Professional Documents
Culture Documents
6, December 2019
https://doi.org/10.5573/IEIESPC.2019.8.6.506 506
School of Computer Science and Engineering, Kyungpook National University, Daegu, Korea yongtae@knu.ac.kr
Received August 26, 2019; Accepted September 16, 2019; Published December 30, 2019
* Short Paper
Abstract: This paper proposes an approximate adder that employs a novel carry speculation scheme
to enhance the computation precision of the existing error tolerant adder (ETA) designs with
extremely little hardware overhead. The proposed carry prediction technique leverages two input
bits to increase the prediction accuracy while the conventional ones do only one bit. This leads to a
reduction of the carry prediction error rate from 25% to 18.75%. Compared to the existing ETA
design, the proposed adder reduces normalized mean error distance (NMED) and mean relative
error distance (MRED) by up to 10% and 28%, respectively, at the cost of only a two-input OR
gate. Moreover, the proposed design outperforms the conventional ETAs when jointly evaluating
hardware cost and computation accuracy. Specifically, the new design allows 11% and 17%
reductions of area-power-NMED and power-NMED products, respectively, compared to the
traditional ETA.
Keywords: Approximate adder, Error tolerant computing, Error tolerant adder (ETA), Carry predicting error
tolerant adder (CPETA)
k-bit
Precise Cin
Adder
Fig. 1. Block diagram of the proposed enhanced carry predicting error tolerant adder (ECPETA).
error occurs when both input bits of the n-k-2 LSB position Table 1. Comparison of the proposed adder with two
are 1. It is written as other approximate adders when n=16 and k=8.
n − k −1
Area Power PDP
Adder Delay (ns)
⎛3⎞ (µm2) (µW) (fJ)
ERECPETA (n, k ) = 1 − ⎜ ⎟ (2)
⎝4⎠ ETAI 108 1.09 26.3 28.7
CPETA 114 1.14 28.0 31.9
ECPETA 116 1.26 29.7 37.4
3. Experimental Results
The proposed approximate adder, together with the 1 for all the adders. Those of the proposed adder are
ETAI and CPETA, were designed in the Verilog HDL and enhanced by up to 8% to 10% and 54% to 56%,
synthesized with a 65-nm CMOS standard cell library in respectively, compared to the CPETA and ETAI. In
order to evaluate and compare circuit performance. The addition, the MRED of all adders is reduced by between
parameters of n=16 and k=8 were considered and the 0.5 and 1 when k increases by 1. The proposed design
precise adder was implemented with an RCA for all these allows the MRED to decrease by 9% to 28% and 17% to
approximate adders. Also, software simulations were 43% compared to the CPETA and ETAI, respectively.
conducted to analyze the computation accuracy of the Thanks to the proposed carry speculation method, in short,
adders, and ER, CPER, MRED, mean error distance this design outperforms the other two ETA designs from
(MED), and normalized MED (NMED) were extracted all computation accuracy perspectives.
under 10 million (i.e., 107) uniformly generated random To validate the CPER calculated in Section 2, the
input patterns. CPER of the proposed design and the CPETA at the
Table 1 summarizes the circuit performance of the different sizes of k were also simulated. Fig. 3 shows the
proposed adder and the other two approximate adders. The CPER of the two 16-bit adders under different sizes of k.
lack of a carry prediction method makes the ETAI the Please note that the ETAI has no carry prediction scheme,
fastest adder among them. Also, it is the most efficient and hence, it was not included in this simulation. As
design with respect to area, power, and power delay discussed in Section 2, the CPER of the CPETA reaches
product (PDP), although it exhibits the worst accuracy, around 25% (24.97% to 24.22%), whereas the proposed
which will be investigated later in this section. Since the design maintains a CPER of about 18.75% (18.75% to
proposed ECPETA necessitates an OR gate to more 17.99%), regardless of k. As an example, the CPER of the
accurately speculate the carry, it causes a longer critical CPETA and the ECPETA are 24.81% and 18.56%,
path delay than the others. Unfortunately, this also makes respectively, for k=8, resulting in a CPER reduction of
the proposed design the most energy-consuming (i.e. PDP) 25.19%. In fact, this improvement in the carry prediction
adder. It is very important to note that the critical path accuracy enables the existing ETA designs to enhance
delay of these adders depends on either parameter k or the their overall computation accuracy.
delay of the precise adder (i.e., tpa), as seen in (1). It means To jointly evaluate the power and accuracy of the
that the overall critical path delay could be the same for all approximate adders, the power-NMED product was
three adders when k is small enough. From the area and presented in [11]. This represents the overall efficiency of
power perspectives, the proposed design is comparable to approximate adders, considering both hardware cost and
the CPETA. precision. Furthermore, the area-power-NMED product
Table 2 provides a computation accuracy comparison can be defined to take are into account, as well. Fig. 4
of the 16-bit approximate adders for different sizes of k. It exhibits the products for the three approximate adders with
shows that the accuracy of the approximate adders clearly n=16 and k=8. To effectively compare these metrics
improves as the size of the precise adder k increases. The among the adders, they were normalized by the ETAI, and
ER of the CPETA and ECPETA is identical, as explained the values outside of the bars were inserted. The results
in Section 2, and is slightly lower than the ETAI. The clearly show that the proposed design is superior to the
MED and NMED decrease by nearly half as k increases by ETAI and CPETA in terms of both the power-NMED
Table 2. Accuracy comparison of the 16-bit approximate adders for different sizes of k.
Acknowledgement
Fig. 3. CPER of the 16-bit approximate adders for This research was supported by Kyungpook National
different sizes of k. University Research Fund, 2018.
References