This review is built from Qualcomm’s and Intel’s published specifications, third-party reviews and community benchmarks rather than hands-on testing. Figures are attributed in the text, and some speed figures are modelled estimates, labelled as such.
The Snapdragon X2 Elite and Intel’s Panther Lake, sold as Core Ultra Series 3, are the two chips behind most new thin-and-light Windows laptops in 2026. Both are marketed heavily on AI: Qualcomm on its 80 TOPS NPU, Intel on a platform figure of up to 180 TOPS. If you want a slim laptop that can also run language models locally, these are the chips you will be choosing between.
Here is the honest answer. Both are excellent laptop chips, and both are modest local LLM machines. Text generation is limited by memory bandwidth, not TOPS, and these chips have between roughly 120 and 230 GB/s, a fraction of a discrete GPU. They run small models, from 3B to around 14B, at usable speeds, and mixture-of-experts models better than their size suggests. They will not replace a desktop GPU or a 128GB unified-memory machine. Buy one for battery life and portability, with small local models as a bonus.
Table of Contents
- The Two Chips at a Glance
- Snapdragon X2 Elite TOPS vs Real Tokens
- Memory Bandwidth: Snapdragon X2 Elite Leads
- Snapdragon X2 Elite NPU in Practice
- Panther Lake for Local Models
- Snapdragon X2 Elite and Panther Lake Speeds
- Snapdragon X2 Elite and Panther Lake Memory Options
- Snapdragon X2 Elite Software vs x86
- Snapdragon X2 Elite Battery Life vs Panther Lake
- Models That Suit a Snapdragon X2 Elite Laptop
- Snapdragon X2 Elite vs Strix Halo, RTX Spark and Apple
- Everyday AI Uses Worth Running Locally
- Snapdragon X2 Elite Buying Advice
- Who Should Buy What
- FAQ
- Setting Up Local AI on Either Laptop
- Verdict
- Sources and Further Reading

The Two Chips at a Glance
| Snapdragon X2 Elite Extreme | Intel Core Ultra X9 388H (Panther Lake) | |
|---|---|---|
| CPU | Up to 18 Arm cores | 16 x86 cores (4+8+4) |
| NPU | Hexagon, 80 TOPS | NPU 5, 50 TOPS |
| Integrated GPU | Adreno | Arc B390, 12 Xe3 cores |
| Memory bandwidth | About 228 GB/s | About 154 GB/s with LPDDR5X-9600 |
| Maximum memory | Up to 64GB LPDDR5X | Up to 96GB LPDDR5X or 128GB DDR5 |
| Operating system | Windows 11 on Arm | Windows 11 or Linux, x86 |
MadeByAgents listed the X2 Elite Extreme at 228 GB/s with up to 64GB of LPDDR5X. LLMRun’s profile of the Core Ultra X9 388H showed 32GB of LPDDR5X-9600 on a 128-bit bus at 153.6 GB/s. HotHardware reported Intel’s platform limits of up to 96GB of LPDDR5X or 128GB of DDR5.
Snapdragon X2 Elite TOPS vs Real Tokens
TOPS measures peak arithmetic. It is a good guide for vision models, image effects and other workloads that do a lot of maths per byte of data. It is a poor guide for language model generation, which does very little maths per byte and spends most of its time reading weights from memory.
WhySoGeek put it plainly: LLM text generation is bound by memory bandwidth, not raw TOPS, and most mainstream runtimes do not even use the NPU yet. LLMRun was more direct still in its profile of the Qualcomm chip: the 80 TOPS NPU is not used by llama.cpp or Ollama, so local LLM speed depends on the 228 GB/s memory bandwidth, with models running on the CPU or Adreno GPU.
So when a laptop advertises 80 or 180 TOPS, read it as a measure of how well it handles Windows AI features, not how fast it chats.
Memory Bandwidth: Snapdragon X2 Elite Leads
On bandwidth, Qualcomm leads. Its top chip offers about 228 GB/s against about 154 GB/s for Panther Lake with the fastest supported LPDDR5X. For the same model, that suggests the Snapdragon X2 Elite can generate text roughly 45% faster, if the software uses the bandwidth well.
A Framework community thread on a Panther Lake laptop showed how much memory speed matters within Intel’s range: theoretical bandwidth rises from 89.6 GB/s with DDR5-5600 to 119.5 with LPDDR5X-7467, 136.5 with LPDDR5X-8533 and 153.6 with LPDDR5X-9600. The thread’s conclusion was that LLM workloads are so demanding of bandwidth that more wattage or compute often does not help unless you can feed it more bandwidth.
For comparison, a used desktop RTX 3090 has 936 GB/s and AMD’s Strix Halo laptops have 256 GB/s with up to 128GB. These thin laptops sit well below both.
Snapdragon X2 Elite NPU in Practice
For a long time the NPU on Snapdragon laptops was a marketing feature for local LLMs. BestLLMfor described it as sitting mostly idle for 18 months, because llama.cpp, Ollama and LM Studio targeted CPUs or CUDA rather than Qualcomm’s runtime.
That changed in late 2025. BestLLMfor reported that tools such as the Nexa SDK and Microsoft’s Foundry Local can now run models on the NPU, with Qualcomm’s March 2026 NPU driver as the minimum version. Its picks included Qwen3-8B in a hybrid INT4 format through Nexa for chat on battery, and Phi-4-mini through Foundry Local for zero-setup app integration.
The NPU’s real advantage is efficiency. Qualcomm claims up to 3.8 times better NPU performance per watt than Windows competitors. For a laptop running a small model all day on battery, that matters more than peak speed.
Panther Lake for Local Models
Intel launched Panther Lake at CES 2026 as the first chips on its 18A process. HotHardware’s review was enthusiastic, describing the 12-core Xe3 GPU as a huge leap and NPU 5 as built for current local AI workloads. Intel claimed its NPU runs LLM inference 4.3 times faster than AMD’s XDNA 2 NPU in the Ryzen AI 9 HX 370, and twice as fast as the NPU in the previous Core Ultra 9 285H.
The Arc B390 integrated GPU is the more useful part for local models today, because llama.cpp can use it through Vulkan or SYCL. PC Gamer’s analysis described the Xe3 architecture as keeping an Arc A770 desktop card in its sights on some measures, while noting its biggest weakness is reliance on LPDDR5X memory latency and bandwidth.
Panther Lake’s other strength is x86 compatibility. Every Windows and Linux AI tool runs natively, which removes the Arm compatibility questions that still occasionally affect Qualcomm laptops.
Snapdragon X2 Elite and Panther Lake Speeds
Measured results on these exact chips are still limited. These figures give a reasonable picture.
| Chip and model | Speed | Source and type |
|---|---|---|
| Snapdragon X Elite, 8B model, CPU | About 5 to 10 tok/s | WhySoGeek, first-generation chip |
| Panther Lake X9 388H, Qwen3.5 9B Q4 | About 12 tok/s | LLMRun estimate |
| Panther Lake X9 388H, Qwen3.6 35B-A3B Q4 | About 22 tok/s | LLMRun estimate |
| Panther Lake X9 388H, 27B dense Q4 | About 4.4 tok/s | LLMRun estimate |
| Panther Lake vs Lunar Lake, llama.cpp Vulkan | 4.8 vs 3.9 tok/s | Linux test via LAVX |
The X2 Elite should run ahead of the first-generation figure thanks to its higher bandwidth, though comparable measured numbers are scarce. On Geekbench AI’s quantised INT8 test, Tech Insider reported that the X2 Elite dominated Panther Lake, while trailing in FP32 work.
The pattern is consistent with bandwidth: 8B to 9B models run at reading pace or faster, mixture-of-experts models with a few billion active parameters run surprisingly well, and dense 27B models are too slow for comfortable use.
