AMD launches Strix Halo kit for $3,999 to compete with Nvidia DGX Spark

AMD has officially launched its Strix Halo development kit, a mini-PC system designed to compete directly with the Nvidia DGX Spark. Priced at $3,999, the Strix Halo undercuts Nvidia's model by $700, which recently rose to $4,699. Equipped with the Ryzen AI Max+ 395 processor and 128 GB of unified memory, this kit targets AI developers seeking a powerful yet more accessible platform. This move by AMD is not an isolated event: it is part of a broader strategy to challenge Nvidia's leadership in the AI hardware market, which has been dominated by the A100 and H100 series GPUs in data centers, and more recently by the DGX Spark in the desktop workstation segment. The Strix Halo represents AMD's first serious foray into this niche, combining its unified CPU/GPU architecture with an aggressive price point.

What is the Strix Halo kit?

The Strix Halo is a compact development system that integrates the AMD Ryzen AI Max+ 395 processor, a unified CPU/GPU/NPU with 40 RDNA 3.5 compute units and an XDNA 2 NPU delivering up to 60 TOPS. The 128 GB of unified memory allows handling large language models and AI workloads without dedicated VRAM. Unlike the DGX Spark, which uses Linux and CUDA, the Strix Halo offers native Windows 11 support, easing integration with popular development tools like Visual Studio Code, PyTorch with DirectML, and TensorFlow with ROCm. Additionally, the kit includes USB4 ports, 2.5 GbE Ethernet, and support for up to three 4K monitors, making it a versatile workstation for developers who also need to perform general productivity tasks. According to Tom's Hardware, the Ryzen AI Max+ 395 is built on a 4 nm process and features 16 Zen 5 cores, giving it competitive CPU performance, while the 60 TOPS NPU surpasses the 45 TOPS NPU of the DGX Spark (based on Blackwell), though direct comparison is complex due to architectural differences.

Strategic importance

This launch is significant because AMD directly challenges Nvidia's dominance in the AI workstation market. While Nvidia relies on its CUDA ecosystem and Linux, AMD bets on Windows compatibility and a lower price. Moreover, the 128 GB of unified memory is an advantage for models requiring high bandwidth, such as large-scale transformers, which often need more than 80 GB of VRAM for inference with large batches. The DGX Spark, on the other hand, offers 128 GB of unified LPDDR5X memory but with a bandwidth of 1 TB/s compared to the estimated 800 GB/s of the Strix Halo, according to preliminary specifications. However, the real performance difference will depend on software optimization. AMD has improved its software support with ROCm 6.0, which now offers compatibility with PyTorch and TensorFlow, though it still lacks the maturity and broad library support of CUDA. This launch also comes at a time when Nvidia has increased the DGX Spark's price from $3,999 to $4,699, making the Strix Halo even more attractive from a cost perspective.

Market implications

Competition could pressure Nvidia to adjust prices or improve its offering. For developers, the choice between Strix Halo and DGX Spark will depend on the preferred ecosystem: Windows vs Linux, and ROCm vs CUDA. AMD has improved its software support, but still lags behind Nvidia in framework optimization. The Strix Halo could also appeal to companies already using AMD hardware seeking consistency, as well as independent developers who want to avoid dependency on the CUDA ecosystem. In the broader market, this launch could accelerate the adoption of AI hardware in workstations by offering a more affordable option. However, the true impact will be seen when the first independent benchmarks compare real-world performance on typical AI workloads, such as fine-tuning language models or inference with transformers. Historically, Nvidia has dominated this segment since the launch of the DGX Station in 2017, and the DGX Spark (originally Project DIGITS) was announced at CES 2025 as an affordable platform for developers. AMD, for its part, has found success with its APUs in consoles and laptops, but has never before launched a product so focused on desktop AI.

What readers should know

• Price: $3,999 vs $4,699 for the DGX Spark. A $700 difference that can be allocated to peripherals or software licenses.
• Performance: The Ryzen AI Max+ 395 delivers up to 60 TOPS in NPU, but real-world AI performance depends on software. It is expected to compete in inference tasks with models up to 70B parameters, though the DGX Spark may have an edge in training due to its higher memory bandwidth.
• Availability: The kit is expected to reach developers in the second quarter of 2025, according to AMD. Pre-orders may open in April.
• Ecosystem: Windows 11 with support for DirectML and ROCm, while Nvidia uses CUDA on Linux. AMD also offers compatibility with ONNX Runtime and Hugging Face, easing model migration.

"The Strix Halo represents an important step in democratizing AI development, offering a more affordable and familiar alternative for Windows developers," notes an analyst from Tom's Hardware. However, he warns that "ROCm's lack of maturity compared to CUDA could be a hurdle for researchers relying on specialized libraries."

In summary, AMD's Strix Halo is a solid option for developers prioritizing cost and Windows compatibility, though Nvidia's CUDA ecosystem remains the de facto standard in AI research. The final decision will depend on each developer's specific needs: if they primarily work in Windows and want to save money, the Strix Halo is attractive; but if they require maximum performance and compatibility with cutting-edge AI frameworks, the DGX Spark might be the better choice despite its higher price. Time will tell if AMD can close the software gap and become a serious contender in this market.