Thai
While AMD’s Strix Halo devices offered 128GB unified memory for local AI long before NVIDIA’s entry, NVIDIA defined the desktop local-AI category with the one-liter DGX Spark. Launched alongside the Ryzen AI Max PRO 400 Series in May 2026, the $3,999 AMD Ryzen AI Halo serves as AMD’s direct counterpart to the DGX Spark, with pre-orders opening in June via Micro Center and in-store availability starting July 10. Matching the Spark’s compact footprint and 128GB unified memory, the Halo features key design differences that redefine its use cases.
Core Specifications & Hardware Design
Powered by the 16-core/32-thread Zen 5 Ryzen AI Max+ 395 APU, the Halo integrates Radeon 8060S graphics (40 RDNA 3.5 CUs) and an XDNA 2 NPU delivering 50 TOPS, with a combined system AI throughput of up to 126 TOPS. Its 128GB LPDDR5x-8000 memory provides 256GB/s bandwidth, supporting local inference of up to 200B-parameter models, and runs on a 120W single USB-C power input.
The 150×150×45.4mm, under-1.2kg aluminum chassis adopts a high-airflow ventilation design with a status indicator ring. Rear I/O includes three USB-C data ports, one USB-C power port, HDMI 2.1b, 10GbE Ethernet and a Kensington lock slot. Internally, it uses dual-fan cooling and a standard M.2 2280 SSD slot (max 8TB upgradeable), unlike the Spark’s rare 2242 drive. AMD pre-sets maximum Variable Graphics Memory across Windows 11 and its custom Linux developer image, eliminating manual tuning for large AI models. A key tradeoff is the absence of high-speed fabric, limiting multi-node clustering versus the Spark.
Key Differentiators vs. NVIDIA DGX Spark
The Halo’s biggest advantage is dual Windows/Linux OS support, addressing the top user complaint about the Linux-only DGX Spark. It offers broader aftermarket storage compatibility and user-friendly visual status lighting, while matching the Spark’s core large-model local inference capability at a lower price. The Halo retails for $3,999 with 2TB storage, while the updated DGX Spark costs ~$4,700; third-party Grace Blackwell systems sit at a comparable $4,000 baseline.
Performance Benchmarks
Windows Workstation & AI Performance
Tested against identical-spec HP Z2 Mini G1a and ZBook Ultra G1a systems, the Halo delivers top-tier comprehensive performance. It scores 37,316 in Cinebench R23 multi-core and 184.2 GIPS in 7-Zip compression/decompression, leading all rivals. In UL Procyon AI tests, it achieves industry-leading scores: 1,192 for Phi text generation and 937 for Stable Diffusion 1.5 image generation, with drastically reduced first-token latency. It also leads in SPECworkstation 4 financial, media and product design workloads, with consistent advantages in Geekbench 6 single/multi-core and GPU OpenCL performance.
Linux CPU & Storage Performance
On Linux Phoronix benchmarks, the Halo outperforms the DGX Spark in CPU-centric tasks: 11% faster 7-Zip compression, 38% faster decompression and 14% quicker LLVM compilation. For storage, the Halo’s Gen5 SSD is locked to Gen4 speeds, creating a bandwidth gap. It excels at low-queue-depth latency for random I/O but falls behind the Spark in high-load sequential throughput, saturated random write IOPS and sustained high-concurrency storage performance.
LLM Inference (vLLM)
The Spark dominates high-concurrency LLM serving, with 2–4x higher throughput in most scenarios, peaking at 8.8x faster performance in prefill-heavy 120B-parameter model tasks. The Halo remains competitive in decode-heavy workloads (within 10–12% of the Spark) and lightweight 8B-model inference, making it viable for daily local AI development.
Conclusion
The Ryzen AI Halo is a well-balanced compact x86 AI workstation rather than a direct Spark competitor. It outperforms all tested Ryzen AI Max+ 395 devices in Windows professional and AI workloads and beats the DGX Spark in Linux CPU tasks. Its core strengths are dual-OS support, flexible upgradability, user-friendly tuning and mainstream compatibility at a competitive price.
The DGX Spark remains superior for high-throughput LLM serving and multi-node clustering. Meanwhile, the Halo is ideal for developers needing Windows-Linux dual-environment AI development and unified professional workloads. AMD also plans a 2026 Q3 upgrade to 192GB unified memory, extending the platform’s support for larger AI models.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!