Google 内部芯片 "Frozen v2" 将 Gemini 架构固化到硬件，效率提升 6-10 倍

谷歌正在内部开发一款新服务器芯片，名为“Frozen v2”，它将Gemini AI模型的架构直接嵌入硅片中。

根据《The Information》引用的消息来源，这款芯片在提供AI响应的效率上可能比谷歌当前的TPU芯片高6到10倍：https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently。谷歌计划从2028年开始部署这款芯片，并将Frozen v2视为针对专用芯片的测试试运行，其生产量将低于TPU系列。

与能够处理多种模型的谷歌TPU不同，Frozen v2将Gemini模型的一部分结构直接内置到硬件中。其命名逻辑与AI模型中“冻结”参数的概念相同，即锁定数值，使其不再变化。在Frozen v2中，模型的一部分被永久嵌入到芯片中，从而减少计算步骤并加快响应速度。

据报道，最初的想法来自谷歌DeepMind首席科学家Jeff Dean。他最初的Frozen设计方案是将模型权重直接嵌入芯片。权重是决定AI模型如何响应查询的具体设置。谷歌放弃了这种方法，因为该芯片只能用于单一Gemini版本，很快就会过时。

Frozen v2采取了更灵活的路径，通过嵌入模型架构而不是权重，也就是说嵌入的是底层蓝图而非调优参数。新的权重仍然可以加载到芯片上。据《The Information》报道，目前尚未决定实际将有多少架构会被硬编码。

由于该芯片只能在谷歌坚持相同模型架构的情况下工作，它可能不会成为面向外部客户的产品。谷歌已经将其 TPU 租赁给 Meta：https://the-decoder.com/meta-signs-multi-billion-dollar-deal-to-rent-googles-tpus-in-a-direct-challenge-to-nvidias-ai-chip-dominance/，提供给外部云客户：https://the-decoder.com/google-unveils-8th-gen-tpus-agent-platform-and-workspace-ai-layer-at-cloud-next-26/，并通过“TPU@Premises”计划将其定位为 Nvidia 的替代方案：https://the-decoder.com/the-mere-existence-of-google-tpus-reportedly-saved-openai-30-on-nvidia-chips/，其内部目标是争取 Nvidia 年收入的百分之十：https://the-decoder.com/google-cloud-aims-to-capture-ten-percent-of-nvidias-annual-revenue-with-tpus/。相比之下，Frozen v2 的目的是缓解谷歌内部在 AI 计算能力上的紧张：https://www.theinformation.com/articles/inside-balancing-act-googles-compute-crunch。广告

如果该芯片兑现其承诺，它仍可能成为重大的竞争优势。在 AI 业务中，公司对推理成本的优化程度正日益决定其利润率。谷歌可以使用 Frozen v2 以更低的价格运行强大的模型，并从 OpenAI 和 Anthropic 手中争取市场份额。

保持对 AI 的关注。清晰、有用，无废话。

关注 The Decoder 获取 AI 新闻、背景故事和专家分析。

The Decoder：https://the-decoder.com/

Google is building a new server chip internally called "Frozen v2" that embeds the Gemini AI model's architecture directly into silicon.

The chip could be 6 to 10 times more efficient at serving AI responses than Google's current TPU chips, according to sources cited by The Information：https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently. Google plans to deploy it starting in 2028 and sees Frozen v2 as a test run for specialized chips, with a smaller production volume than its TPU line.

Unlike Google's TPUs, which work with many models, Frozen v2 has parts of Gemini's model structure built right into the hardware. The name follows the same logic as "freezing" parameters in AI models, where you lock values so they stop changing. With Frozen v2, a portion of the model gets permanently frozen into the chip itself, which cuts down on compute steps and speeds up responses. Ad

The original idea reportedly came from Jeff Dean, Google Deepmind's chief scientist. His first Frozen design called for embedding the model weights directly into the chip. Weights are the specific settings that determine how an AI model responds to queries. Google scrapped that approach because the chip would have only worked with a single Gemini version and would have become outdated too quickly. Ad DEC_D_Incontent-1

Frozen v2 takes a more flexible path by embedding the model architecture instead of weights, meaning the underlying blueprint rather than the tuned parameters. New weights can still be loaded onto the chip. How much of the architecture will actually be hardcoded hasn't been decided yet, according to The Information.

Because the chip only works as long as Google sticks with the same model architecture, it probably won't become a product for outside customers. Google already leases its TPUs to Meta：https://the-decoder.com/meta-signs-multi-billion-dollar-deal-to-rent-googles-tpus-in-a-direct-challenge-to-nvidias-ai-chip-dominance/, offers them to external cloud customers：https://the-decoder.com/google-unveils-8th-gen-tpus-agent-platform-and-workspace-ai-layer-at-cloud-next-26/, and positions them through its "TPU@Premises" program as an alternative to Nvidia：https://the-decoder.com/the-mere-existence-of-google-tpus-reportedly-saved-openai-30-on-nvidia-chips/ with an internal goal of capturing ten percent of Nvidia's annual revenue：https://the-decoder.com/google-cloud-aims-to-capture-ten-percent-of-nvidias-annual-revenue-with-tpus/. Frozen v2, by contrast, is meant to ease Google's internal crunch on AI compute capacity：https://www.theinformation.com/articles/inside-balancing-act-googles-compute-crunch. Ad

If the chip delivers on its promise, it could still become a major competitive edge. In the AI business, how well companies optimize inference costs increasingly determines their margins. Google could use Frozen v2 to run powerful models at lower prices and take market share from OpenAI and Anthropic.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/