The three
chips.
Every model you'll ever use runs on physical silicon. Three kinds of chip do the heavy lifting for AI — the GPU, the FPGA, and the ASIC — and each one trades flexibility against raw speed in a different way. Here's the honest map.
Training and inference are both just enormous piles of the same simple math — multiplying big grids of numbers, billions of times per second. The whole story of AI hardware is a race to do that one kind of math faster and cheaper. Where the chips differ is how much they're willing to specialize to get there.
Why not just a CPU?
The CPU — the general-purpose brain in every computer — is a brilliant generalist. It has a handful of very fast, very flexible cores designed to do one complicated thing at a time, in order. That's perfect for running an operating system or a spreadsheet. It's terrible for AI, because AI doesn't need one hard thing done quickly — it needs the same simple thing done a few trillion times in parallel. Ask a CPU to do that and its few cores become a bottleneck. So the industry reached for chips built around massive parallelism instead.
GPU — the parallel powerhouse
A GPU (Graphics Processing Unit) was originally built to draw video-game graphics, which — it turns out — is the same problem as AI: do a lot of simple math on a lot of numbers all at once. Instead of a handful of cores, a GPU has thousands. That makes it wildly good at the grid-multiplication that models live on.
GPUs are the default for a reason: they're powerful, widely available, and wrapped in a mature software ecosystem (NVIDIA's CUDA especially) that almost every AI tool is built to use. The tradeoffs are cost and power draw — the top data-center GPUs are expensive and hungry, and the whole world is competing to buy them.
GPU = flexible + fast + proven, but power-hungry and pricey. If you're not sure what to run AI on, the answer is almost always a GPU.
FPGA — the chip you can rewire
An FPGA (Field-Programmable Gate Array) is the strange, clever middle option. Most chips have their circuitry locked in at the factory. An FPGA doesn't — its internal logic can be reconfigured after it's manufactured, in the field (that's the "field-programmable" part). You're not writing software that runs on fixed hardware; you're rewiring the hardware itself to match your exact problem.
That buys two things: very low, predictable latency (great when a response must land in microseconds — think trading, signal processing, real-time systems) and the ability to build custom accelerators without committing to permanent silicon. The cost is effort: FPGAs are harder to program and rarely beat a GPU on raw throughput. They shine in specialized, latency-critical niches more than in mainstream model training.
People sometimes say "FPU" here — but an FPU is a Floating-Point Unit, a small math component inside a CPU, not a class of AI chip. The reconfigurable chip is the FPGA. Easy to mishear, worth getting right.
ASIC — built for one job, unbeatable at it
An ASIC (Application-Specific Integrated Circuit) is the opposite of an FPGA: a chip designed and permanently etched to do exactly one task as fast and efficiently as physically possible. No flexibility, no wasted transistors — every part of the silicon serves the one job. For that job, nothing is faster or more power-efficient.
Google's TPU (Tensor Processing Unit) is an ASIC purpose-built for neural networks. The chips in Bitcoin miners are ASICs too. The catch is the price of admission: designing and fabricating an ASIC costs enormous money and time up front, and once it's made you can't change it — if the workload shifts, the chip is a paperweight. ASICs only pay off at massive, stable scale, which is why you see them inside the biggest AI companies rather than on a desk.
ASIC = fastest and most efficient possible — for one fixed job, at a huge up-front cost and zero flexibility.
Flexible ⟷ fast
Line the four up and the pattern is clean: the more you specialize a chip, the faster and more efficient it gets at its target job — and the less it can do anything else.
Most flexible
Does anything, in order, with a few powerful cores. Weakest at AI's parallel math.
Flexible + parallel
Thousands of cores, mature software. The mainstream AI workhorse. Power-hungry.
Reconfigurable
Rewire the hardware per task. Low latency, custom fit. Harder to program.
Most specialized
Etched for one job. Fastest + most efficient at it. Costly, permanent (e.g. TPU).
You'll also hear NPU (Neural Processing Unit) — a small AI-specialized block now built into phones and laptops to run models efficiently on-device. It's really just an ASIC-style accelerator that lives inside a consumer chip. Same idea, smaller scale.
You're almost certainly choosing a GPU
For every business short of a hyperscaler, the practical choice is a GPU — rented in the cloud, or owned in a rack in your office. FPGAs and ASICs matter at the extremes; the middle of the market runs on GPUs. The real decision isn't which chip, it's whose chip, and where it lives — a rented frontier model in someone else's data center, or your own hardware under your own roof. That's the final piece.
Common questions
What chips are used for AI?
The three main AI chips are the GPU, FPGA, and ASIC. GPUs are the mainstream workhorse (thousands of parallel cores, mature software); FPGAs are reconfigurable chips for low-latency custom work; and ASICs are fixed-function chips built for one job at maximum speed (Google's TPU is an ASIC). CPUs handle general control but are weak at AI's parallel math.
What's the difference between a GPU and an ASIC?
A GPU is flexible — thousands of cores that handle many kinds of parallel math, which makes it the default for training and inference. An ASIC is etched to do exactly one task as fast and efficiently as physically possible, with no flexibility. ASICs beat GPUs on their target job but cost enormous money to design and can't be changed afterward.
Is an FPGA better than a GPU for AI?
Usually not for mainstream AI. An FPGA can be rewired after manufacturing, giving very low, predictable latency for specialized real-time tasks, but it's harder to program and rarely beats a GPU on raw throughput. Most AI runs on GPUs; FPGAs shine in latency-critical niches.
What is a TPU or NPU?
A TPU (Tensor Processing Unit) is Google's ASIC built specifically for neural networks. An NPU (Neural Processing Unit) is a small AI accelerator built into phones and laptops to run models efficiently on-device. Both are specialized, ASIC-style chips rather than general-purpose processors.