Google Unveils TPU 8tx and TPU 8i Max, Widening Its AI Chip Lineup

Google added two artificial intelligence computing systems on Monday, the TPU 8tx and the TPU 8i Max, filling out an eighth generation Tensor Processing Unit lineup that now gives customers four configurations for training and running AI models. Google says it serves nearly 100 TPU customers, from AI laboratories to universities, and it announced both systems in a single year of four TPU introductions, a first for the company.

The announcement matters because of what each system is built to do. One is built around privacy while the work runs. The other is built around memory, the limit Google says now holds back large AI jobs.

TPU 8tx: privacy during the work

The TPU 8tx uses hardware level confidential computing to isolate customer data and proprietary model parameters even while a job is operating, according to the announcement as reported by Decrypt. It combines TPU 8t silicon with fifth generation AMD EPYC processors and Google's Boardfly network.

Google claims the design can cut latency by up to half compared with the TPU 8t, with almost no performance penalty. That comparison is Google's own, measured against its earlier hardware. The reports are clear on one limit of the announcement: Google's performance figures compare the new systems with earlier Google hardware, not with Nvidia's latest products.

TPU 8i Max: built for the memory wall

The TPU 8i Max combines two TPU 8i chips into one package. That doubles computing power, on-chip SRAM and high bandwidth memory throughput compared with the TPU 8i, according to the reports. TechTarget reporting put the memory figures at 768MB of SRAM and 17.2TB/s of high bandwidth memory throughput.

Amin Vahdat, Google's chief technologist and senior vice president of AI and infrastructure, framed the system in one line quoted in the reports: "Our fourth system, TPU 8i Max, is built for a different problem: the memory wall."

Google also said it is open-sourcing TPU Sync, software that moves AI working memory between chips, servers and storage. Putting that software in the open is part of the same pitch: customers who build on Google hardware get tools to move the working memory their models depend on.

What Google did not say

Google did not disclose pricing for the two new configurations, and it did not give availability dates. Customers weighing the systems against rival hardware get the architecture and Google's own performance claims on Monday, but not the two numbers that decide most purchases: what the systems cost and when they can be delivered.

The timing gave the launch a market backdrop. Nvidia shares fell 9.3 percent on Thursday and fell another 5 percent in Monday trading, in reporting that links the pressure on the stock to the Nvidia board investigation involving its chief executive. Google's announcement landed in that window, with its two new systems aimed at the same customers who buy AI computing power in large amounts.

Attribution matters here. The latency and memory numbers above are claims and reported specifications from Google's announcement, carried by Decrypt and TechTarget reporting. They are not independent test results, and Monday's announcement did not include a head to head comparison with Nvidia hardware.