AI accelerators
H100, B200, MI300X, the TPU, Trainium, Gaudi, Cerebras WSE and the NPUs of phones and laptops. 96 chips on EveryChip so far.
An AI accelerator is a chip designed to train or run neural networks, the layered models behind image recognition, speech and large language models. Most of that work is multiplying large tables of numbers, called matrices. Accelerators pack in thousands of multiply units, feed them from fast memory and use low-precision number formats that save power. Nvidia's data centre GPUs, Google's TPU and many start-up chips belong here.
What makes them fast
A neural network layer multiplies a matrix of inputs by a matrix of learned weights. Accelerators have dedicated matrix units, such as Nvidia's tensor cores or the systolic array of Google's TPU, in which numbers flow through a grid of multipliers. They use formats with fewer bits, such as 16-bit, 8-bit and even 4-bit numbers, so each operation costs less energy and more fit on the chip.
Memory and networks matter as much
A large model has billions of weights that must be read again and again. That is why AI accelerators sit beside stacks of High Bandwidth Memory, and why training clusters link thousands of chips with fast networks. Nvidia's H100 of 2022 has 80 billion transistors; its Blackwell B200 joins two dies with 208 billion. Demand for these chips since 2023 has strained supplies of HBM and advanced packaging.
Famous AI accelerator chips
Further reading
Photo credits on this page
The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.
- AMD Instinct: Мой Компьютер, CC BY 3.0, source
- Tesla Dojo: Steve Jurvetson, CC BY 2.0, source
- Tensor Processing Unit: Norman P. Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan,, CC BY 4.0, source
- Hopper: 极客湾Geekerwan, CC BY 3.0, source
- AI Engine: Dmitry Grigoriev, CC BY 2.0, source
- Blackwell: 极客湾Geekerwan, CC BY 3.0, source