Google is intensifying its chip development efforts to lessen its reliance on external chip manufacturers for its cloud computing services. On Monday, Google introduced two new versions of its in-house Tensor Processing Units (TPUs), designed for AI tasks, expanding its chip capabilities to include secure inference and increased memory for extensive workloads.
These new units, the TPU 8tx and TPU 8i Max, are part of the company’s eighth generation of chips, which were initially unveiled in April of this year. They build upon the existing offerings by providing enhanced inference capacity and improved support for demanding training requirements.
The TPU 8tx, co-developed with AMD and based on the TPU 8t chip, is specifically engineered for confidentially training and operating AI models, as detailed in a Google blog post. The company explained that this chip “integrates security directly into the hardware layer, ensuring sensitive data and model weights are protected and isolated even during active use.”
Additionally, the chip features a custom inferencing network that allows it to perform reasoning concurrently with training, facilitating the real-time distillation and training of smaller models. According to Google, this functionality could be especially valuable for companies looking to train a model using their proprietary data without needing to transfer that data between different devices.
Meanwhile, the TPU 8i Max essentially combines two of the inference-focused TPU 8i chips, which Google released in April, into a single package. This effectively doubles its memory capacity, processing power, and bandwidth. The outcome is a chip that enables AI models to handle a broader spectrum of information during queries.
This chip launch occurs as major cloud providers like Microsoft and Amazon, along with AI research organizations such as OpenAI and Anthropic, are developing their own AI-specific semiconductors and accompanying software to reduce their reliance on Nvidia’s market dominance.
The news also comes at a particularly sensitive time for Nvidia, following its CEO Jensen Huang’s recent admission over the weekend of having an affair with an individual alleged to be a Chinese spy.