Summary
Alex Ziskind tests the Kadas Mind Pro — a palm-sized PC powered by Intel’s Panther Lake architecture with an ARC B390 integrated GPU and 64GB of RAM — to see how fast it can run Qwen 3 27B locally. The machine is evaluated in two configurations: standalone iGPU and paired with a dock housing a discrete GPU. Using llama.cpp, Ziskind benchmarks all available inference backends (CPU, Vulcan, SYCL, and OpenVINO), measuring both prefill and decode speeds to establish a baseline before any optimization.
The core of the video is a systematic effort to extract maximum token-generation throughput from constrained hardware. Speculative decoding with a bundled 1.5GB draft model more than doubles iGPU decode speed from roughly 5 to 11 tokens per second. GPU layer offloading via llama.cpp’s NGL flag pushes further to 14 tokens/sec at an optimal 56 layers — but carelessly using NGL99 when the model doesn’t fully fit in VRAM causes a catastrophic 17x slowdown. Dropping one quantization level from Q4KM (18GB) to IQ4XS (13GB) then allows the entire model to reside in VRAM and roughly doubles both prefill and decode speeds.
Key takeaways for local AI practitioners: actual measured memory bandwidth on this device is approximately 100 GB/s, blindly following tutorial flags like NGL99 can silently destroy performance, and quantization choice often matters more than any software tuning. An essential reference for edge inference engineers and local AI enthusiasts evaluating Intel Panther Lake hardware for serious workloads.
📺 Source: Alex Ziskind · Published September 13, 2026
🏷️ Format: Benchmark Test
