Rohan Paul@rohanpaul_ai · XAI芯片制造商Cerebras正计划在美国进行IPO，目标融资高达35亿美元，每股定价区间为115至125美元。该公司采用独特的晶圆级芯片技术，将几乎整个硅晶圆集成为一个巨型处理器，旨在通过将海量计算和内存置于单一硅片上，减少芯片间数据移动的瓶颈，从而为特定大型AI工作负载加速训练和推理过程。其技术路线与依赖多GPU协作、受限于数据移动速度的传统AI硬件形成对比。此次IPO是公司在撤回先前尝试后的新一轮努力，其竞争对手包括英伟达和AMD，后者同时也是其支持者。
Reuters： Cerebras is looking to raise as much as $3.5B through its U.S. IPO. It competes with Nvidia and AMD， which is also a backer.
Cerebras builds wafer-scale AI chips， i.e. uses an almost entire silicon wafer as one huge processor instead of cutting the wafer into many smaller chips.
Normal AI hardware spread work across many GPUs， so speed depends not only on raw compute but also on how fast data can move between chips， memory， and networking gear.
Cerebras is betting that putting massive compute and memory on one piece of silicon can reduce that data movement problem and make training and inference faster for certain large AI workloads.