谷歌宣布扩展其第八代张量处理单元(TPU)产品组合,新增专为智能体时代设计的TPU 8tx和TPU 8i Max系统 1。这是谷歌首次在单一年份内发布四款TPU系统,另外两款为TPU 8t和8i 1。
在硬件层面,TPU 8tx侧重于硬件级安全与机密推理 1。该芯片基于TPU 8t硅片,搭配第5代AMD EPYC CPU,提供硬件级机密计算能力,其延迟比TPU 8t低高达2倍 1。TPU 8i Max旨在通过双芯片封装解决大规模上下文窗口下的内存瓶颈问题 1。它将两个低延迟TPU 8i芯片合并在一个封装中,使得计算能力、SRAM和HBM带宽均为TPU 8i的两倍 1。
此外,谷歌还发布了一系列软件生态更新,以降低开发门槛并提升AI基础设施的性能与灵活性 1。具体包括:
- TorchTPU开源:使PyTorch代码和Hugging Face模型可原生运行在TPU上 1。
- 开放TPU代理套件(预览版):支持Claude Code、Cursor等工具自动进行TPU配置和内核调优 1。
- Cloud Pathways全面可用:该服务现已在Google Kubernetes Engine (GKE)上全面可用,允许用户像操作本地机器一样编排大规模TPU系统 1。
- GKE动态切片功能全面可用:利用光路交换技术消除僵硬的节点池重建周期,作业启动时间缩短多达5倍 1。
- TPU Sync开源:作为模型状态层被开源,可将KV传输延迟和首Token生成时间(TTFT)降低90%以上 1。
Google has announced an expansion of its eighth-generation Tensor Processing Unit (TPU) product line, introducing two new systems designed for the agentic era: the TPU 8tx and the TPU 8i Max 1. This launch marks the first time Google has released four TPU systems within a single year, adding these variants to the existing TPU 8t and TPU 8i models 1. The new hardware targets distinct computational needs, with the TPU 8tx focusing on hardware-level security and confidential inference, while the TPU 8i Max addresses memory bottlenecks associated with large context windows through dual-chip packaging 1.
The TPU 8tx is built on the TPU 8t silicon but pairs it with fifth-generation AMD EPYC CPUs to provide hardware-level confidential computing capabilities, achieving latencies up to twice as low as the standard TPU 8t 1. In contrast, the TPU 8i Max integrates two low-latency TPU 8i chips into a single package, effectively doubling the compute capacity, static random-access memory (SRAM), and high-bandwidth memory (HBM) bandwidth compared to the original TPU 8i 1. These hardware advancements are supported by a suite of software ecosystem updates aimed at lowering development barriers and enhancing performance flexibility 1.
Key software innovations include the open-sourcing of TorchTPU, which enables PyTorch code and Hugging Face models to run natively on TPUs 1. Google also previewed the TPU Agent Suite, allowing tools such as Claude Code and Cursor to automatically handle TPU configuration and kernel tuning 1. Furthermore, Cloud Pathways is now generally available on Google Kubernetes Engine (GKE), permitting users to orchestrate large-scale TPU systems similarly to local machines 1. To optimize job efficiency, GKE’s dynamic slicing feature, which uses optical switching to eliminate rigid node pool rebuild cycles, is also generally available, reducing job startup times by up to five times 1. Additionally, the model state layer TPU Sync has been open-sourced, claiming to reduce KV transmission latency and time-to-first-token (TTFT) by more than 90% 1.
评论
还没有评论,欢迎留下第一条。