Snapdragon X2 Elite and Panther Lake Memory Options
Memory configuration matters more on these laptops than almost any other specification for AI.
On Intel, the full Arc B390 GPU only appears with fast LPDDR5X. ComputerBase reported that if an OEM ships slower memory, Intel brands the graphics simply as Intel Graphics rather than Arc. TweakTown reported a test where the B390 ran about 16 to 18% slower on DDR5-5600 SO-DIMMs than on LPDDR5X. Look for the Arc B390 name and LPDDR5X-9600 in the specification.
On Qualcomm, look at memory capacity. Many laptops ship with 16GB, which leaves little room for models after Windows takes its share. LLMRun’s profile used a 48GB configuration, with about 44GB usable for models. For local AI, 32GB is a sensible minimum on either platform.
Snapdragon X2 Elite Software vs x86
Panther Lake laptops run x86 Windows or Linux, so llama.cpp, Ollama, LM Studio, OpenVINO and every other tool work natively. Intel’s OpenVINO can also target the NPU for supported models.
Snapdragon laptops run Windows 11 on Arm. Most mainstream AI tools now have Arm builds, and LM Studio and Ollama run on them, though often on the CPU. To use the NPU, you need Qualcomm-aware runtimes such as Nexa or Foundry Local, which support a narrower set of models. Older x86-only Python packages may still cause friction.
Linux support on Qualcomm laptops remains limited compared with Intel. If you work mainly in Linux, Panther Lake is the safer choice.
Snapdragon X2 Elite Battery Life vs Panther Lake
Both chips are very efficient. Phoronix found the Core Ultra X7 358H averaged about 15W across 300 benchmarks, peaking at 64W, and often delivered the best performance while drawing less than AMD Strix Point laptops. NoobFeed praised the MSI Prestige 14 AI+ for exceptional battery life.
Qualcomm has built its reputation on battery life and consistent performance on battery. HotHardware noted that Qualcomm works with partners to minimise the gap between plugged-in and battery performance, which many x86 laptops do not.
Running a model continuously drains any laptop faster. The advantage of these chips is that light AI use, such as a small model answering occasional questions, barely dents a full day of battery.
Models That Suit a Snapdragon X2 Elite Laptop
- 3B to 4B models such as Phi-4-mini. Fast on either chip, and the NPU-friendly choice on Qualcomm.
- 8B to 9B models at Q4. Reading-pace or better, good for drafting and summarising.
- 30B to 35B mixture-of-experts models with 32GB or more. About 22 tokens per second estimated on Panther Lake, with strong quality for the speed.
- Dense 14B models. Usable but slower; pick these only when quality matters more than speed.
Our best local LLMs guide lists current options by size.
Snapdragon X2 Elite vs Strix Halo, RTX Spark and Apple
If local models are a priority rather than a bonus, other laptops do far more. AMD’s Strix Halo laptops, such as the ROG Flow Z13, offer 256 GB/s and up to 128GB. NVIDIA’s RTX Spark laptops, led by the Surface Laptop Ultra, bring CUDA and up to 128GB. Apple’s MacBook Pro M5 Max reaches 614 GB/s.
Those machines are heavier, more expensive and use more power. The Snapdragon and Panther Lake laptops win on weight, price, battery life and quiet operation. Decide which matters more to you before comparing AI numbers.
Everyday AI Uses Worth Running Locally
It helps to be specific about what a small local model on a thin laptop is actually good for. Drafting and rewriting emails, summarising a long document you cannot paste into a cloud service, extracting action items from meeting notes, translating short passages and answering questions about a local folder of files are all well within reach of an 8B model at reading speed.
Coding help is more mixed. Autocomplete and explanations of short functions work well with small coding models. Agentic coding, where a tool reads large parts of a codebase and plans changes, sends long contexts on every step, and these laptops process long prompts slowly. For that, a cloud model or a desktop GPU is still the better tool.
