AMD Built the DGX Spark Rival I Predicted… But There's a Catch
Alex Ziskind ·2026-07-06 ·2 min read
Summary written by us from the video's transcript. The video, and everything in it, is Alex Ziskind's work.
AMD’s new Ryzen AI Halo box offers a compact, $4,000 AI workstation that matches Nvidia’s DGX Spark on many tasks but has distinct trade‑offs in performance, software readiness, and platform compatibility.
Takeaways
- The Halo matches the DGX Spark on token‑generation speed but lags significantly in compute‑heavy prefill and image/video tasks.
- Its out‑of‑the‑box software stack (LLM tools, NPU support) and x86 Windows/Linux compatibility simplify setup compared to the Spark’s ARM‑only Linux environment.
- For developers focused on code compilation or casual LLM chat, the Halo offers no cost advantage over cheaper alternatives; its value lies in a ready‑to‑run AI workstation for tinkering.
- The inclusion of an NPU provides a low‑power inference path, though current software ecosystems have limited support.
- Benchmark presentations from AMD emphasize favorable token‑generation numbers while downplaying prefill performance where the Spark still leads.
What the Halo Box Is
The Halo is a tiny dedicated AI appliance built around AMD’s Ryzen AI Max Plus 395 chip – the same Stricks Halo silicon found in 2025 desktop boxes like the Beink GTR9 and Framework Mini. It ships with 128 GB of unified memory, a fan‑cooled chassis with top vents, and is priced at $4,000, the same price Nvidia originally charged for its DGX Spark.
Target Audiences
The reviewer splits potential buyers into three groups: (1) developers who need CPU‑bound workloads such as code compilation; (2) tinkerers who experiment with LLMs and image generation; and (3) a third group not covered in this video but hinted at for future discussion. The Halo is not aimed at developers looking only for cheap compile machines, nor at high‑end production users who need raw pre‑fill speed.
CPU‑bound Performance
Running a multi‑core Python benchmark (Mandelbrot) showed the Halo and Spark drawing similar power (~164 W) and delivering comparable scores: 18.4 vs 15.4 points, with a Mac at 17.3. The differences are minor, confirming that for pure CPU work the Halo offers no advantage over cheaper alternatives.
LLM Inference – Token Generation vs Prefill
Using Llama CPP on Gemma‑412B, token‑generation (memory‑bandwidth bound) was 24.6 t/s on the Halo versus 26.4 t/s on the Spark; the Apple M4 Pro topped both at 33.8 t/s. Prefill (compute‑heavy) showed a larger gap: the Spark achieved ~2,000 units while the Halo managed ~650, roughly one‑third the speed. However, software improvements since last year have narrowed this gap for the same chip. For interactive chat or agent use, the token‑generation parity makes the Halo adequate; heavy context dumps or large image/video generation still favor the Spark.
Image & Video Generation
In a stable‑diffusion test the Spark delivered 2.9 iterations per second versus the Halo’s 1.3. A WAN 2.2 video task took ~5½ minutes on the Spark but 75 minutes on the Halo, highlighting the Spark’s advantage in compute‑intensive GPU workloads.
Software & Platform Advantages
The Halo ships with pre‑installed AI tools (LM Studio, Ollama, Comfy UI, Lemonade Server) and works out‑of‑the‑box on both Windows and Linux x86, avoiding the ARM‑only Linux environment of the Spark. This makes driver installation, Python environments, and IDEs like Visual Studio run without emulation. An NPU (up to 50 TOPS) is also included for hybrid inference, though software support is still emerging.
Benchmark Claims vs Reality
AMD’s marketing highlights token‑generation gains (4–14 % faster on several models) and image generation wins against the Apple M4 Pro, but omits prefill results where the Spark remains three times faster. The price comparison notes the Spark’s fluctuating retail price ($4,000‑$4,700), affecting “tokens per dollar” metrics.