Qualcomm Snapdragon X Elite
Copilot+ PC reference SoC. 12-core Oryon CPU + Adreno GPU + Hexagon NPU at 45 TOPS INT8. The first ARM Windows laptop with serious NPU compute; runs Phi Silica and other on-device Microsoft AI features.
Affiliate disclosure: as an Amazon Associate and partner of other retailers, we earn from qualifying purchases. The verdict on this page is our editorial opinion; affiliate links never influence what we recommend.
Sub-scores sum to 40 / 1000. Headline = 40 × 0.70 (Estimated-confidence discount) = 28. This is an algorithmic performance-tier score — distinct from, and often lower than, the editorial “Our verdict” below, which weighs value and real-world fit (especially for hardware we haven’t measured yet). How scoring works →
Insufficient data — VRAM 0GB, bandwidth ? GB/s.
Plain-English: Doesn't fit modern chat models usefully — vision models won't fit.
Verdicts extrapolated from catalog VRAM + bandwidth + ecosystem flags. Hover any chip for the rationale. Want measured numbers? Submit your own run with runlocalai-bench --submit.
What it does well
The Qualcomm Snapdragon X Elite is the flagship Windows-on-ARM laptop SoC and Qualcomm's first credible Copilot+ PC platform — 12 Oryon ARM CPU cores + Adreno X1 iGPU + dedicated Hexagon NPU rated at 45 TOPS. Ships in laptops from Microsoft Surface Pro/Laptop, Lenovo Yoga, ASUS Vivobook S, HP OmniBook, Dell XPS at $999-$2,499 retail. The platform delivers genuinely impressive battery life (15-20 hours real productivity work) and silent operation. For local AI on the NPU + iGPU, throughput on sub-7B models is competitive with Apple M4-class chips. Microsoft's tooling (DirectML, ONNX Runtime, Phi Silica APIs) targets the SDX Elite NPU first-class.
Where it breaks
- Windows-on-ARM software ecosystem is still maturing. x86 emulation works for most apps but performance ranges from "good" to "abysmal" depending on workload. Native ARM64 Windows apps + drivers + dev tooling has improved through 2024-2025 but remains thinner than Apple Silicon's mature macOS-on-ARM story.
- No CUDA, no ROCm, no Metal. DirectML + ONNX Runtime + IPEX-style approach. CUDA-locked stacks don't run.
- iGPU memory bandwidth limits decode. Shared LPDDR5X-8533 at ~136 GB/s — dramatically below discrete GPU bandwidth.
- Hard ceiling on model size. Memory caps at 32 GB unified in most laptops. 14B Q5 fits with limited context.
- NPU framework support is thin for LLMs. 45 TOPS is impressive on paper but real LLM throughput on the NPU is limited by software.
- Linux support is improving slowly. Windows is the right path in 2026.
Ideal model range
- Sweet spot: 7B FP16 / Q5 inference at ~20-35 tok/s on the iGPU + NPU.
- Sweet spot: Microsoft Copilot+ PC features (Phi Silica, Recall, Cocreator) tuned to the NPU.
- Sweet spot: Battery-life-friendly local AI for traveling professionals.
- Stretch: 13B Q4 with limited context (slow but functional).
- Bad fit: 14B+ FP16, 32B+ anything, fine-tuning, CUDA-required.
Verdict
Buy this (in laptop form) if you want the best Windows-on-ARM platform with credible AI features, exceptional battery life, and your stack is Windows + ONNX/DirectML compatible. SDX Elite is the right pick for the "Windows productivity laptop with AI features" segment.
Skip this if you need 14B+ models (jump to discrete GPU laptop), CUDA-locked, x86-locked Windows apps are critical (emulation can be rough), or you can use macOS (MacBook Pro ecosystem is more mature).
