字号·· | 护眼
彭博社

谷歌推出AI芯片专用版本,加剧与英伟达的竞争Google Escalates Nvidia Rivalry With Specialized Versions of AI Chips

点「原文对照」整页切到原文,或双击某段只看那段的原文。

今年早些时候,在加利福尼亚州山景城,成架的托盘正在测试谷歌的TPU。Racks of trays testing Google’s TPUs in Mountain View, California, earlier this year.

谷歌公司正在推出更广泛的人工智能芯片阵容,旨在利用其作为英伟达公司挑战者的快速崛起之势,在一个价值数千亿美元的市场中分一杯羹。

Inc.’s Google is rolling out a broader array of artificial intelligence chips, aiming to capitalize on its rapid ascent as a challenger to Nvidia Corp. in a market worth hundreds of billions of dollars.

该公司正在推出其张量处理单元(TPU)的两个新变体,这些定制半导体有助于开发和运行AI软件。最新芯片将满足机密计算的需求,并帮助模型在处理查询时访问更广泛的信息。更广泛地说,它们旨在服务于一个正转向使用AI智能体并面临安全挑战的行业。

The company is introducing two new variations of its tensor processing units, or TPUs, custom-made semiconductors that help develop and run AI software. The latest chips will address the need for confidential computing and help models access a wider range of information when handling queries. More broadly, they’re meant to serve an industry that’s shifting toward the use of AI agents and confronting security challenges.

此次推出标志着谷歌芯片部门产品开发的加速。该业务曾每两年发布一款新的AI处理器。今年,它同时推出了两款。而即将于周一宣布的最新变体将为这一对芯片增加更多规格选择。

The rollout marks an acceleration in product development for Google’s chip unit. The business once released a new AI processor every two years. This year, it put out two at the same time. And the latest variations, due to be announced Monday, will add additional flavors to that pair of chips.

谷歌芯片负责人阿明·瓦赫达特在采访中表示,这四款TPU产品将针对不同需求进行优化。“这确实遵循了我们在谷歌的理念,即未来将关乎专业化,而专业化的优势依然显著。”谷歌正处于AI芯片行业动荡之际确立其地位。对基础设施支出潜在放缓的担忧动摇了科技股,而民主党在美国中期选举中获胜引发了对数据中心项目的新疑虑。但对英伟达AI处理器的需求仍大于其供应能力,这为谷歌和其他竞争者留出了建立可观业务的空间。

The four TPU offerings will optimize different needs, Google chip chief Amin Vahdat said in an interview. “This really follows the thesis that we have here at Google, which is that the future is going to be about specialization, and the upsides for specialization continue to be significant.”Google is staking its claim at a tumultuous time for the AI chip industry. Concerns about a potential slowdown in infrastructure spending have roiled tech stocks, and US midterm elections that swept Democrats into power have raised fresh doubts about data center projects. But the demand for AI processors from Nvidia remains larger than it can satisfy, and that’s left room for Google and other contenders to build a sizable business.

瓦赫达特表示,新款TPU 8tx和8i Max变体并非全新的硅设计——而是经过重新配置,以承担客户一直在要求的任务。

The new TPU 8tx and 8i Max variations aren’t completely new silicon designs — rather, they’re reconfigured to take on tasks that customers have been requesting, Vahdat said.

8tx 旨在为政府机构和受监管行业的公司等客户处理安全、机密的工作。该系统与 Advanced Micro Devices Inc. 的芯片配合使用,这是两家公司合作的一部分。

The 8tx is aimed at handling secure, confidential work for customers like government agencies and companies in regulated industries. The system works with chips from Advanced Micro Devices Inc., part of a collaboration between the two companies.

谷歌芯片负责人阿明·瓦赫达特正在发布公司张量处理单元的新版本。Amin Vahdat, Google’s chip chief, is delivering new versions of the company’s tensor processing units.

