Google Unveils Two New AI Chips For the 'Agentic Era'

Source: Slashdot

Google’s new TPU lineup for the “agentic era”

Google announced two new Tensor Processing Units (TPUs) aimed at the emerging “agentic era.” The chips are purpose‑built, with one dedicated to training and the other to inference.

“With the rise of AI agents, we determined the community would benefit from chips individually specialized to the needs of training and serving,” said Amin Vahdat, Google senior vice president and chief technologist for AI and infrastructure, in a blog post. Both chips are expected to become available later this year.

Performance claims

• The training chip delivers 2.8 × the performance of the seventh‑generation Ironwood TPU (announced in November) at the same price.
• The inference processor is 80 % faster than its predecessor.

These figures come from a CNBC report that highlights Google’s shift from a single chip handling both training and inference to distinct processors, positioning the company against Nvidia in the AI‑hardware market.

Architecture details and competition

• Nvidia has hinted at its upcoming Groq 3 LPU hardware, which will rely heavily on static random‑access memory (SRAM).
• Cerebras, another AI chipmaker that recently filed to go public, also uses SRAM in its designs.
• Google’s inference chip, dubbed TPU 8i, similarly relies on SRAM, featuring 384 MB—three times the amount found in Ironwood.

According to Sundar Pichai, CEO of Alphabet, the architecture is designed “to deliver the massive throughput and low latency needed to concurrently run millions of agents cost‑effectively.”

References

• Google’s blog announcement:
• CNBC article: