[Retracted] A Post-training Quantization Method for the Design of Fixed-Point-Based FPGA/ASIC Hardware Accelerators for LSTM/GRU Algorithms

Table 3

Comparison of quantization results on the PTB dataset.

	Model	#Layers	#Units	Quantization method	Weights bits	Activation bits	FP model PPW	Quantized model PPW	PPW variation
[13]	LSTM	1	300	In-training	3	3	89.8	87.9	−1.9
[1]	LSTM	1	300	In-training	4	4	109	114	+5
[41]	LSTM	1	300	In-training	4	4	97	100	+3
[42]	LSTM	1	300	In-training	2	2	97.2	110.3	+13.1
Our work	LSTM	1	300	Post-training	11	10	92.8	93.7	+0.9
[13]	GRU	1	300	In-training	3	3	92.5	92.9	+0.4
[1]	GRU	1	300	In-training	4	4	100	102	+2
Our work	GRU	1	300	Post-training	8	3	91.3	90.6	−0.7