Compact AI workstations in comparison: Nvidia DGX Spark meets AMD Ryzen AI Max+ 395

Nvidia announced the DGX Spark platform first. AMD delivered the direct response with the Strix Halo architecture and, interestingly, brought the corresponding chips to the market even earlier than the competitor. As a direct opponent to the Nvidia GB10, the AMD Ryzen AI Max+ 395 is typically also paired with 128 GB of memory, enabling the execution of large local models. In various AI benchmarks and pure inference speed, the chips are nearly on par, especially in FP16 and FP64 tasks. The memory bandwidth and many other performance figures are also identical on paper. Therefore, it is worth considering systems like the HP ZGX Nano G1n AI Station as well as systems like the Bosgame M5.

The two systems' underlying processor architectures are fundamentally different from one another. While Nvidia uses an ARM-based Grace module for the GB10 Superchip, AMD relies on the classic x86 architecture with Zen 5 cores for the Ryzen AI Max+ 395. This difference has a significant impact on software compatibility. AMD's x86 platform scores points with extensive support for established legacy applications and slips seamlessly into the Windows ecosystem. In contrast, Nvidia's ARM strategy is optimized only for the Linux-based DGX operating system and heavily parallelized artificial intelligence workloads, limiting its applicability to traditional desktop tasks.

AMD takes jet another architectural path with the integration of a dedicated NPU. This delivers 50 INT8 TOPS and allows smaller models or background tasks to run in an energy-saving manner. Projects like FastFlowLM benefit from this architecture, as the system does not have to engage the compute-intensive main chip for every AI task. Nvidia, however, retains a massive memory advantage with the Blackwell architecture and native FP4 support, which is missing from AMD in this form.

The decisive differences become apparent when looking at the software ecosystems. In order to maintain its position, Nvidia relies on the well-established CUDA ecosystem. AMD counters this with its own ROCm platform for the RDNA architecture. In terms of compatibility in many specialized applications, this does not yet quite match Nvidia's software stack.

Ultimately, the decision comes down to weighing the budget against the ecosystem. Nvidia charges noticeable premiums for the DGX Spark systems, offering the industry standard in return. Preparing code for large data centers makes CUDA almost unavoidable. For pure inference tasks that primarily require a lot of local memory and can do without Nvidia's proprietary features, the Ryzen AI Max+ 395 represents a powerful and often more cost-effective alternative.