与此同时,8i Max 旨在运行需要长上下文窗口的 AI 应用——即 AI 系统在单次请求中可考虑的数据量。Vahdat 表示,这对于药物发现和数学等任务是必要的,对编码应用也很重要。

The 8i Max, meanwhile, is meant to run AI applications that require long-context windows — the amount of data an AI system can consider in a single request. That’s necessary for things like drug discovery and math but is also important for coding applications, Vahdat said.

TPU 在由 Nvidia 的图形处理单元(GPU)主导的市场中竞争。十多年前,Google 着手开发这项技术,并与 Broadcom Inc. 合作,后者协助设计和协调生产。最初的想法是更高效地应对这家搜索巨头日益增长的计算需求。随后,Google 意识到 TPU 在 AI 工作方面的潜力,并已开始在其云计算业务之外销售这些芯片。

TPUs compete in a market dominated by Nvidia’s graphics processing units, or GPUs. Google set out to develop the technology more than a decade ago, alongside Broadcom Inc., which helped with design and coordinating production. The initial idea was to more efficiently handle the search giant’s burgeoning computing needs. It then realized the potential of TPUs for doing AI work and has now begun selling them outside of its cloud-computing operation.

Google 作为芯片销售商的新角色正证明是有利可图的。据 Bloomberg Intelligence 预测,到 2028 年,包括相关软件和硬件在内的 TPU 销售额将超过 1350 亿美元。这比今年估计的 120 亿美元有所上升,且不包括客户通过 Google Cloud 使用芯片产生的销售额。

Google’s new life as a chip seller is proving lucrative. By 2028, TPU sales will top $135 billion, including related software and hardware, according to Bloomberg Intelligence. That’s up from an estimated $12 billion this year, and it doesn’t include sales from clients using the chips through Google Cloud.

TPU 已赢得一些知名客户,包括 Apple Inc.、Meta Platforms Inc. 和 Oracle Corp.。Vahdat 表示,提供芯片的专用版本应有助于 Google 取得进一步进展。他说,这些变体将于明年推出。

TPUs have netted some high-profile clients, including Apple Inc., Meta Platforms Inc. and Oracle Corp. And offering specialized versions of the chips should help Google make further progress, Vahdat said. The variations will be available next year, he said.

“我们可以制造一款试图覆盖整个市场的单一芯片,或者通过制造芯片的专用版本,我们实际上能提供更大的价值,”他说。

“We could build a single chip that tries to address the whole market, or by building specialized versions of the chips, we actually can deliver much more value,” he said.

即使是英伟达,这家历史上一直将其 GPU 推广为适用于所有 AI 任务产品的公司,也开始提供更专业化的芯片系统。这包括提供其自家的中央处理器(CPU),这一品类由英特尔公司开创。这些处理器不负责创建人工智能模型所需的繁重数据运算,但它们对于管理 AI 智能体很有用。

Even Nvidia, which has historically promoted its GPUs as a product for all AI tasks, is starting to offer more specialized chip systems. That includes offering its own central processing units, or CPUs, a category pioneered by Intel Corp. Those processors don’t handle the heavy-duty data crunching needed to create artificial intelligence models, but they’re useful for managing AI agents.

英伟达还提供更专注于 AI 推理阶段的芯片——即已训练好的 AI 模型执行任务的阶段——利用的是从初创公司 Groq 收购的技术。

Nvidia also is offering chips more focused on the inference stage of AI — where already-trained AI models carry out tasks — using technology it acquired from startup Groq.

谷歌的 8tx 搭配来自 AMD 的 CPU,后者是英伟达和英特尔的竞争对手。这项技术允许在芯片层面对数据进行加密,使其更安全。Vahdat 表示,通常情况下,信息必须在内存中解锁,计算机才能对其进行处理。8i Max 将两颗 TPU 芯片封装在一个封装内,以提供更多算力和内存。

Google’s 8tx pairs with a CPU from AMD, a rival to both Nvidia and Intel. The technology allows data to be encrypted at the chip level, making it more secure. Normally, the information has to be unlocked in memory while a computer processes it, Vahdat said. The 8i Max puts two TPU chips in a single package to deliver more computing power and memory.

