首页> 外国专利> Accelerators for deep neural networks

Accelerators for deep neural networks

机译:深神经网络的加速器

摘要

A method is described for reducing operations that are not valid in processing layers of systems, integrated circuits, and neural networks. One or more tiles perform operations, each tile receives an input neuron, an offset, and a synapse, and each input neuron has an associated offset. Each tile produces an output neuron, and there is an activation memory that communicates with the tile via a dispatcher and encoder for storing the neuron. The dispatcher reads the neuron from the activation memory, communicates the neuron to the tile, reads the synapse from the memory, and conveys the synapse to the tile. The encoder receives output neurons from the tiles, encodes them and communicates the output neurons to the activation memory. The offset is processed by the tile to perform operations only on non-zero neurons. Optionally, synapses can be similarly processed to skip operations that are not valid. [Selection] Figure 5B
机译:描述用于减少在处理系统,集成电路和神经网络的处理层中无效的操作的方法。一个或多个图块执行操作,每个图块接收输入神经元,偏移和突触,并且每个输入神经元具有相关的偏移。每个瓦片产生输出神经元,并且存在通过调度器和编码器与图块通信的激活存储器,用于存储神经元。调度程序从激活内存中读取神经元,将神经元传递给瓦片,从内存中读取突触,并将突触传递给瓦片。编码器从瓦片接收输出神经元,对它们进行编码并将输出神经元传送到激活存储器。偏移由图块处理,仅在非零神经元上执行操作。可选地,可以类似地处理突触以跳过无效的操作。 [选择]图5B

著录项

相似文献

  • 专利
  • 外文文献
  • 中文文献
获取专利

客服邮箱:kefu@zhangqiaokeyan.com

京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有
  • 客服微信

  • 服务号