创建

登录 ReadmeX

或

了解我们如何处理你的信息: 隐私政策

ReadmeX

陪你读懂 AI 动态

今天,想先了解什么?

选一份动态总结,或直接聊聊你的问题。

社区和关注动态总结只属于你自己。

资讯

黄仁勋离任或为Google TPU带来机会

AI 总结

9to5Google认为,黄仁勋辞去Nvidia CEO职务后,客户可能更愿意寻找CUDA生态之外的替代方案,Google TPU将因此受益。报道提到Google正开放TPU第三方销售、计划自2027年向Oracle Cloud提供容量,并推出第八代TPU及TorchTPU工具,但TPU 8i Max的原始性能仍不及Nvidia的Vera Rubin。

深度见解

Google正扩大TPU的外部可用范围,并通过兼容PyTorch的工具降低迁移门槛,为AI基础设施采购者提供Nvidia之外的替代选择。

摘自汇总《Jensen’s exit could be Google’s TPU moment》

⭐️ Starred

In the immediate fallout of Jensen Huang’s humiliating resignation from Nvidia, coverage will focus on how the company plans to continue without its beloved former CEO, the world’s biggest champion of open-source AI and the proliferation of GPUs across borders. But in the long term, there’s an argument to be made that Google will be one of the largest winners of the shake-up, with a market opening up to competition outside of Nvidia’s CUDA moat.

While Colette Kress, the interim CEO, is unlikely to make any drastic changes to the company’s roadmap, the reality is that Nvidia’s competitors in the semiconductor space have desperately been working to unseat the market leader, and without Jensen at the helm, customers may be more likely to pursue alternatives.

For years, Google has offered its TPUs, or Tensor Processing Units, as a solution for AI training and inference, but they’ve been largely locked to its Google Cloud infrastructure. That’s changing now that the company is opening up sales to third parties and has announced it will bring TPU capacity to Oracle’s cloud beginning in 2027.

Google’s latest generation of TPUs is specialized; here’s how the company describes the lineup in its recent announcement post:

While flexible across workloads, together the four systems in our eighth-generation TPUs give customers a purpose-built accelerator for every stage of the agentic lifecycle. TPU 8t for scale-up training, TPU 8i for low-latency reasoning (inference), TPU 8tx for confidential computing, and TPU 8i Max for massive context windows and the highest compute density per rack.

Google isn’t comparing its new TPUs directly to Nvidia’s offerings, and clock-for-clock, even the most powerful TPU 8i Max doesn’t touch Vera Rubin for sheer performance. But this generation should be much more competitive for a few reasons: it’s more power efficient in certain configurations, and Google made it easier to translate coding workflows optimized for CUDA to work with its TPUs.

Back in October, Google open-sourced TorchTPU, which it co-developed with Meta’s PyTorch team, to let developers use familiar PyTorch workflows on TPUs with minimal code changes. That won’t make every custom CUDA operation portable or eliminate performance tuning, but reducing the effort required to try an alternative may change the buying decision. Customers don’t need to abandon Nvidia outright for Google to increase its TPU market share, or for Nvidia to face more pressure on price.

Alongside Broadcom and MediaTek, Google is taking on the Nvidia behemoth, and in the short term, it will almost certainly see a large upside in sentiment. But in the longer term, with scaling pressure lowering margins for Nvidia, Google’s specialized and more efficient TPUs may be a primary beneficiary of this sea change. They could also offset a slowdown in Search as Google’s shift to Cloud continues apace.


#Google#Nvidia#Oracle#Meta#Google TPU 8t#TPU 8i#TPU 8tx#TPU 8i Max#TorchTPU#PyTorch#CUDA#Vera Rubin

原文约 2 分钟读完阅读原文
群聊

还没有评论,来说说你的看法。