彭博情报分析师 Mandeep Singh 表示,TPU 的部分吸引力在于其性价比。它们的成本比英伟达的 GPU 低约 30%,但在吞吐量——即单位时间内可处理的数据量——方面表现具有竞争力。

Part of the appeal of TPUs is their price for performance, said Mandeep Singh, an analyst at Bloomberg Intelligence. They cost about 30% less than Nvidia’s GPUs but do a competitive job in terms of throughput — the amount of data that can be processed over a given time.

这种方案吸引了各类客户,包括金融等领域的用户。谷歌在上个月的财报发布中宣布,Jump Trading 是 TPU 客户。

The approach appeals to a variety of clients, including users in finance and other fields. Google announced during its earnings report last month that Jump Trading is a TPU customer.

Jump Trading 首席技术官 Alex Davies 在采访中表示,该公司在与谷歌签署更大规模合同前(约一年前),已花费数年时间试用这些芯片。

That company has spent several years trying out the chips beforeg a larger contract with Google about a year ago, Jump Trading Chief Technology Officer Alex Davies said in an interview.

谷歌在其第八代TPU芯片的基础上推出新的变体。Google is building on its version 8 TPU chip with new variations.

这家交易公司拥有数千颗 TPU,用于训练 AI 模型,用途包括交易研究。Davies 表示,Jump 仍主要使用英伟达进行训练,但一部分“直言不讳的少数派”员工非常喜欢用 TPU 完成他们的任务。

The trading firm has thousands of TPUs, which it uses to train AI models for things like trading research. Jump still mainly uses Nvidia for training but a “vocal minority” of employees really like TPUs for their tasks, Davies said.

TPU 在某些特定任务上表现非常出色,他说。戴维斯表示,使用不同种类的芯片来应对挑战是有好处的。“实际上,让某些问题以不同方式解决是很有吸引力的,”他说。“你能学到东西。”他说,在 TPU 与英伟达芯片不相上下或更胜一筹的领域,TPU 还具有更高的能效优势,从而降低了成本。

TPUs work very well for a subset of tasks, he said. There’s a benefit to using different kinds of chips to address challenges, Davies said. “It’s actually attractive to have some problems being done differently,” he said. “You learn stuff.”For the areas where TPUs are as good as or better than Nvidia chips, they also offer the benefit of being more power-efficient, which lowers the cost, he said.

上个月,甲骨文宣布已签署协议,将从明年起首次让其云客户使用 TPU。此前,甲骨文云客户只能在英伟达和 AMD 的 AI 芯片之间选择。

Last month, Oracle said it had signed a deal to let its cloud customers use TPUs for the first time starting next year. Oracle cloud customers previously had a choice of AI chips from Nvidia and AMD.

“我们与谷歌合作,是因为随着 AI 工作负载变得更加繁重,客户需要更多选择,”负责监管甲骨文云基础设施的高级副总裁卡兰·巴塔在声明中说。“这项合作扩展了我们帮助客户在最适合其需求的基础设施上构建和运行 AI 的能力。”Claude 聊天机器人及其他服务的制造商 Anthropic PBC 也是 TPU 客户。根据近期的一份监管文件,该公司有权在 2028 年至 2031 年间购买高达 1500 亿美元的 TPU 芯片。彭博新闻社报道,OpenAI 也计划购买 TPU。

“We partnered with Google because customers need more choice as AI workloads become more demanding,” Karan Batta, a senior vice president overseeing Oracle’s cloud infrastructure, said in a statement. “This partnership expands our ability to help customers build and run AI on the infrastructure that best fits their needs.”Anthropic PBC, maker of the Claude chatbot and other services, is another TPU customer. It has an option to purchase as much as $150 billion worth of the chips between 2028 and 2031, according to a recent regulatory filing. And OpenAI is planning to purchase TPUs as well, Bloomberg News has reported.

