Modal 计划在 Nvidia 基于 Arm 架构的 Vera CPU 上运行其 Sandboxes 产品，这是其首次摆脱 x86 芯片的尝试。在 Modal 自行进行的测试中，当服务器装满容器时，Vera 能处理的任务数量是 Intel 服务器的两倍。这些结果尚未经过独立验证。
Modal plans to run its Sandboxes product on Nvidia’s Arm-based Vera CPU, its first step away from x86 chips. In Modal’s own tests, Vera handled up to twice as many tasks as an Intel server when packed with containers. The results have not been independently checked.
Modal 计划在 Nvidia 新推出的 Vera CPU 上运行其 AI 代理沙盒。在其自行进行的测试中，该芯片处理的任务数量是 Intel 服务器的两倍。这家 AI 云公司于周一在一篇博客文章中宣布了这一计划。
Modal plans to run its AI agent sandboxes on Nvidia’s new Vera CPU. In its own tests, the chip handled up to twice as many tasks as an Intel server. The AI cloud company announced the plan in a blog post on Monday.
这将是 Modal 首次摆脱 x86 芯片——即 Intel 和 AMD 所生产的那类芯片。Vera 采用的是 Arm 的架构，因此 Modal 需要调整其基础设施以及用于构建软件镜像的系统。Nvidia 上周将公司首台 Vera 服务器亲自送达了其旧金山办公室，Modal 表示。
It will be Modal’s first move away from x86 chips, the kind made by Intel and AMD. Vera uses Arm’s design instead, so Modal will need to change its infrastructure and the system it uses to build software images. Nvidia hand-delivered the company’s first Vera server to its San Francisco office last week, Modal said.
沙盒是隔离的计算机，AI 代理在此运行它们编写的代码。该产品每年可创造 3 亿美元收入，Modal 在上周宣布融资 7.5 亿美元、估值达 157.5 亿美元时透露了这一数据。
Sandboxes are isolated computers where AI agents run code they write. The product makes $300M a year, Modal said last week when it raised $750M at a $15.75B valuation.
沙盒工作负载难以规划，Modal 表示。一个代理可能在等待 AI 模型时保持空闲，随后在软件安装过程中会使用十倍以上的计算能力。每台服务器上会并行运行数百个这样的沙盒，因此一个客户的使用激增可能会拖慢另一个客户的工作。
Sandbox workloads are hard to plan for, Modal said. An agent may sit idle while it waits for an AI model, then use ten times more computing power during a software install. Hundreds of these run side by side on each server, so one customer’s spike can slow down another’s work.
为测试该芯片，Modal 构建了一个包含 50 项编码任务的基准测试，例如编译软件和搜索文件。逐项运行时，Vera 通常比 Intel、AMD 和亚马逊的芯片更快，但并非总是如此，且优势并不显著。
To test the chip, Modal built a benchmark of 50 coding tasks, such as compiling software and searching files. Run one at a time, Vera was often faster than chips from Intel, AMD and Amazon, but not always, and not by much.
当 Modal 在服务器中装满容器时，性能差距逐渐扩大。Modal 表示，Vera 能以不变的速度继续完成任务，而竞争对手的芯片则遇到了性能瓶颈。公司将此归功于 Vera 的高内存带宽。
The gap grew when Modal packed the server with containers. Vera kept finishing tasks at the same speed, while rival chips hit a ceiling, Modal said. The company credits Vera’s high memory bandwidth.
Modal 表示，这一点对两种用途尤为重要：后台编码代理和强化学习。在强化学习中，昂贵的 GPU 常常处于空闲状态，直到 CPU 工作完成为止。
That matters for two uses, Modal said: background coding agents and reinforcement learning. In reinforcement learning, expensive GPUs often sit idle until the CPU work finishes.
英伟达在3月份推出了Vera，这是一款专为AI代理构建的CPU，拥有88个核心，并将Anthropic和OpenAI列为其首批用户之一。Perplexity表示，在报告称编码任务速度提升了1.5倍之后，它将在该芯片上运行代理工作。竞争对手也在加强自己的沙盒产品，Baseten于9月收购了Blaxel。Modal未透露搭载Vera的沙盒何时将面向客户推出。
Nvidia unveiled Vera in March as a CPU built for AI agents, with 88 cores, and has named Anthropic and OpenAI among its first users. Perplexity has said it will run agent work on the chip, after reporting 1.5 times faster coding tasks. Rivals are building up their own sandbox offerings too, with Baseten buying Blaxel in September. Modal did not say when the Vera-powered sandboxes would reach customers.