Practitioner Benchmarks: DGX Spark Crushes Strix Halo on Prompt Processing (1,723 vs 339 tok/s) but Token Gen Is Near-Even

Hardware Corner·medium signal

June 2026 local-inference hardware comparisons across the 128GB class — NVIDIA DGX Spark (~$4k), AMD Strix Halo/Ryzen AI Max+ 395 (~$2–3k), and M5 Max 128GB (~$5k) — show DGX Spark delivering 1,723 tok/s prompt processing vs Strix Halo's 339 tok/s, while token generation lands surprisingly close at ~34–38 tok/s on 120B models. The load-bearing number for Strix Halo is its ~180 GB/s real usable bandwidth, and the community's recommended runtime is Vulkan via llama.cpp, not ROCm. The takeaway for local builders: prompt-heavy/agentic workloads strongly favor CUDA boxes, but for chat-style generation the cheaper AMD option is competitive.