Karan Diwan

High-Performance Adder Architectures

Six 64-bit adder topologies, characterized on identical terms.

Binary addition can be built a dozen ways, and each one trades area, speed, and power differently depending on how the carry chain is computed. This project implements and characterizes six 64-bit adders on identical terms so that trade-off is concrete rather than theoretical: Ripple-Carry (RCA), Carry-Select (CSA), Carry-Bypass (CBA), Carry-Lookahead (CLA), Kogge-Stone (KSA), and Brent-Kung (BKA).

Primary characterization uses the Sky130HD standard-cell library, with Nangate45 used for cross-library comparison, and design-space exploration performed on selected architectures to study how block width trades area against timing within a single topology.

Adder architecture performance comparison chart
Estimated Fmax across all six topologies, Sky130HD.

Post-synthesis characterization, Sky130HD

TopologyAreaCrit. pathFmaxPowerADPPDP
Ripple-Carry (RCA)1,761.7 µm²25.00 ns~40 MHz925 µW44,04223,125
Carry-Select (CSA)2,635.0 µm²7.23 ns~138.3 MHz1,580 µW19,05111,423
Carry-Bypass (CBA)2,875.3 µm²24.09 ns~41.5 MHz1,400 µW69,26533,726
Carry-Lookahead (CLA)7,204.4 µm²3.22 ns~310.6 MHz2,590 µW23,1988,340
Kogge-Stone (KSA)3,080.5 µm²3.88 ns~257.7 MHz1,620 µW11,9526,286
Brent-Kung (BKA)2,058.2 µm²12.89 ns~77.6 MHz1,020 µW26,53113,148
ADP (area-delay product) and PDP (power-delay product): lower is better for both. Kogge-Stone gives the best balance on both metrics of the six; Carry-Lookahead wins outright on raw speed at roughly 4× the area of Ripple-Carry.

Relative to Ripple-Carry, Sky130HD

TopologyAreaFmaxPowerADPPDP
Carry-Select1.50×3.46×1.71×0.43×0.49×
Carry-Bypass1.63×1.04×1.51×1.57×1.46×
Carry-Lookahead4.09×7.77×2.80×0.53×0.36×
Kogge-Stone1.75×6.44×1.75×0.27×0.27×
Brent-Kung1.17×1.94×1.10×0.60×0.57×

Block-width study

For block-structured topologies (CLA, CSA, CBA), block width is a free parameter: wider blocks reduce the number of stages the carry has to ripple through but grow the logic inside each block. Sweeping this parameter on the Nangate45 library shows the trade-off directly.

CLA block width versus timing and area chart
Maximum combinational delay and area versus CLA block width, Nangate45.

Cross-library comparison, Nangate45

TopologyAreaCrit. pathFmaxPower
RCA391.6 µm²2.10 ns~476.2 MHz246 µW
CSA519.8 µm²1.11 ns~900.9 MHz446 µW
CBA551.2 µm²2.90 ns~344.8 MHz330 µW
CLA1,441.2 µm²0.85 ns~1,176.5 MHz729 µW
KSA653.3 µm²0.95 ns~1,052.6 MHz462 µW
BKA416.6 µm²1.83 ns~546.4 MHz285 µW
Toolchain: Verilog, Icarus Verilog, Yosys, OpenSTA, GTKWave, Sky130HD, Nangate45
View repository on GitHub →
← Harvard Processor Next: High-Performance Multipliers →