Google is ramping up its chip manufacturing efforts to reduce its dependence on chips from other companies for its cloud computing business. Google on Monday announced two new variants of its tensor processing unit (TPU). This is a chip developed in-house for AI workloads and expands the chip portfolio to additional use cases: confidential inference and increased memory capacity for large-scale workloads.
The new units, TPU 8tx and TPU 8i Max, are part of the company’s 8th generation chips introduced in April of this year, and build on what those chips had to offer: more inference power and support for large-scale training requirements.
Designed in collaboration with AMD, the TPU 8tx is based on the TPU 8t chip and is specialized for confidential training and execution of AI models, Google said in a blog post. The chip “builds security directly into the hardware layer, protecting and isolating the weight of sensitive data and models even during active use,” the company wrote.
The chip also features a custom inference network that can perform inference in parallel with training, so it can be used to extract and train small models on the fly. Google says this feature could be useful for businesses that want to train models on their own data without moving data between devices.
The TPU 8i Max, on the other hand, essentially combines the two inference-focused TPU 8i chips the company launched in April, doubling the memory capacity, processing power, and bandwidth, all in a single package. The result is a chip that allows AI models to process a wide range of information during queries.
The chip’s launch comes as hyperscalers like Microsoft and Amazon, as well as AI labs like OpenAI and Anthropic, are working to develop their own AI-specific semiconductors and accompanying software to reduce their dependence on NVIDIA’s dominance.
The news comes at a particularly precarious time for NVIDIA. NVIDIA CEO Jensen Huang admitted over the weekend that he had an affair with an alleged Chinese spy.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
