
新系統針對機密推論與龐大脈絡視窗運算需求設計,擴大AI基礎設施布局
Google發表TPU 8tx與TPU 8i Max兩款新系統,加上4月推出的TPU 8t與8i,首度在同年推出4款第八代TPU。TPU 8tx針對機密推論與安全運算優化,TPU 8i Max提供雙倍運算與記憶體頻寬,客戶數接近100家。
AI-generated summary
Google最初十多年前為內部語音辨識與語言翻譯需求開發TPU,如今發展至第八代,應用範圍從Gemini模型訓練與推論擴展至AI代理工作負載與Google Cloud客戶雲端運算需求,近期更開放客戶將TPU部署於自有資料中心。
Google宣布擴充第八代張量處理器(TPU)產品陣容,推出兩款全新系統,包括專為安全與機密推論設計的「TPU 8tx」,以及針對龐大脈絡視窗與高密度運算打造的「TPU 8i Max」,加上今年4月推出的TPU 8t與8i,Google首度在同一年內推出4款TPU系統,進一步擴大AI基礎設施布局。
Google表示,TPU最初是十多年前為滿足內部語音辨識、語言翻譯等需求而開發,如今已發展至第八代,應用範圍也從Google自家Gemini模型的訓練與推論,逐步擴展至AI代理工作負載、Google Cloud客戶的雲端運算需求,近期更開放客戶將TPU直接部署於自有資料中心。目前TPU已成為Google快速成長的業務之一,客戶數接近100家,涵蓋前沿AI研究機構、交易公司、零售商及大學等。
今年4月舉辦的Google Cloud Next大會,Google已率先發表兩款第八代TPU系統,其中TPU 8t主要負責大規模AI模型訓練,TPU 8i則鎖定即時推論需求。Google指出,相較前一代TPU,TPU 8t的訓練效能最高提升2.7倍,TPU 8i的推論性價比則提高80%。
Google此次進一步擴大第八代TPU產品陣容,推出兩款全新系統,包括專為處理龐大脈絡視窗及提升單機櫃運算密度設計的TPU 8i Max,以及強化安全與機密推論能力的TPU 8tx。
Google說明,過去TPU主要與Google DeepMind共同設計,但隨著外部企業的應用需求日益多元,Google也開始更頻繁地將客戶意見納入產品開發。尤其當AI運算從人類操作規模逐步邁向大量AI代理協同運作,企業對於能夠處理龐大脈絡視窗的AI加速器需求持續增加,部分受到高度監管的產業,也開始針對特定工作負載提出專屬加密功能的需求。
為因應這些新興需求,Google此次推出兩款全新TPU系統,並搭配近期發布的PyTorch相關更新,進一步降低開發者及Google Cloud客戶使用TPU的門檻。透過結合模組化的自研晶片,以及開放、易於使用的軟體架構,Google希望能更快速地因應AI產業發展,推出符合不同運算需求的硬體與系統配置。
TPU 8tx是第八代的第三款系統,針對安全與機密推論進行了最佳化,同時也非常適合小規模訓練,是專為受高度監管的企業與政府機構所打造,讓他們能在處理高機敏資料的同時,安全地執行前沿或專屬模型及自主AI代理。
Google指出,一般資料在儲存或透過網路傳輸時會進行加密,但在電腦處理資料的過程中,必須先在記憶體中解密。TPU 8tx採用與TPU 8t相同的矽晶片,並搭配第五代AMD EPYC CPU主機,將安全機制直接內建於硬體層中,能確保敏感資料與模型權重即使在活躍使用狀態下,也能全程受到保護並維持隔離。
此外,藉由採用專為即時推論所打造的相同低延遲Boardfly網路,TPU 8tx 能讓多步驟AI代理保持順暢運作,在幾乎零效能損耗的情況下,提供比TPU 8t低高達2倍的延遲表現。
將訓練等級的運算能力與高速推論網路結合,意味著企業可以在單一系統上同時執行並客製化自己的模型,TPU 8tx讓團隊可以使用專屬資料微調模型、將大型模型蒸餾為較小模型,並透過強化學習訓練他們的AI代理,所有過程皆無需在不同叢集間搬移敏感資料。
另一款新系統TPU 8i Max,則是鎖定解決AI代理協同運作時面臨的記憶體瓶頸。當數十個AI代理同時針對整個軟體程式碼庫或長達數小時的影片協同運作時,記憶體頻寬很快就會成為瓶頸,為此,Google將兩顆低延遲的TPU 8i晶片整合封裝在單一模組上。
TPU 8i Max 同步擴展了運算與記憶體資源,提供雙倍的運算能力、雙倍的靜態隨機存取記憶體(SRAM),以及雙倍於TPU 8i的高頻寬記憶體(HBM)頻寬,確保處理器隨時備妥關鍵資料,避免大型工作負載發生停滯。此外,由於TPU 8i Max在現有的機櫃空間容納了兩倍的矽晶片,資料中心每平方英尺與每瓦電力能部署的運算能力也因此大幅提升。
Google表示,第八代TPU的4款系統各自針對不同運算需求進行設計,涵蓋AI代理應用生命週期的各個階段。其中,TPU 8t主要負責大規模模型訓練,TPU 8i鎖定低延遲推論,TPU 8tx著重機密運算,而TPU 8i Max則專為處理龐大脈絡視窗及提升單機櫃運算密度打造。
Google特別強調,此次推出的是兩款全新TPU「系統」,而非重新設計的「晶片」,兩者在開發方式與所需時間上有所不同。一般而言,新晶片的設計涉及數十億個微型電晶體的布局,製造流程複雜,往往需要耗費多年時間;相較之下,新系統則是以既有晶片為基礎,整合運算、網路及儲存資源,並調整相關封裝與配置,以更快速地回應客戶需求及新興研究應用。
Google指出,早在十多年前開發第一代TPU時,就已預見AI軟體的演進速度將超越晶片製造週期,因此採取模組化設計,將TPU作為AI超級電腦(AI Hypercomputer)的基礎運算模組,讓工程師能依據不同需求,靈活調整晶片封裝、高速記憶體、光學網路及安全防護機制。透過這種設計方式,Google得以將專用超級電腦的交付時間,從原本可能需要數年縮短至數個月。
AI outlook — possibilities, not facts
Google將繼續擴展TPU產品陣容,針對不同AI工作負載推出更多系統配置
Likely · Within months

