Published 2026-09-27 · Last updated 2026-09-27

AI PC Glossary 2026

The short version: an NPU is your laptop's AI engine, TOPS measures its speed, memory bandwidth decides how fast local models generate text, and quantization decides how big a model fits in RAM. Every term that matters for buying and using a 2026 AI PC is defined below — 36 entries, no jargon.

NPU & Acceleration

NPU (Neural Processing Unit)#: A dedicated processor block built for neural-network math. Unlike the CPU or GPU, it runs AI inference at a fraction of the power, which is why AI features can run all day without killing battery life.
TOPS#: Tera-operations per second — the headline measure of NPU throughput. One TOPS equals one trillion operations per second. Compare TOPS only at the same precision (see INT8 TOPS vs FP16 TOPS).
INT8 / INT4 TOPS#: NPU throughput measured on 8-bit or 4-bit integer math. Most vendor TOPS figures are INT8; a few vendors quote FP16 or sparse numbers, which inflate comparisons. The Copilot+ 40-TOPS bar is an INT8 figure.
Copilot+ PC (40+ TOPS)#: Microsoft's certification tier requiring at least 40 NPU TOPS (INT8), 16GB RAM, and 256GB storage. It unlocks Recall, live captions and translation, Cocreator, and other on-device Windows AI features.
NPU vs iGPU vs dGPU#: The NPU is the efficiency king (low power, sustained AI), the integrated GPU adds flexible parallel compute, and a discrete GPU delivers raw speed at the cost of battery. For all-day AI features the NPU wins; for heavy generative work the GPU wins.
DirectML#: Microsoft's hardware-accelerated DirectX 12 layer for machine learning on Windows. It lets models run across GPUs and NPUs from any vendor without vendor-specific code paths.
ONNX Runtime#: Microsoft's cross-platform inference engine. Models exported to ONNX format can target the NPU, GPU, or CPU with execution providers, and it is the most common path for shipping NPU-accelerated Windows apps.
Windows ML#: Windows' built-in ML inference API that automatically picks the best local execution hardware (NPU, GPU, or CPU) for a model. A key part of how Windows 11 routes AI work on Copilot+ machines.
XDNA 2 (AMD)#: AMD's second-generation NPU architecture in Ryzen AI 400 chips, delivering about 60 TOPS with block-FP16 support. It competes with Snapdragon's Hexagon on both throughput and efficiency.
Hexagon NPU (Qualcomm)#: Qualcomm's NPU architecture inside Snapdragon X and X2 chips. The X2 generation reaches 80 TOPS, the highest sustained NPU figure in a shipping Windows laptop, and drives Copilot+ features plus local inference via the QNN/FastRPC stack.

Memory & Bandwidth

Unified Memory Architecture#: A design where CPU, GPU, and NPU share one physical pool of memory instead of copying data between separate pools. Apple Silicon popularized it on laptops; it is the main reason a 64GB MacBook can run a 70B model that no 64GB Windows laptop with discrete graphics can match for simplicity.
LPDDR5X#: Low-power double-data-rate memory soldered to the board, typically LPDDR5X-8533 in 2026 laptops. It is fast and efficient but not upgradeable — which is why RAM decisions at purchase time are permanent.
Memory Bandwidth#: How many gigabytes per second the memory system can feed to compute. For local LLMs it is the single best predictor of generation speed: tokens/sec ≈ bandwidth ÷ model size in GB. The 2026 spread runs from ~80GB/s thin-and-lights to ~540GB/s M5 Max-class machines.
CAMM2 / LPCAMM2#: The replaceable memory module standard replacing soldered LPDDR on some 2026 laptops. It brings upgradeability back to high-speed memory, though availability remains limited and modules cost more than soldered RAM.
KV Cache#: The memory a model keeps for previously processed tokens so generation stays fast. It grows linearly with context length and can rival the model weights themselves at 128K context. Calculator formula: 2 × layers × KV heads × head dim × 2 bytes per token.
Grouped-Query Attention (GQA)#: An attention optimization where multiple query heads share fewer key/value heads, shrinking the KV cache by 4–8x versus classic multi-head attention. Every current open model (Llama 3.x, Qwen 2.5, Mistral) uses it.
RAMageddon (2026 memory shortage)#: The 2026 DRAM and NAND shortage driven by AI datacenter demand, which pushed laptop RAM prices sharply up with no predictable end. Practical consequence: buy the RAM you need upfront — upgrading later costs more, if the option exists at all.
DDR5 vs LPDDR5X#: DDR5 uses replaceable SO-DIMM modules and trades speed and power for upgradeability; LPDDR5X is soldered, faster, and more efficient. Performance-focused 2026 laptops increasingly ship LPDDR5X only, making the RAM choice at purchase permanent.

