TPU
An AI accelerator chip designed by Google. It is an ASIC dedicated to the matrix calculations of neural networks and is not sold as a card.
TPU stands for Tensor Processing Unit. It is a chip designed by Google specifically to run neural networks. Google has used it in its own data centers since 2015 and unveiled it publicly in 2016. Several generations have followed since, each more powerful and more efficient.
A TPU is in fact an ASIC, just like a Bitcoin miner: instead of being versatile like a GPU, it is built for a single family of calculations, the matrix multiplications at the heart of AI, which it chains together with great energy efficiency. TPUs are grouped by the hundreds or thousands into clusters connected by a very fast network, to train and run very large models.
For an individual or a business, the difference is very concrete: data center TPUs are not available to buy. Google runs them mainly in its own facilities and makes them available remotely; only small chips intended for embedded electronics are sold on the open market. If you want AI at home, on your own hardware, the answer is GPUs, which also benefit from the CUDA software ecosystem, compatible with almost all tools and models. It is the same reasoning as ASIC versus GPU in mining: specialisation versus versatility.
We offer NVIDIA GPU workstations and servers to host your AI locally, without depending on an outside service.