OpenAI、AMD 用 AI 将芯片设计周期缩短至 5 个月

OpenAI 利用其 AI 模型优化了 AMD 的旗舰 GPU 架构,相较于经过高度优化的基线设计,在功耗、性能和面积(PPA)上最多提升 14%。
OpenAI 在一篇博客文章中透露,此次合作在不到五个月内完成了大约一年的芯片设计进展,AMD 已将部分由模型生成的改动纳入近期送交制造的某款设计。
这项工作涉及使用 OpenAI 的模型重写 SystemVerilog 寄存器传输级(RTL)代码,这类代码描述芯片电路如何执行计算、管理控制操作以及传输数据。
这些模型探索了替代的电路实现方案与流水线编排方式,找到了在降低硬件需求的同时提升性能的机会。
OpenAI 表示,这些模型的水平相当于一位资深微架构师,它们使用 Synopsys 电子设计自动化(EDA)工具来验证电路改动是否保持了原有功能,并评估其对 PPA 的影响。
AMD 已将部分改动纳入近期流片的设计,另一些改动则被加入其量产设计文件,供未来版本使用。
OpenAI 总裁 Greg Brockman 在 X 上发帖称,这项工作利用其模型在不到五个月内为 AMD 的芯片带来了“大约额外一年的进展”。他称这一成果在芯片设计前沿领域“前所未有”。
AMD 还确认,在其即将推出的下一代旗舰 CPU 的关键模块中,采用 OpenAI 模型生成的设计实现了个位数百分比的 PPA 提升。AMD 独立验证了这些提升。
两家公司预计,随着模型持续针对现有基线进行优化,GPU 和 CPU 设计都将取得进一步改进。
这项芯片设计合作建立在 OpenAI 与 AMD 去年 10 月宣布的一项更广泛协议的基础之上。
根据这项多年期协议,OpenAI 承诺为其 AI 基础设施部署最高 6 吉瓦的 AMD Instinct GPU,首批将于 2026 年下半年部署 1 GW 的 MI450 GPU。
AMD 还向 OpenAI 授予了最多 1.6 亿股的认股权证,其归属与部署、技术和商业里程碑挂钩。
此后,这一合作已扩展至硬件与软件的联合开发。
在 AMD 于 7 月举办的 Advancing AI 活动上,两家公司公布了相关计划:利用 OpenAI 的 Triton 框架,并结合 AMD 的 ROCm 软件,在 AMD 的 MI455X GPU 和 Helios 机架级系统上优化 GPT 级工作负载。
OpenAI 预计将于 2026 年第四季度开始将 Helios 系统上线,并在 2027 年全年加快部署。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(OpenAI, AMD Use AI to Cut Chip Design Process From a Year to Under 5 Months)
OpenAI has used its AI models to optimise AMD's flagship GPU architecture, achieving up to 14% improvements in power, performance, and area (PPA) over heavily optimised baseline designs.
In a blog post, OpenAI revealed that this collaboration delivered roughly a year of chip-design progress in less than five months, with AMD incorporating some model-generated changes into a design recently sent for manufacturing.
The work involves using OpenAI's models to rewrite SystemVerilog Register Transfer Level (RTL) code, which describes how a chip's circuits perform computations, manage control operations, and move data.
The models explored alternative circuit implementations and pipeline arrangements, identifying opportunities to improve performance while reducing hardware requirements.
OpenAI said the models operated at the level of an expert microarchitect, using Synopsys electronic design automation (EDA) tools to verify that circuit modifications preserved the original functionality and to assess their PPA impact.
AMD has already incorporated some changes into a recently taped-out design, while others have been added to its production design files for future versions.
Greg Brockman, OpenAI's president, said in a post on X that the work had delivered "roughly an additional year of progress" on AMD's chip in less than five months using its models. He described the results as "unprecedented" at the frontier of chip design.
AMD also confirmed single-digit percentage PPA improvements across key blocks of its upcoming next-generation flagship CPU, using designs generated by OpenAI's models. AMD independently verified the gains.
The companies expect further improvements across both GPU and CPU designs as the models continue to optimise against existing baselines.
The chip design collaboration builds on a wider agreement between OpenAI and AMD announced last October.
Under the multi-year agreement, OpenAI committed to deploying up to 6 gigawatts of AMD Instinct GPUs for its AI infrastructure, beginning with an initial 1 GW deployment of MI450 GPUs in the second half of 2026.
AMD also issued OpenAI a warrant for up to 160 million shares, with vesting tied to deployment, technical, and commercial milestones.
The partnership has since expanded into hardware and software co-development.
At AMD's Advancing AI event in July, the companies outlined plans to optimise GPT-class workloads on AMD's MI455X GPUs and Helios rack-scale systems using OpenAI's Triton framework alongside AMD's ROCm software.
OpenAI expects to begin bringing Helios systems online in the fourth quarter of 2026, with deployments accelerating through 2027.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 Analytics India Magazine(OpenAI, AMD Use AI to Cut Chip Design Process From a Year to Under 5 Months),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。