跳到正文
原文
TechCrunch · AI· Ram Iyer·· 6 天前AI 评分50

Google 发布 TPU 8tx 与 TPU 8i Max 两款专用 AI 芯片

Google broadens its AI hardware portfolio with two new specialized chips

文章要点
  • 1Google 发布 TPU 8tx 与 TPU 8i Max 两款第八代 TPU 新变体,扩展 AI 芯片适用场景
  • 2TPU 8tx 与 AMD 合作设计,主打机密训练与推理,安全保护直接写入硬件层
  • 3TPU 8i Max 将两颗 TPU 8i 合并为单一封装,内存容量、算力和带宽翻倍
  • 4Microsoft、Amazon、OpenAI、Anthropic 均在自研 AI 芯片,以减少对 Nvidia 的依赖
决策要点

Google 把自研 TPU 从通用推理扩展到机密推理与大内存场景,是云厂商摆脱对外部芯片依赖路线的延续。

  • 影响:需要处理敏感数据的企业多了一条硬件级隔离的训练与推理路径;单封装大内存可降低大模型查询部署的芯片数量与复杂度;超大规模厂商自研芯片会改变云算力的供给与议价格局。
  • 参考:可考虑让安全合规团队评估 TPU 8tx 的机密推理能力是否匹配自身数据边界要求;有大模型推理负载的企业值得先对比单封装大内存方案与多芯片方案的吞吐和成本。
  • 依据:两款芯片的能力与定位来自 Google 官方博客,属官方口径,暂无独立评测;行业自研芯片趋势为 TechCrunch 对行业的归纳。
正文

Google is notching up its chipmaking game as the company seeks to reduce its dependence on chips made by other companies for its cloud computing business. Google on Monday unveiled two new variants of its tensor processing units (TPUs), in-house developed chips meant for AI workloads, that broaden its chip portfolio to additional use cases: confidential inference and more memory capacity for large workloads.

The new units, TPU 8tx and TPU 8i Max, are part of the company’s eighth generation of chips introduced in April this year, and build on what those chips offered, which was more inference capacity and support for large training requirements.

TPU 8tx, designed with AMD, is based on the TPU 8t chip, and is geared specifically for confidentially training and running AI models, Google said in a blog post. The chip “builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use,” the company wrote.

The chip also has a custom inferencing network that lets it run reasoning alongside training, so it can be used to distill and train small models on the fly. That ability, according to Google, could prove useful for cases where a company might want to train a model on its proprietary data without having to move the data between devices.

Meanwhile, TPU 8i Max basically combines two of the inference-focused TPU 8i chips the company launched in April, doubling the memory capacity, processing power and bandwidth, but in a single package. The result is a chip that lets AI models handle a larger range of information during queries.

The chip launch comes as hyperscalers like Microsoft, Amazon, and even AI labs like OpenAI and Anthropic are working on their own lines of AI-specific semiconductors and accompanying software in order to reduce their reliance on the dominance of Nvidia.

The news also comes at a particularly vulnerable moment for Nvidia, whose CEO Jensen Huang over the weekend admitted to having an affair with someone alleged to be a Chinese spy.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Ram is a financial and tech reporter and editor. He covered North American and European M&A, equity, regulatory news and debt markets at Reuters and Acuris Global, and has also written about travel, tourism, entertainment and books.

You can contact or verify outreach from Ram by emailing [email protected].

View Bio

来源:TechCrunch · AI · techcrunch.com