Explain like I'm five · Systems

What is an AI Accelerator (TPU/NPU)?

AI accelerators like TPUs and NPUs, told as a story about a robot arm built for exactly one calculation.

even the GPU crew strains under one specific calculation, repeated endlessly'multiply and add' has to repeat hundreds of millions of timesthe general crew (GPU) carries features this job never uses, so multiply-and-add alone wastes effort

AI work pours out one calculation — multiply and add — by the hundreds of millions.The general crew (GPU) carries features this job never uses, so multiply-and-add alone wastes effort.

adding GPU crew adds power and chip area for unused features toofor multiply-and-add alone, the GPU's hands are 'too versatile' for the job

Adding GPU crew also adds power spent on features this job never uses.For multiply-and-add alone, the GPU's hands are too versatile.

an AI accelerator = a robot arm built only for multiply-and-add (matrix math)the arm suits little else — but on the same power it usually does more matrix math than a GPU

An AI accelerator = a robot arm built only for multiply-and-add (matrix math).The dedicated arm (AI accelerator) suits little else — but on the same power it usually does more matrix math than a GPU.

Built for Matrix Multiplythe core move behind neural-network math
Specializationdoing only one thing gets more math out of the same power
NPUthe small dedicated arm inside your phone
Tradeoffthe dedicated arm is hard to use for anything else
FLEXIBILITY vs SPEEDCPU: does anything, slowGPU: many similar jobs, fasterTPU: Google's matrix-multiply armNPU: the small arm inside your phonefurther down the board, each chip does fewer kinds of work, usually on less power per calculation

On the same power, the more kinds of work a chip can do, the slower it is at any one of them.Further down the board, each chip does fewer kinds of work, usually on less power per calculation.

the arm takes on the matrix mathother logic and control still belong to the chief (CPU)the dedicated arm works alongside the CPU and GPU — it doesn't do everything alone

The dedicated arm doesn't replace the CPU or GPU.The arm takes the matrix math; the CPU and GPU take the rest.

In one breath

AI accelerators (TPUs, NPUs) = dedicated hardware built for exactly one kind of calculation (mainly matrix multiplication), made to do that one thing on less power than a CPU or GPU.

A TPU (Tensor Processing Unit) is an ASIC Google designed specifically for the matrix and tensor math behind neural networks. NPU (Neural Processing Unit) is a broad name for neural-network accelerators, widely used to run inference on-device — in a smartphone, say — at low power. Think of these chips as a spectrum: CPU for generality, GPU for massive parallelism, TPU/NPU for hardware optimized to one specific computation pattern. Sacrificing generality buys power efficiency on that one pattern: far better than a CPU, and usually better than a GPU.

When grown-ups say it

TPU
Google's robot arm built for matrix multiply. Specialized for neural-network training and inference.
NPU
The small dedicated arm inside your phone. Runs inference at low power.
Specialization
Built to do just one thing. Trades generality for speed and efficiency.
GPU/GPGPU
A crew doing the same simple task all at once. One step earlier on the accelerator spectrum. → the hundred-hand helper crew
CPU Scheduling
The rule deciding who gets the lift first. The program usually picks what runs on the accelerator; the OS scheduler sets CPU turns. On that page the chief is the OS scheduler, not the CPU. → who gets the lift first
Cache Memory
A small toolbox holding only the frequent tools. The dedicated arm has its own small toolbox too (on-chip memory). → the toolbox kept right beside you
Time Complexity
The ruler for how time grows as the amount of math grows. The matrix-multiply workload itself is unchanged — dedicated hardware just shrinks the constant factor. → how much slower as cars pile up
GPU in the AI Field
The fast hardware (perch) that runs an AI model (the parrot). Covers accelerators briefly, from the inference-cost angle. → AI inference
Next story
One Car, Many Hands →
Want another term as a picture book? Request →