A：在实验覆盖的规模范围内（从6亿到69亿参数），记忆模块越大带来的提升越明显，且对较小基础模型的增益幅度更大。目前实验还没有观察到明显的效果天花板，但研究也指出，记忆模块的训练预算（训练时间和数据量）同样重要，训练越充分效果越好，说明规模和训练量两个维度都有提升空间。

As embedded systems take on more sophisticated applications, they also require advanced external memory systems, such as DDR3, in order to offer adequate throughput. At times it can be very helpful to ...