Google rounds out its eighth generation of AI chips with a sealed-off version for sensitive data and a doubled chip for very long inputs
With the TPU 8tx and the TPU 8i Max, Google is targeting governments and tightly regulated companies, as well as agents that have to process huge amounts of text or video.
Google on Monday unveiled two new variants of its own AI chips, the Tensor Processing Units. That brings the eighth-generation family to four systems, Nasi Lemak Tech writes: alongside the existing 8t for training and 8i for fast reasoning come the 8tx and the 8i Max.
The TPU 8tx uses the same chip as the 8t, paired with fifth-generation AMD EPYC processors. According to Analytics India Magazine, security is built into the hardware so that data and model weights stay shielded even during computation. That should let tightly regulated sectors, such as governments, train and run their own models on sensitive data. Google claims up to two times lower latency than the 8t, with hardly any loss of performance.
The TPU 8i Max combines two 8i chips in a single package, offering twice the computing power, fast memory and memory bandwidth in the same rack space. It is intended for agents working with contexts of a million tokens, such as entire codebases or hours of video. Google gave no prices or delivery date. The company does say its TPU business has nearly a hundred customers and that it also wants to install the chips directly in customers' data centres.