JVET-AA0080 proposes to use CP decomposition plus fusing adjacent 1x1 convolution to reduce the complexity of NNLF. In addition, it also proposes to split architecture for luma and chroma components between input network and output network. Simulation shows that using JVET-X0140 low complexity model (EE1-1.4.1) as baseline, whose worse case block level complexity is 33.6 KMAC/Pixel, the proposal can achieve the AI luma gain about 5% with worst case complexity of 17.7 KMAC/pixel for (24L,8C) split. This contribution applies the same techniques proposed in JVET-AA0080 and investigates the RA case. The same baseline model shows the BD-rate saving of (Y, Cb, Cr, respectively) for RA compared to EE1. For the proposed complexity reduction techniques, when compared against EE1 , the BD-Rate saving for RA (Y, Cb Cr, respectively) is: