Databricks’ former AI chief thinks he can cut AI’s power bill by 1,000x
Databricks’ former AI chief thinks he can cut AI’s power bill by 1,000x
Databricks 前 AI 负责人认为他能将 AI 的能耗降低 1000 倍
The drive to discover the next big thing in AI has funded some pretty ambitious projects — but one company is taking it as a chance to rebuild computing architecture from the ground up. Led by Naveen Rao, formerly the head of AI at Databricks, Unconventional AI promises to make inference processing vastly more power efficient. 为了发掘 AI 领域的下一个重大突破，许多雄心勃勃的项目获得了资金支持，但有一家公司将其视为从底层重构计算架构的契机。在曾任 Databricks AI 负责人的 Naveen Rao 的领导下，Unconventional AI 公司承诺将大幅提升推理处理的能效。
The secret weapon: a new kind of oscillator-based computer architecture. On Thursday, the company released its first model AI — called Un-0 — an image-generation system tool that shows for the first time how the company’s technology can replicate conventional AI systems. 其秘密武器是一种基于振荡器的新型计算机架构。周四，该公司发布了其首个 AI 模型——名为 Un-0。这是一个图像生成系统工具，首次展示了该公司的技术如何能够复刻传统的 AI 系统。
In an accompanying new paper, the company’s research team details how they built a fully functional image-generation model using a software simulation of the new architecture — one that performs just as well as state-of-the-art diffusion models. “This is the ‘hello world’ of a new kind of computer,” Rao told TechCrunch. “Over the next year, you’re going to start seeing some pretty interesting news around this.” 在随附的一篇新论文中，该公司的研究团队详细介绍了他们如何利用新架构的软件模拟构建了一个功能完备的图像生成模型，其性能表现与最先进的扩散模型不相上下。“这是新型计算机的‘Hello World’，”Rao 在接受 TechCrunch 采访时表示，“在接下来的一年里，你将会看到一些关于此的非常有趣的消息。”
The output from the new Un-0 model is similar to that of image-generation models like Stable Diffusion or OpenAI’s GPT Image 1. The impressive part is how it arrives at that performance. The model is built on an oscillator-based architecture that is completely different from the chips that power conventional computing and traditional LLMs. Un-0 模型生成的输出结果与 Stable Diffusion 或 OpenAI 的 GPT Image 1 等图像生成模型相似。令人印象深刻的是它实现这一性能的方式。该模型构建于一种基于振荡器的架构之上，这与驱动传统计算和传统大语言模型（LLM）的芯片完全不同。
The advantages of the oscillator-based computing are complex, but Rao believes it will ultimately reduce power use by as much as 1,000 times. Much of the infrastructure to get there is still being built. The current version of Un-0 runs on a software simulation of Unconventional’s oscillator chips, but the company plans to release schematics for an actual chip soon. 基于振荡器的计算优势非常复杂，但 Rao 相信它最终能将能耗降低多达 1000 倍。实现这一目标所需的大部分基础设施仍在建设中。目前版本的 Un-0 运行在 Unconventional 公司振荡器芯片的软件模拟上，但该公司计划很快发布实际芯片的原理图。
From there, the plan is to build an entire inference stack from the ground up, with Unconventional AI eventually supplying compute capacity just like any other provider. “We will build a new kind of system composed of our chips,” says Rao. “We will run AI models there, and we will have a network cable where prompts come in and inferences go out, but it’ll be done at 1/1000 of power.” 此后，他们的计划是从零开始构建完整的推理技术栈，最终让 Unconventional AI 像其他供应商一样提供计算能力。“我们将构建一种由我们的芯片组成的新型系统，”Rao 说，“我们将在上面运行 AI 模型，通过网线输入提示词并输出推理结果，但其能耗仅为原来的千分之一。”
It’s a stunningly ambitious goal, particularly for a company that still counts less than 50 employees. But given the scale of the AI buildout and the anticipated cost of meeting the growing demand for inference, it may be one of the few efforts to meet the scale of the problem. 这是一个极其宏大的目标，对于一家员工人数不足 50 人的公司来说尤其如此。但考虑到 AI 建设的规模以及满足日益增长的推理需求所带来的预期成本，这可能是少数能够应对这一规模问题的尝试之一。
As Rao sees it, the available supply of power will be one of the hard limits for AI in the years to come — and Unconventional is one of the few projects able to address it. “AI scaling is hard because of energy. It’s going to be the fundamental limit in the next few years. You just can’t go past it. It’s going to be an energy-limited problem, at the end of the day,” he says. 在 Rao 看来，可用的电力供应将是未来几年 AI 发展的硬性限制之一，而 Unconventional 是少数能够解决这一问题的项目。“AI 的扩展之所以困难，是因为能源问题。这将是未来几年最根本的限制。你无法绕过它。归根结底，这将是一个受能源限制的问题，”他说。