Cerebras 宣布将其晶圆级引擎(Wafer Scale Engines)与 GPU 结合,推出生产环境的“解耦推理”架构 1。该方案的架构分工明确:GPU 负责处理提示词预填充(Prefill),而 Cerebras 系统负责高速生成解码(Decode)1。
数据显示,相比仅使用 Cerebras CS-3 系统的配置,新方案在保持极速生成的同时,将每兆瓦的 Token 吞吐量提升了 5 倍 1。硬件优势方面,Cerebras 单芯片拥有数十 GB SRAM 和数十 PB/s 内存带宽,这分别是典型 GB300 GPU 的 350 倍 SRAM 容量和 5400 倍带宽 1。
目前,Amazon Web Services、Fireworks AI、Gimlet Labs 和 General Compute 等早期合作伙伴已接入该解决方案 1。在生产环境的应用实例中,测试使用了 AMD Instinct GPUs 配合 Cerebras CS-3 系统 1。Cerebras 计划在 2027 年底前,在数百兆瓦规模上广泛部署解耦推理 1。
Cerebras has announced a production-grade "disaggregated inference" architecture that combines its Wafer Scale Engines with GPUs to optimize AI model performance. 1 In this setup, GPUs handle the prefill phase of processing prompts, while Cerebras systems manage the high-speed generation or decoding phase. 1 This division of labor aims to maximize token throughput per megawatt without sacrificing generation speed. 1
The new approach delivers a fivefold increase in token throughput per megawatt compared to configurations relying solely on Cerebras CS-3 systems. 1 A key hardware advantage driving this efficiency is the massive memory capacity and bandwidth of Cerebras' single-chip design, which features tens of gigabytes of SRAM and tens of petabytes per second of memory bandwidth. 1 These specifications represent 350 times the SRAM capacity and 5,400 times the bandwidth of a typical GB300 GPU. 1 Early partners such as Amazon Web Services, Fireworks AI, Gimlet Labs, and General Compute have already adopted the solution, with tests involving AMD Instinct GPUs working alongside Cerebras CS-3 systems. 1 Looking ahead, Cerebras plans to widely deploy disaggregated inference across hundreds of megawatts by the end of 2027. 1
评论
还没有评论,欢迎留下第一条。