom’s Hardware. That does narrow the difference against higher-priced Strix Halo systems, but it does not make all Ryzen AI Max machines equivalent to a DGX Spark.
A fair purchase comparison needs to account for what comes with each machine:
- A DGX Spark includes 128GB of coherent memory, a 4TB NVMe SSD, ConnectX-7 networking, and Nvidia’s dedicated AI software platform.
- A Ryzen AI Max+ 395 mini PC may offer Windows, conventional x86 application support, gaming capability, additional networking, and PCIe expansion, but its ports, storage, cooling, power limit, and operating-system image are determined by the OEM.
- The memory in current Strix Halo systems is soldered LPDDR5X. Choosing 32GB, 64GB, or 128GB at purchase is therefore a long-term capacity decision, not an upgrade a sysadmin can schedule later.
The practical result is that a $2,000 128GB Strix Halo box, when available, is an unusually strong local-AI value. A $3,600 to $4,300 Strix Halo system must instead be evaluated against DGX Spark on software, expansion, serviceability, and workload behavior—not merely on memory capacity.
Windows users should treat 96GB as the planning limit
The report’s discussion of a Linux “unlock” to expose 110GB to 120GB of GPU-addressable memory is based on community tuning rather than an AMD-supported configuration. Linux users have reported manipulating AMD GPU memory-management parameters to increase what applications can allocate, but that is not the same thing as AMD certifying a 120GB VRAM configuration across distributions, kernels, BIOS versions, and AI frameworks.
That distinction is more than legal fine print. Kernel parameters that improve model fit can affect system stability, memory pressure behavior, suspend and resume, display allocation, and supportability. An enterprise IT administrator should not size a deployment around an unofficial 120GB target unless it has been tested on the exact hardware and software image being deployed.
For Windows planning, use the documented 96GB Variable Graphics Memory ceiling. A machine bought to host a 70B model can be designed around that limit with room for context and the host operating system. A machine bought specifically to run a model that requires more than 96GB of accelerator-visible memory should be tested under Linux before procurement, rather than assuming a community configuration will translate to production.
This also puts the 128GB model-fit claims in perspective. A heavily quantized 120B mixture-of-experts model may run well because it fits within the available shared memory and activates only a subset of parameters per token. A 120B dense model is a different capacity and bandwidth problem. Parameter count alone is not sufficient sizing information; architecture, quantization format, context length, KV cache, and backend all matter.
One platform can game; the other is built to cluster
Strix Halo’s broadest advantage is that it remains an ordinary x86 PC. A Radeon 8060S-equipped system can run Windows applications, Linux development tools, native PC games, creative software, containers, and local inference on one machine. AMD’s official platform support includes Windows 11 and mainstream enterprise Linux distributions, which makes it easier to incorporate into a mixed workstation fleet.
The Radeon 8060S is capable integrated graphics, but buyers should be careful with desktop-GPU comparisons. Performance varies dramatically by OEM power limit and cooling design. A 45W tablet or thin laptop will not behave like a 120W desktop mini PC, despite sharing the same processor name. RAM allocation also does not turn the Radeon 8060S into a high-end discrete GPU: extra memory helps capacity-limited workloads, while shader throughput and memory bandwidth still set the performance ceiling.
DGX Spark has the opposite focus. It is not a sensible gaming purchase, as testing has shown, and its Arm-based environment is less convenient for legacy desktop software. Its differentiator is the ConnectX-7 interconnect, which Nvidia says can connect two systems for work with models up to 405 billion parameters. That is a defined, vendor-supported path to a 256GB paired environment that Strix Halo mini PCs do not currently match.
A pair of AMD mini PCs can be networked, including through 10GbE on some models, but networking two hosts does not create a transparent 256GB unified memory pool. Distributed inference is possible in many forms, but it is a software and orchestration project rather than the appliance-like clustering proposition Nvidia sells.
The better choice depends on what must remain local
Strix Halo is not a DGX Spark replacement in the broadest sense. DGX Spark remains the more specialized local-AI appliance, particularly for CUDA-native development, long-context prompt processing, batch inference, containerized Nvidia workflows, and two-node clustering.
But the Ryzen AI Max+ 395 has opened a category conventional PC hardware largely ignored: local models too large for consumer GPU VRAM, housed in a compact Windows-and-Linux-capable computer that can still serve as a developer workstation. The decisive advantage is not a marketing TOPS figure. It is the ability to buy 128GB once and use a large fraction of it for local AI without adding a discrete accelerator.
For Windows users, the disciplined purchasing rule is simple: price the exact 128GB configuration, confirm its sustained power limit and cooling, plan around 96GB of supported graphics allocation, and verify the required software stack before ordering. If those checks pass, a Ryzen AI Max+ 395 system can be one of the most flexible ways to keep a large local model on a desk.
