Recent benchmarks reveal that NVIDIA’s DGX Spark and AMD’s Ryzen AI Max+ 395 excel in different AI workload phases, with software support and memory architecture playing crucial roles in their real-world performance and value for users.
NVIDIA’s DGX Spark and AMD’s Ryzen AI Max+ 395 are often presented as rival compact AI workstations, but the comparison only looks simple until the workload is broken into its two distinct parts: prompt processing and token generation. In recent benchmark material discussed by HardwareCorner and IntuitionLabs, the Spark was only modestly ahead on decode speed with GPT-OSS 120B, yet far ahead on prompt throughput. That split matters because the two systems are not just competing on raw performance; they are competing on which phase of an AI task they accelerate best. Tom’s Hardware and Notebookcheck both frame the DGX Spark as a specialist AI system, while comparison coverage of Strix Halo machines emphasises the AMD platform’s broader PC utility and lower cost.
The decisive point is that benchmark ratios can move sharply with software stack, quantisation and backend choice. In the material supplied for this comparison, the same model produced materially different results depending on whether it ran through llama.cpp, Vulkan or ROCm, and whether testing was done with short prompts or long-context workloads. Notebookcheck’s comparison makes the same broad point: on these machines, software support and implementation details are part of the performance story, not a footnote. That makes any single “winner” label misleading, because the apparent gap can widen or shrink without any change to the silicon itself.
The hardware specifications also explain why the two systems behave differently. The DGX Spark’s GB10 platform pairs 128GB of unified memory with NVIDIA’s Blackwell-based acceleration stack, while Ryzen AI Max+ 395 systems rely on AMD’s Strix Halo architecture with shared LPDDR5X memory and a Radeon 8060S integrated GPU. That shared-memory design is attractive for compact machines, but it also means both platforms are operating in a memory-bound regime rather than the traditional discrete-GPU model. Tom’s Hardware’s review of the DGX Spark highlights the machine’s AI focus, thermals and power envelope, while AMD-focused mini PC guides stress that Strix Halo systems are more varied in implementation and pricing.
Memory capacity is another point where the headline figures can mislead. AMD’s Ryzen AI Max+ 395 systems are widely sold with 128GB of RAM, but Windows imposes a 96GB cap on what can be assigned as graphics memory, which changes how much of that memory is realistically available for large models in that environment. By contrast, the DGX Spark’s memory is presented as a coherent pool for CPU and GPU workloads, making its usable model footprint more straightforward, especially for users working close to the upper limit of what can fit locally. That distinction is central for anyone targeting models in the roughly 95GB to 110GB range.
Pricing has also shifted enough that the old “half the price” argument no longer holds cleanly. The supplied comparison notes that DGX Spark pricing has risen, while 128GB Ryzen AI Max+ 395 systems now vary widely by OEM and memory configuration, with current market prices often far above their original launch level. TechRadar’s reporting on newer, even larger Ryzen AI Max-based mini PCs shows how quickly this category is being pushed upwards by memory costs and positioning, which further complicates direct price comparisons. The result is a narrower premium for NVIDIA than buyers might expect from earlier discussions.
That leaves the practical question: what kind of AI work do these machines actually support best? If the workload is short, interactive and mostly decode-heavy, the AMD box can look like the better buy because generation performance is often close enough to the Spark to matter more than the difference in list price. If the task is prompt-heavy, involves long-context inference, or depends on NVIDIA’s CUDA and TensorRT-LLM ecosystem, the Spark’s higher prefill throughput and stronger software support make its premium easier to defend. Tom’s Hardware’s review of the Spark and Notebookcheck’s head-to-head both point to that same conclusion: the choice is less about a universal winner than about matching platform strengths to workload shape.
For most buyers, then, the real decision is not which machine is faster in the abstract, but which one aligns with the way their tokens are spent. Compact AI workstations are now good enough that software, memory allocation and deployment model matter almost as much as the chip itself. In that context, the DGX Spark is the more specialised platform, while Ryzen AI Max+ 395 systems remain the more flexible and usually cheaper route for users who need a capable general-purpose PC as well as local AI hardware.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