The strongest argument for running locally on either a Snapdragon X2 Elite or Panther Lake laptop is privacy on the move. On a train, in a client’s office or on a flight without Wi-Fi, a small local model keeps working and keeps your documents on the device. Treat it as a capable assistant for light tasks, not a replacement for frontier models.
Snapdragon X2 Elite Buying Advice
For the Qualcomm side, choose the X2 Elite Extreme or a high-bin X2 Elite with at least 32GB of memory. Check that the laptop supports the latest NPU drivers, and that the tools you rely on have Arm builds.
For Intel, choose a Core Ultra X7 or X9 model with Arc B390 graphics and LPDDR5X-9600, and 32GB or more. Avoid models that list Intel Graphics without the Arc name if you plan to use the GPU for models.
Who Should Buy What
| If you are | Start with | Why | Link |
|---|---|---|---|
| Wanting the best battery life with small local models | Snapdragon X2 Elite laptop, 32GB+ | Higher bandwidth, efficient NPU | Check laptops |
| Needing x86 or Linux compatibility | Panther Lake with Arc B390 and LPDDR5X | Every tool runs natively | Check laptops |
| Wanting big models in a laptop | Strix Halo laptop | Up to 128GB at 256 GB/s | Check price |
| Wanting the fastest laptop for local AI | MacBook Pro M5 Max | 614 GB/s, up to 128GB | Check price |
Frequently Asked Questions About Snapdragon X2 Elite and Panther Lake
Which is better for local LLMs?
The Snapdragon X2 Elite has more memory bandwidth, about 228 GB/s against roughly 154 GB/s, so it should generate text faster. Panther Lake has better software compatibility.
Does the NPU speed up chat models?
Only with NPU-aware runtimes such as Nexa or Foundry Local. llama.cpp and Ollama mostly use the CPU or GPU.
How much memory do I need?
At least 32GB for local AI. 16GB leaves little room for models.
What size models run well?
3B to 9B models comfortably, and 30B-class mixture-of-experts models with enough memory. Dense 27B models are slow.
Can these replace a desktop GPU?
No. A used desktop GPU is roughly ten times faster on an 8B model.
Which works better with Linux?
Panther Lake. Linux support on Snapdragon laptops is still limited.
Setting Up Local AI on Either Laptop
On Panther Lake, install LM Studio or Ollama and enable GPU offload through Vulkan, or use Intel’s OpenVINO tools for NPU-supported models. Update graphics drivers first.
On Snapdragon, update firmware and NPU drivers, then try Foundry Local for the simplest NPU setup, or the Nexa SDK for a wider choice of models. LM Studio works too, on the CPU and GPU.
On both, start with a small model, check speed on battery and plugged in, then move up in size.
Verdict
Choose a Snapdragon X2 Elite laptop if you want the best battery life and the fastest small-model performance in a thin Windows machine, and your tools run on Arm. Choose Panther Lake with Arc B390 and LPDDR5X if you need x86 or Linux and the widest software support.
The honest caveat is that both are modest local AI machines. TOPS figures overstate their LLM ability, bandwidth limits them to small and mixture-of-experts models, and many speed figures are still estimates.
Next step: decide whether local AI is a bonus or a priority. If it is a bonus, pick between these two on battery, weight and software. If it is a priority, look at Strix Halo, RTX Spark or an M5 Max instead.
Sources and Further Reading
- LLMRun: Snapdragon X2 Elite Extreme, for bandwidth, memory and the NPU caveat.
- BestLLMfor: local LLMs on Snapdragon, for NPU runtimes and model picks.
- HotHardware: Core Ultra X9 388H review, for Panther Lake architecture and results.
- Phoronix: Panther Lake on Linux, for efficiency across 300 benchmarks.
- Framework community: LLM performance on 358H, for memory bandwidth by configuration.
On this site: compare bigger laptops in our Surface Laptop Ultra, ROG Flow Z13 and MacBook Pro M5 Max reviews, and size memory with how much VRAM for AI.