Huida CEO Huang Renxun admitted that he had an extramarital affair with Huang Le, a suspected Chinese female spy, triggering national security concerns. Ma Yixiang, secretary-general of Taiwan Foundation, said that scientific and technological leaders have become national-level intelligence targets, calling for the establishment of a counter-intelligence protection system for key scientific and technological personnel and the establishment of a joint early warning mechanism with the United States.

Nvidia CEO Jensen Huang is under internal company investigation for allegedly having an extramarital affair with Huang Le, a woman accused of being a Chinese spy. Huang Le was the former CEO of a Singapore data center company, and his company was accused of possibly smuggling Nvidia chips into China. This matter involves US-China AI competition and national security. Nvidia's stock price fell but did not plummet significantly. Huang Renxun still received support from the board of directors and Trump.

At the first World Robot Taishan Competition in Tai'an, Shandong Province, the same Yushu G1 development version robot used in the Spring Festival Gala withdrew from the competition due to overheating, which aroused public attention on the actual working capabilities of humanoid robots. Experts pointed out that the "heatstroke" of robots is mainly caused by insufficient heat dissipation rather than high external temperatures; there is a gap between stage performances and continuous operations under real loads. The industry believes that the biggest bottleneck of humanoid robots is not the "brain", but the ability to continue to operate autonomously. They may become reliable helpers in specific positions in the next two or three years. It will still take longer for general-purpose humanoid robots to independently undertake housework or high-risk rescue.
Zhu Feng, a professor at the School of International Relations at Nanjing University, said at the "International Forum on Mutual Learning of Civilizations 2026" sub-forum that the uncertainty and technical challenges brought by the artificial intelligence revolution far exceed those of the previous three industrial revolutions. Each country's regulatory rules are different and governance mechanisms lag behind. The international community should jointly manage and control AI risks based on multilateralism to ensure that technology serves the common welfare of all mankind.

As a computing power service, AI tokens were launched on e-commerce platforms such as Tmall and JD.com during Double 11. Large model manufacturers such as Zhipu, MiniMax, Alibaba Cloud, and Qianwen opened stores. Consumers can purchase token packages like recharging phone bills. The transaction volume of Taobao's Ci Yuan category increased by 450% month-on-month on the first day of sale, and the sales volume of JD.com's Token package products increased 20 times month-on-month within 2 hours. Experts believe that Ciyuan's move towards the C-side marks the beginning of C-side demand, but issues such as measurement and accounting, pricing mechanisms, security control and compliance governance still need to be resolved, and the payment conversion rate for ordinary users is still low.

Huida CEO Jen-Hsun Huang's extramarital affair with suspected Chinese spy Huang Le triggered an investigation. Huida's board of directors has launched an independent investigation, and the U.S. Congress has demanded clarification. Zhang Kunyang, a researcher at the Jamestown Foundation, analyzed that the intelligence risk lies not only in the leakage of secrets, but also in affecting Huang Renxun's judgment on Taiwan's security and export controls to China. He also revealed that many Chinese universities and military-related institutions are seeking to purchase Huida H200 and other chips.