TL;DR
There are two technical routes for the AI PC: the 'NPU low-power small-model' route led by Qualcomm Snapdragon X Elite (Hexagon NPU at 45 TOPS, up to 64GB memory, on-device generative AI at about 30 tokens/s), and the 'GPU high-compute' route led by NVIDIA RTX Spark (up to 1 PFLOPS FP4, up to 128GB unified memory, running a 125B-parameter model locally). The AI metrics differ (TOPS is INT8, PFLOPS is FP4) and cannot be converted or compared directly.
Introduction
Qualcomm route (NPU small models): since May 2024, Snapdragon X Elite has been the exclusive launch platform for Microsoft Copilot+ PCs, with the built-in Qualcomm Hexagon NPU officially rated up to 45 TOPS (INT8), meeting Microsoft's 40+ TOPS Copilot+ PC requirement (per Qualcomm's official site).
NVIDIA route (GPU high compute): on October 7, 2026, Microsoft and NVIDIA unveiled the RTX Spark platform integrating a Grace CPU and a Blackwell RTX GPU, with officially published up to 1 PFLOPS FP4 and up to 128GB unified memory, able to run the 125B-parameter Qwen 3.8 Flash Next locally (per ifanr citing official materials).
Core Spec Comparison
| Dimension | Qualcomm NPU route (Snapdragon X Elite) | NVIDIA GPU route (RTX Spark) |
|---|---|---|
| Launch | May 2024 (Copilot+ PC debut) | October 7, 2026 (Surface Laptop Ultra) |
| AI accelerator | Qualcomm Hexagon NPU | Blackwell RTX GPU |
| AI performance (official) | Up to 45 TOPS (INT8) | Up to 1 PFLOPS FP4 (FP4) |
| Metric note | INT8 integer ops | FP4 floating-point ops; not directly comparable to TOPS |
| On-device models | 13B+ parameter LLMs, generative AI about 30 tokens/s (official) | 125B-parameter Qwen 3.8 Flash Next (official) |
| Memory | Up to 64GB LPDDR5x (135GB/s bandwidth) | Up to 128GB unified memory |
| Positioning | Low power, long battery, light AI | High compute, heavy local AI and agent workloads |
| Software ecosystem | Hexagon SDK / Windows on ARM | CUDA, DLSS, G-SYNC, ray tracing |
| Representative device | First Copilot+ PCs (2024) | Surface Laptop Ultra (2026) |
Key Differences
- Metrics are not directly comparable: Qualcomm's 45 TOPS is INT8 integer performance; NVIDIA's 1 PFLOPS is FP4 floating-point performance. Neither vendor publishes a cross-metric conversion, so they cannot be directly subtracted or divided (per both official specifications).
- Memory and model scale: RTX Spark goes up to 128GB unified memory and runs a 125B-parameter model locally; Snapdragon X Elite caps at 64GB and is officially positioned for 13B+ LLMs at about 30 tokens/s (per official data).
- Power and form: the Qualcomm route targets thin, light, low-power devices with long battery life; the RTX Spark route targets high compute, with the flagship Surface Laptop Ultra at about 2kg and under 18mm thin (per ifanr).
- Ecosystem: Qualcomm offers Hexagon SDK and Windows on ARM; RTX Spark natively supports CUDA, DLSS and G-SYNC, compatible with existing NVIDIA/Windows tools (per official materials).
Buying Advice
- Thin and light, long battery, daily AI assistance: choose the Qualcomm NPU route (Snapdragon X Elite Copilot+ PC) for note-taking, meeting transcription and light generation tasks (per Qualcomm's official positioning).
- Large local models, CUDA toolchain, heavy agent workloads: choose the RTX Spark route, which runs 100B+ parameter models locally and reuses the CUDA ecosystem (per ifanr citing official positioning).
- When comparing AI numbers: check the metric first (INT8 vs FP4); TOPS and PFLOPS under different metrics cannot be compared directly (per both official specifications).
Sources
- Qualcomm: Snapdragon X Elite specs (Hexagon NPU 45 TOPS)
- Qualcomm (May 2024): Snapdragon X Series is the Exclusive Platform to Power the Next Generation of Windows PCs with Copilot+
- ifanr (October 2026): RTX Spark platform official data (citing The Verge)