天龙八部

有了TPU还不够，谷歌开始为Gemini 独造芯片：推理能效或达6到10倍_我的网站

一 |     

    Foreign ministers and top diplomats of the Pacific Islands countries attend a meeting of the Pacific Islands Forum (PIF) in Suva, Fiji, on August 7, 2026, ahead of the full leaders meeting scheduled for September 2026 in Palau. Photo: VCG
As foreign ministers and senior representatives from Pacific nations convened in Suva for the Pacific Islands Forum (PIF) to discuss key regional issues, New Zealand's foreign minister reportedly pushed to secure a joint statement condemning China's missile tests in July, however, his effort ultimately fell flat. A Chinese expert described the outcome as a clear rebuke to Canberra and Wellington's ingrained habit of speaking for the Pacific, saying the two countries must come to terms with a shifting reality: Pacific Island states are increasingly unwilling to let Australia and New Zealand dictate who their friends, foes and perceived security threats should be.
Pacific Island foreign ministers failed to agree on Friday to a joint response to China's missile launch in July, with further discussions expected at the Pacific leaders meeting at the end of August, said New Zealand Foreign Minister Winston Peters, according to Reuters. 
According to a report from Radio New Zealand, there has not been a unanimous statement from Pacific nations about China's test-launch in July, and Peters said New Zealand had been pushing for one. Peters claimed that the Pacific missed an opportunity not to issue a joint statement soon after China's launch.
Australia's Foreign Minister Penny Wong said before the meeting that she hoped China took note of the "very clear response" already sent from the Pacific about the test, according to AFP.
Peters' push for a joint PIF statement on China's missile tests, together with Wong's remarks, are less about technical security concerns than an effort to turn their own anxieties about China into a regional agenda and pressure Beijing under the forum's collective banner. His description of Pacific Island countries' refusal to single out China as a "missed opportunity" also reveals little respect for their independent judgment, Chen Hong, director of the New Zealand Studies Centre, East China Normal University, told the Global Times.
The result is a clear rebuke to Canberra and Wellington's long-standing habit of arrogantly "speaking for" the Pacific, as island states are increasingly unwilling to let Australia and New Zealand decide who their friends, foes or security threats should be, said Chen Hong. 
The Chinese People's Liberation Army Navy said on July 6 that one of its strategic nuclear submarines successfully launched a strategic missile toward relevant high seas of the Pacific Ocean, which landed precisely within the designated waters. China notified relevant countries in advance of the test launch, a move that demonstrates the openness and transparency of the Chinese armed forces, Chen Xi, a spokesperson for the Ministry of National Defense, said at a press conference on July 7. 
Chen Hong also noted that Peters' recent China-related rhetoric has become increasingly populist and hawkish as elections approach in New Zealand. "Such an approach risks undermining New Zealand's long-standing reputation for pragmatic, independent, and rational diplomacy," Chen Hong said. 
In July, during a parliamentary debate on New Zealand's Covid-19 pandemic response, Peters, who is also leader of the populist New Zealand First party, told a Chinese-born lawmaker to "go back to your own country" while mocking his English proficiency. His remarks drew wide criticism in New Zealand.
Yet Australia and New Zealand continued to amplify missile-related issues and great-power rivalry, pushing the "China threat" narrative to the center of the regional agenda while sidelining the priorities that matter most to Pacific Island nations, said Chinese analysts. 
Chen Hong noted that for island states, the most urgent concerns are not geopolitical competition but everyday challenges: climate change, energy security, development financing, infrastructure and economic resilience. By shifting the focus from these concerns to strategic rivalry among major powers, Australia and New Zealand risk distorting the region's real priorities and undermining the aspirations of Pacific Island countries.
According to an introduction on the forum website, "The Forum's Pacific Vision is for a resilient region of peace, harmony, security, social inclusion and prosperity, so that all Pacific people can lead free, healthy, and productive lives."
At the opening of the meeting, Rick Houenipwela, foreign minister of the Solomon Islands and chair of the meeting, said that the Pacific's agenda must continue to be "conceived ⁠by the Pacific, shaped by the Pacific and only by the Pacific," according to Reuters. 
Chen Hong said Pacific Island nations have made their position clear: differences with China do not mean signing up to the strategic agendas of Australia and New Zealand. A growing number of island countries refuse to be trapped in the black-and-white binary choice of siding with either China or the West.
If Canberra and Wellington keep pushing rhetoric beyond the regional consensus, they risk undermining the very institution they claim to support. The PIF's strength lies not in issuing confrontational statements, but in preserving Pacific unity. Turning that consensus into a tool for pressuring China would erode the forum's cohesion and weaken its central role in the region, said Chen Hong.
。    谷歌正在尝试把 Gemini 与芯片更紧密地结合起来。         据《The Information》援引知情人士消息，谷歌母公司 Alphabet 正在研发一款服务器 AI 芯片，内部非正式代号为“Frozen v2”。这款芯片针对 Gemini 模型设计，计划最早于 2028 年部署。谷歌工程师预计，以单位功耗所能处理的 Token 数量计算，其效率可能达到公司最新一代自研 AI 芯片的 6—10 倍。         谷歌目前尚未正式确认 Frozen v2 项目，也没有公布芯片规格。公司在回应媒体时表示，其团队一直在研究和试验提升性能与效率的新方案，但并非所有项目最终都会进入生产阶段。谷歌同时强调，从底层共同设计硬件和软件，是其打造高度集成系统的重要方式。

二 | 但报道发布后，Alphabet 股价在当地时间 7 月 20 日早盘上涨约 3.3%。         与现有 TPU 相比，Frozen v2 的主要区别，是它进一步收窄了芯片的适用范围。它的定位是一颗更极致的“专用芯片”。

