Lemonade NPU Crushes TTFT, Vulkan Wins Decode

Hot take: if your workload is dominated by giant prompt ingestion, the NPU is the right place to spend silicon; if you live in steady-state chat generation, Vulkan on the iGPU still looks stronger.

• –Lemonade NPU/hybrid wins hard on TTFT, with the Qwen3 4B test showing a dramatic first-token advantage over llama.cpp Vulkan.
• –llama.cpp still leads on TPS, especially on the smaller lfm 1.2B run, where Vulkan nearly doubles NPU throughput.
• –The comparison is useful but not perfectly apples-to-apples because the quantization formats and backends differ.
• –The data supports a practical split: NPU/hybrid for RAG-style long-context prefill, iGPU Vulkan for faster decode-heavy interactions.
• –The takeaway is about workload shape, not absolute winner status; long-context latency and decode speed are optimizing for different bottlenecks.