精选理由
XPower用12颗小芯片组卡,跑400B到1.6TB参数的模型还能把token成本压到最低,想省钱搞推理可以看看。
XPower Technology在WAIC上展示了RPP架构的AE7100E芯片,其单颗芯片仅指甲大小。一张加速卡可集成12颗AE7100E,通过分布式计算支持400B到1.6TB参数的模型推理。官方声称在同等参数量下可实现最低token成本。
原文 · Pandaily
XPower Technology Defies Big-Chip Orthodoxy: Small RPP Architecture Chips Use Distributed Computing to Beat Giants at AI Inference
XPower Technology unveils RPP architecture AE7100E chips at WAIC: fingernail-sized chips with 12 on one card, supporting 400B-1.6TB parameter models through distributed computing at lowest token cost.