How it compares
- vs Apple M4 Pro / M4 → Apple Silicon has more mature ARM ecosystem (5+ years of macOS-on-ARM polish vs Microsoft's 2-3 years), faster iGPU compute, slightly less battery life. Pick by OS preference.
- vs AMD Ryzen AI 9 HX 370 → AMD x86 has more raw iGPU compute + 50 TOPS NPU at higher power draw. SDX Elite wins on battery + ARM efficiency. Pick by power priorities.
- vs Intel Lunar Lake 258V → Intel x86 with 48 TOPS NPU at slightly less battery. Both target Microsoft Copilot+ PC. Pick by laptop OEM availability.
- vs Snapdragon X Plus → SDX Plus is the lower-tier sibling at modest discount.
Overview
What the Snapdragon X Elite actually is, in local-AI terms
The Snapdragon X Elite is Qualcomm's first serious assault on the Windows-on-ARM laptop market, paired with a 45-TOPS-class Hexagon NPU that defines the floor of "Copilot+ PC" hardware in 2026. Up to 32 GB of LPDDR5X unified memory at ~135 GB/s, twelve Oryon CPU cores at high single-thread performance, and an Adreno GPU that's competitive with a low-power discrete card on graphics workloads but lags every comparable mobile NVIDIA / Apple chip on raw FLOPs.
For local AI specifically, the X Elite is NPU-first hardware. The CPU is fine, the GPU is okay, the NPU is the differentiator. That positioning makes the X Elite a strong target for on-device AI features and 1B-7B-class LLM inference but a weak target for the 32B-class workloads that define what "serious local AI" means on a RTX 4090 or Apple M4 Max.
Where it fits in the hardware ladder
The 2026 NPU-laptop ladder:
| Chip | NPU TOPS | Unified mem |
|---|---|---|
| Intel Lunar Lake (258V) | ~48 TOPS | up to 32 GB |
| Snapdragon X Elite | ~45 TOPS | up to 32 GB |
| AMD Ryzen AI 9 HX 370 | ~50 TOPS | up to 64 GB |
| Apple M4 Max NPU | ~38 TOPS | up to 128 GB |
The X Elite's NPU is competitive with the field on TOPS but its memory ceiling (32 GB) and bandwidth (135 GB/s) limit what models it can practically host.
vs cross-architecture laptop options:
| Chip | Mem | LLM ceiling |
|---|---|---|
| Snapdragon X Elite | 32 GB | 7B-13B comfortably |
| Apple M4 Max | 128 GB | 32B comfortably |
| RTX 5090 Mobile | 24 GB | 13B class plus |
Best use cases
- Copilot+ PC on-device AI features. Windows 11 ships first-party on-device features that target the NPU directly. The X Elite was designed for this.
- 1B-7B-class LLM inference on battery. The NPU + LPDDR5X unified memory let small models run with very low power draw — a real win vs forcing the same model through a CPU or laptop GPU.
- Embedding pipelines for desktop RAG. Sentence-transformers exported to ONNX run extremely fast on the Hexagon NPU via the QNN EP. See /stacks/offline-rag-workstation.
- Battery-life-first Windows laptop. The X Elite chassis runs cool and quiet for hours; a NVIDIA-laptop equivalent does not.
- Cross-platform on-device AI app development. The QNN EP behind ONNX Runtime gives you the same toolchain as Snapdragon mobile NPUs. See /stacks/android-on-device-ai.
What it can run
The realistic working set in May 2026:
| Model class | Quant | Context | Notes |
|---|---|---|---|
| 1B-3B | INT4 | 32K | excellent on NPU |
| 7B | INT4 / Q4_K_M | 16-32K | good on NPU + CPU hybrid |
| 13B | Q4_K_M | 8-16K | works on CPU + GPU; NPU op coverage gaps |
| 32B | — | — | does NOT fit |
| 70B | — | — | does NOT fit |
For the format-by-format runtime picture see /systems/quantization-formats.