三 | 根据现有报道，谷歌计划将 Gemini 的部分架构特征固化在芯片中，针对其计算路径和常用操作进行专门优化。         大模型推理不仅需要完成大量矩阵运算，还要频繁从高带宽内存读取参数。数据在内存与计算单元之间移动，会消耗时间和电力。Frozen v2 希望减少这部分开销，从而提高响应速度，并在相同功耗下处理更多请求。报道所称的 6—10 倍效率提升，具体指每单位功耗能够处理的 Token 数量，并不是芯片峰值算力。

四 |          谷歌现有 TPU 同样是面向机器学习的专用处理器，但需要适配不同模型，并承担训练、推理等多种任务。最新一代 Ironwood 单颗芯片配备 192GB 高带宽内存，单位功耗性能约为上一代 Trillium 的两倍，最多可由 9216 颗芯片组成 Superpod。Frozen v2 不会取代 TPU，而是作为一条更具针对性的产品线，与其并行使用。

五 |          (来源：Google)          但专用芯片也有其劣势，更深的模型与芯片绑定会降低灵活性。芯片从设计到投产通常需要数年，如果此后 Gemini 的基础架构发生较大变化，已经固化的设计可能难以适配。Frozen v2 能否获得长期收益，既取决于实际能效，也取决于谷歌能否在快速迭代模型的同时，维持底层架构的相对稳定。

六 |          那么，谷歌为什么要“铤而走险”？首先一个原因，是为了应对持续增长的算力需求。

七 | 据相关数据，谷歌内部的 AI 计算资源短缺已经引发不同团队之间的资源竞争，Google Cloud 甚至因此拒绝过部分外部客户订单。         大模型的计算需求包括训练与推理两个阶段。训练先进模型需要集中投入大量算力，但随着 Gemini 被加入搜索、办公软件、云服务和智能终端，长期发生的推理可能形成更庞大的成本。单次生成的费用看似有限，乘以数以亿计的用户和持续增长的调用量后，电力、芯片、网络及数据中心投入都会成为重要支出。         Alphabet 已将 2026 年资本开支预期上调至 1,800 亿—1,900 亿美元，主要用于扩建 AI 计算能力、数据中心和相关基础设施。

八 | 在这一背景下，提高每瓦电力能够生成的 Token 数，不仅意味着更低的能源消耗，还意味着同一座数据中心能够承载更多用户请求。

九 | 对于已经拥有大量 Gemini 使用场景的谷歌而言，即使单次推理成本只下降一小部分，规模化之后也可能带来可观节省。         图｜谷歌历年资本开支预期（来源：S&P）          自研芯片还可以加强谷歌对供应链和产品节奏的控制，减少对英伟达等外部芯片供应商的依赖。

十 | 但这种“减少依赖”并不等于完全替代。

十一 | Frozen v2 即使顺利投产，预计规模也将低于 TPU；谷歌的数据中心仍会同时使用 TPU、GPU 及其他计算设备。它更像是为高频且相对稳定的 Gemini 任务增加一类成本更低的专用工具。         谷歌也不是唯一走向自研芯片的 AI 公司。2026 年 6 月，OpenAI 与博通公布首款自研推理芯片 Jalapeño。OpenAI 表示，这款芯片围绕大语言模型推理设计，将结合博通的芯片实现和网络技术以及 Celestica 的服务器系统能力，计划于 2026 年底开始部署。         图｜博通公司首席执行官 Hock Tan 将一片芯片晶圆递交给 OpenAI 公司首席执行官 Sam Altman。（来源：OpenAI）          Anthropic 也被曝已经开始研究自有 AI 芯片，并与三星电子等潜在合作方接触。

十二 | 不过该项目仍处于早期阶段，芯片的用途、性能目标和服务器部署方式尚未明确 Anthropic 同时表示，亚马逊 Trainium、谷歌 TPU 和英伟达 GPU 仍将是其扩张算力的重要基础。中国方面，此前 DeepSeek 和近期的月之暗面也有自研芯片方面的传闻。         除专注于基础模型的公司外，亚马逊、微软、Meta、阿里巴巴、百度和字节跳动等拥有云服务或大规模互联网业务的企业，也在不同程度上推进专用 AI 芯片。其中部分芯片服务于模型训练，部分集中于推理、推荐和广告等高频工作负载。它们共同反映出一个趋势：随着 AI 服务规模扩大，掌握模型的公司越来越希望同时参与芯片和计算系统设计。

十三 |          这些项目属于同一轮趋势：AI 公司希望根据自身模型和业务负载设计芯片，以获得更好的成本、能效和供应保障。但它们并不是完全相同的技术路线。Trainium、Maia、MTIA 以及 OpenAI 的 Jalapeño 主要面向一类或多类 AI 工作负载，仍保留一定通用性；Frozen v2 则称会把 Gemini 的部分架构直接固化进硬件，模型与芯片的绑定程度可能更高。         因此，Frozen v2 所代表的不只是谷歌扩大自研芯片投入，也是一项更激进的软硬件协同试验。答案将取决于芯片能否如期投产，也取决于到 2028 年时，Gemini 是否仍然适合这块为它提前设计的硅片。         参考链接：          1.https://techcrunch.com/2026/07/20/google-is-working-on-a-new-ai-chip-designed-to-make-gemini-more-efficient/          2.https://www.reuters.com/business/google-plans-new-chip-run-gemini-models-more-efficiently-information-reports-2026-07-20/          3.https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/          4.https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-google-tpu-things-to-know/          运营/排版：何晨龙          注：封面由 AI 辅助生成。

Current article:http://cuozubeishenjingli.cfd/thft/20260826/275249.html

Published on:19:54:45