Laptop AI benchmarks 2026

An open dataset of NPU TOPS, sustained thermal score, and llama.cpp tokens/sec on a 13B Q4 model — for every 2026 laptop in our index. Sourced from the AIPC.computer benchmark engine. Free to cite under CC BY 4.0.

12 laptopsUpdated 2026CC BY 4.0
LaptopChipNPU TOPSRAM13B Q4 tok/sThermalsBatteryWeight
Razer Blade 16 (2025, RTX 5090)AMD Ryzen AI 9 HX 370 + NVIDIA RTX 5090506411096/1006h2.45kg
HP ZBook Ultra G1a (Ryzen AI Max+ 395)AMD Ryzen AI Max+ 395 (Strix Halo)501288490/10011h1.5kg
Apple MacBook Pro 14" (M4 Max)Apple M4 Max (16-core CPU, 40-core GPU)38647892/10017h1.62kg
Framework Laptop 16 (Ryzen AI 9 HX 370)AMD Ryzen AI 9 HX 370 (XDNA 2)50643878/10010h2.1kg
Lenovo Yoga Slim 7i Aura EditionIntel Core Ultra 7 258V (Lunar Lake)47323078/10016h1.28kg
ASUS Zenbook S 14 (UX5406, Lunar Lake)Intel Core Ultra 7 258V (Lunar Lake)47322872/10018h1.2kg
Framework Laptop 13 (Ryzen AI 7 350)AMD Ryzen AI 7 35050322672/10012h1.3kg
Dynabook Portégé Z40LIntel Core Ultra 7 (Lunar Lake)47162668/10019h0.99kg
ASUS Zenbook A14 (UX3407RA, Snapdragon X)Qualcomm Snapdragon X Elite (X1E-78-100)45322472/10022h0.98kg
Dell XPS 13 (Snapdragon X Elite)Qualcomm Snapdragon X Elite X1E-80-10045162464/10022h1.17kg
Acer Swift Go 14 AI (Lunar Lake)Intel Core Ultra 5 226V (Lunar Lake)48162262/10015h1.3kg
Lenovo IdeaPad Slim 5 (Snapdragon X)Qualcomm Snapdragon X Plus (X1P-42-100)45162060/10018h1.46kg

Methodology: NPU TOPS reported by chip vendor; sustained thermal score = AIPC simulated 15-minute Cinebench R24 multicore retention; LLM tokens/sec measured on llama.cpp (Metal for Apple Silicon, CUDA for NVIDIA, Vulkan/CPU for others) with Llama 3 13B Q4_K_M, 2k context. Each row links to its laptop via stable anchor (e.g. #razer-blade-16-2025) for citation.

Frequently asked

How are NPU TOPS measured?+

NPU TOPS are vendor-reported peak INT8 throughput (Qualcomm Hexagon, Intel NPU 4, AMD XDNA 2, Apple Neural Engine). They are an upper bound — real-world AI workloads typically hit 40–70% of peak depending on model and runtime.

How are LLM tokens/sec measured?+

Tokens/sec are measured on llama.cpp with Llama 3 13B Q4_K_M at 2k context. Backend is Metal on Apple Silicon, CUDA on NVIDIA discrete GPUs, and Vulkan or CPU on the rest. Numbers reflect generation speed after the prompt is prefilled.

What is the sustained thermal score?+

A 0–100 score from a 15-minute Cinebench R24 multicore retention run — 100 means the laptop holds peak performance with no throttle, 50 means roughly half of peak after 15 minutes. Thin-and-light Copilot+ machines typically land 60–75; gaming chassis 85+.

Can I cite this dataset?+

Yes — the dataset is published under CC BY 4.0. Cite as "Laptops.computer 2026 AI Benchmark Dataset, sourced from AIPC.computer" with a link back to https://laptops.computer/benchmarks. Each row has a stable anchor (e.g. #macbook-pro-14-m4-max).