谷歌自身作为 AI 软件开发者的地位,有助于巩固其芯片业务。TPU 开发人员依靠提前接触谷歌最新的 Gemini 模型,来改进能效等方面并捕捉漏洞。这在半导体设计中至关重要,因为缺陷可能耗资数十亿美元并抹去数月的工作成果。瓦赫达特说,AI 正在降低这些陷阱的发生可能性。

Google’s own status as an AI software developer has helped bolster the chip operation. TPU developers have relied on early access to Google’s latest Gemini models to improve things like power efficiency and catch bugs. That’s critical in semiconductor design, where flaws can cost billions of dollars and erase months of work. AI is making those pitfalls less likely, Vahdat said.

但谷歌也面临其他挑战,包括新进入者。一家名为 Tenspresso 的初创公司由两名前 TPU 工程师共同创立,其首款芯片专注于长上下文窗口推理 —— 这也是 8i Max 瞄准的领域。但彭博社上周报道,该产品距离设计完成至少还有一年时间。

But Google faces other challenges, including newcomers to the market. One startup co-founded by two former TPU engineers, called Tenspresso, is focusing its first chip on long-context window inference — the same area targeted by the 8i Max. But that product is at least a year away from design completion, Bloomberg reported last week.

OpenAI 和 Anthropic 也在开展内部芯片项目——团队由 TPU 资深人士组成——这为未来的竞争提供了另一来源。

OpenAI and Anthropic also are working on in-house chip efforts — staffed with TPU veterans — providing another source of future competition.

此外,AI 基础设施支出放缓也存在风险。彭博社的 Singh 称,对数据中心的反弹可能导致技术公司原本预计在 2027 年支出的 1.4 万亿美元缩减约 5% 至 7%。其中包括芯片、网络设备、土地以及数据中心建设的其他材料支出。

There’s also the risk of a slowdown in AI infrastructure spending. The backlash against data centers threatens to shrink the $1.4 trillion that technology companies were expected to spend by roughly 5% to 7% in 2027, according to Bloomberg’s Singh. That includes spending on chips, networking equipment, land and other materials for data center construction.

彭博社的 Singh 还指出,Google 的 TPU 利润率不如 Nvidia 那样稳健。

And Google’s TPU profit margins are less robust than what Nvidia generates, according to Singh.

与芯片和 AI 基础设施领域的其他公司一样,Google 也面临内存和制造产能等关键部件的短缺。这使得满足需求变得困难。Google 在财报中表示,所有可用的 TPU 已全部签约给客户。

Like the rest of the chip and AI infrastructure space, Google is also facing shortages of key components like memory and manufacturing capacity. That’s made it hard to meet demand. Every TPU that Google has available is already contracted to a customer, the company said in its earnings report.

Vahdat 的团队正尝试在可能的情况下多元化供应来源。这甚至意味着要超越为 TPU 安排制造的 Broadcom。最近几个月,Google 与 Broadcom 竞争对手 Marvell Technology Inc. 签署了一项包括部分 TPU 工作的协议——但公告中对细节阐述不多。还有报道称 MediaTek Inc. 正在为 Google 的 TPU 业务提供服务。Vahdat 对两家公司在 TPU 业务中的角色拒绝置评。

Vahdat’s team is trying to diversify its supply sources where it can. That even means looking beyond Broadcom, which arranges manufacturing for TPUs. In recent months, Google signed a deal with Broadcom rival Marvell Technology Inc. that includes some TPU work — though the announcement was light on specifics. There have also been reports that MediaTek Inc. is doing work for Google’s TPU business. Vahdat declined to comment on either company’s role in the TPU business.

彭博情报的分析师 William Tong 表示,Google 很可能正在制定长期协议,“以确保供应并防止更多瓶颈”。

William Tong, an analyst with Bloomberg Intelligence, said it’s likely that Google is working out longer-term agreements “to get this supply set and prevent more bottlenecks.”

阅读媒体原文 ↗