Google Releases Two New Variants: TPU 8tx and 8i Max

Article body EM_StockImg_Start

EM_StockImg_End

Google released two new variants of its eighth-generation TPU on Monday local time, targeting confidential computing and long-context AI applications respectively, further challenging Nvidia by segmenting its product line. Demand for TPUs is strong and supply has already been fully booked, but component shortages are constraining supply, and the risk of a slowdown in AI infrastructure spending has still not dissipated.

According to a media report on November 9 local time, Google released two new variants of its Tensor Processing Unit (TPU) that day: TPU 8tx and 8i Max. TPU is Google's self-developed series of AI-specific chips, used for training and inference of AI models, competing with Nvidia's GPUs.

Google's chip division used to launch a new generation of TPU roughly every two years. According to information on Google's official website, the company released two chips of the eighth-generation TPU at the same time at the Google Cloud Next conference on April 22 this year, namely TPU 8t, focused on model training, and TPU 8i, focused on inference.

The 8tx and 8i Max this time are derived from these two chips, expanding the eighth-generation TPU to four models, with a noticeably accelerated development pace.

Amin Vahdat, head of Google's chip business, said in an interview that these four products will each meet different needs. He pointed out:

"This is consistent with Google's long-standing judgment that the future belongs to specialization, and the benefits brought by specialization remain very significant."

"We could build one chip that tries to cover the entire market, but launching specialized versions can create much greater value."

Vahdat introduced that the new products are not entirely new chip designs, but products reconfigured according to customer needs. Among them, 8tx targets the confidential computing needs of government agencies and enterprises in regulated industries, working in coordination with AMD's central processing units (CPUs), and can encrypt data at the chip level, avoiding data being stored in memory in a decrypted state during processing.

8i Max integrates two TPU chips in the same package, providing stronger computing power and larger memory, mainly serving AI applications that require long context windows, such as drug research and development, mathematics and programming scenarios.

Vahdat disclosed that the two new products will go on sale next year.

Segmenting chip products by task is gradually becoming an industry trend. Even Nvidia, which has long positioned GPUs as general-purpose AI products, has begun to offer chip systems for specific tasks, including launching its own CPUs and, using technology obtained from AI chip startup Groq, launching chips focused on the inference stage.

Commercialization accelerates

Google began developing TPU more than a decade ago in cooperation with chip design company Broadcom, initially to meet its own growing computing power needs more efficiently, and has now begun selling them externally beyond its cloud business.

According to forecasts by industry research institutions, by 2028, TPU sales, including related software and hardware, will exceed $135 billion, with an estimated value of about $12 billion this year. This figure does not yet include revenue generated by customers using TPUs through Google Cloud.

Institutional analyst Mandeep Singh pointed out that the appeal of TPU lies in its cost-effectiveness: its price is about 30% lower than Nvidia's GPUs, while its throughput is quite competitive.

Currently, tech giants such as Apple, Meta and Oracle are all TPU customers.

Oracle announced last month that it had signed an agreement with Google, allowing its cloud customers to use TPUs for the first time starting next year. Before that, the AI chip options for Oracle cloud customers were limited to products from Nvidia and AMD.

AI startup Anthropic is also a TPU customer. According to a recent regulatory filing, the company has the right to purchase up to $150 billion of TPUs between 2028 and 2031. According to other media reports, ChatGPT developer OpenAI also plans to purchase TPUs.

Google disclosed in its financial report last month that quantitative trading firm Jump Trading is also a customer. The company's chief technology officer, Alex Davies, said the company has deployed thousands of TPUs to train AI models for trading research and other purposes, but training work still mainly relies on Nvidia chips; in areas where TPUs perform no worse than Nvidia chips, their higher energy efficiency also helps reduce costs.

Challenges remain

Along with surging demand, Google also faces competition from new entrants. According to reports, chip startup Tenspresso, co-founded by two former TPU engineers, also aims its first chip at long-context inference, but it will take at least another year to complete the design. OpenAI and Anthropic are also advancing self-developed chips, with teams that include many veteran TPU engineers.

Institutional analyst Singh expects that opposition to data centers may reduce tech companies' originally projected $1.4 trillion in spending for 2027 by about 5% to 7%. He also pointed out that Google's TPU profit margin is lower than Nvidia's.

Like the entire industry, Google also faces shortages of memory chips and manufacturing capacity, making it difficult to fully meet strong demand. Google said in its financial report that all of its available TPUs have been booked by customers. Vahdat's team is diversifying supply sources as much as possible.

Media reports pointed out that in recent months, Google has sought more partners beyond Broadcom, its main TPU partner, and signed an agreement with Broadcom competitor and chipmaker Marvell involving part of the TPU business, but disclosed limited details; other reports said chip design company MediaTek is also participating in Google's TPU business.

Research institution analyst William Tong believes that Google is very likely finalizing longer-term agreements to lock in supply and avoid more bottlenecks.

(Source: Cailian Press) Middle of the article Embedded advertisement