OS support
| OS | Quality |
|---|---|
| Windows 11 ARM64 | excellent — primary target |
| Linux ARM64 | partial — boots, NPU access immature |
| macOS | unsupported (different vendor) |
Software / runtime support
The Snapdragon X Elite's local-AI software story in May 2026:
- ONNX Runtime + QNN EP — the canonical NPU path; production-ready
- Qualcomm AI Hub — the model-conversion toolchain; required for NPU deployment
- llama.cpp — CPU + Adreno (Vulkan / OpenCL) paths; NPU support arriving in 2026 via QNN integration
- DirectML EP in ONNX Runtime — works on the GPU; less tuned than QNN
- LM Studio — full Windows-on-ARM support via llama.cpp Vulkan path
- Ollama — Windows-on-ARM build available; Adreno GPU support via Vulkan
- PyTorch — CPU only; no CUDA, no NPU yet
Models that fit the NPU need to be converted via Qualcomm AI Hub; not every architecture is supported. Llama, Qwen, Mistral, Phi-3 all have well-tested paths.
What breaks first
- NPU op coverage gaps. Not every transformer op runs on the NPU; unsupported ops fall back to CPU and the heterogeneous transfer kills throughput. The QNN EP improving through 2026 is the realistic path.
- Memory ceiling. 32 GB total unified memory is shared with the OS, browsers, and everything else. A 13B model + 16K context + Windows is tight.
- Windows-on-ARM software compatibility. Most of the Python ML ecosystem now ships ARM64 wheels but exotic dependencies still fall through to CPU emulation.
- Driver / NPU SDK drift. Qualcomm ships QNN SDK and driver updates separately; mismatched versions silently disable the NPU plugin.
- Long-context decode on NPU. NPU SRAM budgets are tight; KV-cache for >4K context spills to system RAM.
Alternatives by intent
| If you want… | Reach for |
|---|---|
| Bigger memory in a laptop | Apple M4 Max (128 GB) |
| Higher TOPS NPU on x86 | Intel Lunar Lake 258V — see OpenVINO |
| AMD competitor | AMD Ryzen AI 9 HX 370 |
| Discrete GPU laptop | RTX 5090 Mobile — different stack, more throughput |
| Mobile NPU sibling | Snapdragon 8 Elite (phone) — same QNN toolchain |
Best pairings
- ONNX Runtime + QNN EP + 7B INT4 LLM — the canonical X Elite local-AI setup
- Qualcomm AI Hub for model conversion
- A Copilot+ PC laptop chassis with at least 32 GB unified memory — the smaller SKUs are usable but tight
- LM Studio Windows-on-ARM build for a GUI path
- Apple A18 Pro as the iOS counterpart in cross-platform mobile-AI shipping
Who should avoid the Snapdragon X Elite
- Operators planning to run 13B+ class LLMs daily. The 32 GB memory ceiling is the wrong tier; pick a Mac with 64 GB+ or a NVIDIA laptop.
- Anyone whose stack depends on CUDA. Wrong vendor entirely.
- Researchers doing Python ML iteration. ARM64 software compatibility is good but not as friction-free as x86 on Linux.
- Operators serving multi-user inference. The X Elite is a laptop; not the right shape for a server.
- Apple-ecosystem operators. Stay with Apple Silicon.
Related
- Stacks: /stacks/android-on-device-ai, /stacks/offline-rag-workstation
- System guides: /systems/quantization-formats, /setup
- Tools: Qualcomm AI Hub, ONNX Runtime, llama.cpp
- Errors: /errors/wsl2-gpu-not-detected
Some links above are affiliate links. We may earn a commission at no extra cost to you. How we make money.
Specs
| VRAM | 0 GB |
| System RAM (typical) | 32 GB |
| Power draw (peak) | 23 W |
| Released | 2024 |
| Backends |
Frequently asked
Does Qualcomm Snapdragon X Elite support CUDA?
Where next?
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify hardware specifications.