Silicon & Platforms

Snapdragon X2 Elite / Plus#: Qualcomm's 2026 flagship Windows-on-ARM platform with the third-generation Oryon CPU and an 80-TOPS Hexagon NPU. It leads Windows laptops on battery life and sustained NPU throughput; the X2 Plus ships in Surface Pro 12 and Surface Laptop 13 from October 2026.
Oryon CPU cores#: Qualcomm's custom CPU cores (acquired via Nuvia) used in Snapdragon X-series chips. X2 Elite pairs up to 18 Oryon cores with the 80-TOPS NPU, delivering Apple-class single-thread performance on Windows.
Panther Lake / Core Ultra X9#: Intel's 2026 platform (Core Ultra X9 class) with an NPU around 50 TOPS and the broadest Windows application compatibility, since it runs x86 natively. The safe pick for buyers who cannot risk ARM app-compatibility edge cases.
Ryzen AI 400#: AMD's 2026 mobile platform with the XDNA 2 NPU at about 60 TOPS. Its strength is combining capable AI silicon with iGPU gaming performance — the all-rounder of the 2026 chip race.
Apple M5 / M5 Pro / M5 Max#: Apple's 2026 chips with the Neural Accelerators in each GPU core and roughly 4x the AI performance of M4. M5 starts at 150GB/s-class memory bandwidth; Pro and Max tiers reach ~270 and ~540GB/s, making them the strongest local-LLM laptops you can buy.
Neural Engine (Apple)#: Apple's dedicated NPU block, rated in billions of operations per second rather than TOPS. On M5, AI work also flows through Neural Accelerators embedded in each GPU core, so Apple reports combined performance differently from PC vendors.
FastRPC#: Qualcomm's communication layer between the CPU and Hexagon NPU on Windows-on-ARM. It is how frameworks like llama.cpp reach the NPU on Snapdragon, and the path Linux distributions are building support for through 2026–2027.
WSL2 on ARM#: Windows Subsystem for Linux running on Snapdragon X2. Native ARM Linux toolchains run at full speed; x86-only Linux containers run under emulation with a measurable penalty — a key check for developers before buying ARM.

Quantization & Inference

Quantization#: Compressing a model's weights from 16-bit floats to 8-, 6-, 4- or 3-bit integers. It cuts memory roughly in proportion (FP16 = 2 bytes/param, Q4 ≈ 0.6 bytes/param) and is what makes 70B models fit on a laptop at all. Q4_K_M is the usual quality/speed sweet spot.
GGUF#: The file format used by llama.cpp and most local-LLM tooling. A GGUF file bundles weights, tokenizer, and metadata for a specific quantization level — the '-Q4_K_M.gguf' suffix you see on Hugging Face.
llama.cpp#: The open-source inference engine that made local LLMs practical on consumer hardware. It runs GGUF models on CPU, GPU, and increasingly NPU, and powers most Mac and Windows local-LLM frontends.
MLX#: Apple's machine-learning framework for Apple Silicon. On M5 Macs it is the fastest path to local inference because it exploits unified memory and the Neural Accelerators directly.
Tokens per second#: The standard speed measure for local LLM generation — how many word-pieces the model outputs per second. Roughly equal to memory bandwidth divided by model size in GB, which is why bandwidth predicts speed better than TOPS.
Time-to-First-Token (TTFT)#: How long a model takes to produce its first output token after receiving a prompt. Dominated by prompt processing (compute-bound), unlike generation speed (bandwidth-bound). Matters most for long-document Q&A and agents.
Context window#: How many tokens (≈0.75 words each) a model can consider at once — 4K, 8K, 32K, 128K. Longer contexts need exponentially more KV-cache memory, which is why 'runs on 16GB' claims usually assume a short context.
Local LLM#: A large language model running entirely on your device — no cloud, no subscription, no data leaving the machine. The 2026 threshold experience: 8B models on 16GB laptops, 32B models on 32–64GB, 70B on 64GB+ with high bandwidth.
Distilled model#: A smaller model trained to mimic a larger one — e.g. DeepSeek-R1 32B distilled from the 671B original. Distillation keeps much of the reasoning quality at laptop-runnable size, which is why 32B is the 2026 local sweet spot.
Agentic AI workload#: AI that plans and executes multi-step tasks (browsing, file operations, tool calls) rather than answering once. Agentic loops multiply inference calls, which is why 80-TOPS NPUs and high memory bandwidth matter for the next Windows feature wave.

Put the terms to work

Use the Local LLM & AI VRAM calculator to turn KV cache and quantization into an exact RAM number, check the 2026 AI PC buyer's guide for which chip to buy, and compare raw figures on the benchmarks page.

Definitions reflect platform specifications and shipping products as of 2026-09-27. Vendor TOPS figures are INT8 unless stated.