Google Launches Eighth-Gen TPU Chips with Improved Price-to-Performance, Separates Training and Inference

Gate News message, April 22 — Google Cloud announced the release of its eighth-generation custom-built TPU (Tensor Processing Unit) chips on April 22. The new lineup includes TPU 8t, designed specifically for AI training tasks, and TPU 8i, optimized for AI inference workloads. Both chips will become available later this year. Google also unveiled new tools for building AI agents and announced a $750 million fund to drive enterprise AI adoption.

TPU 8t delivers 2.8 times the performance of Google’s previous-generation Ironwood TPU at the same price point. TPU 8i improves performance by 80% over its predecessor and incorporates static random-access memory (SRAM) architecture to deliver “cost-effective large-scale throughput and low latency, enabling millions of agents to run simultaneously,” according to CEO Sundar Pichai. Compared to Ironwood, both TPU 8t and TPU 8i achieve more than double the performance-per-watt efficiency, with TPU 8t improving by 124% and TPU 8i by 117%. Google optimized power efficiency across the entire technology stack and integrated dynamic power management systems that adjust consumption based on real-time demand.

Google’s first-party models now process over 160 billion tokens per minute through direct customer API calls, up from 100 billion last quarter. AI now generates 75% of all new code at Google, compared to 50% in fall of last year. Gemini Enterprise, Google’s enterprise offering, grew 40% quarter-over-quarter in paid monthly active users. The company expects to invest slightly more than half of its machine learning compute budget into cloud services by 2026 to better serve cloud customers and partners. Google is also expanding its collaboration with Broadcom to develop and supply custom TPU chips for future generations, as major tech firms seek alternatives to expensive and supply-constrained GPUs from NVIDIA and AMD.