This paper presents a charge-domain in-memory computing (IMC) macro for precision-scalable deep neural network accelerators. The proposed Dual-SRAM cell structure with coupling capacitors enables charge-domain multiply and accumulate (MAC) operation with variable-precision signed weights. Unlike prior charge-domain IMC macros that only support binary neural networks or digitally compute weighted sums for MAC operation with multi-bit weights, the proposed macro implements analog weighted sums for energy-efficient bit-scalable MAC operations with a novel series-coupled merging scheme. A test chip with a 16-kb SRAM macro is fabricated in 28-nm FDSOI process, and the measured macro throughput is 125.2-876.5 GOPS for weight bit-precision varying from 2 to 8. The macro also achieves energy efficiency ranging from 18.4 TOPS/W for 8-b weight to 119.2 TOPS/W for 2-b weight.
17 May 2021 in IEEE Journal of Solid-State Circuits
A redundancy eliminated flip-flop (REFF) is proposed targeting wide-range voltage scalability (1–0.3 V). Two types of redundancies are eliminated in the REFF to achieve low-power (LP) and reliable operation even in the sub-threshold voltage regime. First, redundant internal clock transitions are eliminated without degrading reliability by finding the optimal way of generating internally inverted clock to reduce dynamic power consumption. Then redundant transistors are identified and eliminated with a topological and logical method while keeping it fully static and contention-free. The simulation results show that the REFF is currently the only FF that fully eliminates redundancy while maintaining static and contention-free operation, and is reliable down to 0.31 V in Monte-Carlo simulations. The measurement results from a test chip fabricated in 28-nm LP CMOS technology show that the measured power is reduced by 69.7%/58.7% with 0%/10% activity at 1 V and by 70.3%/58.2% with 0%/10% activity at 0.4 V compared to the conventional transmission gate flip-flop (TGFF). A total of 100 dies from five corners were tested to demonstrate the reliability, and the REFF was functional down to 0.28 V.
15 November 2022 in IEEE Journal of Solid-State Circuits
A static contention-free differential flip-flop (SCDFF) is presented in 28-nm CMOS for low-voltage and low-power applications. The SCDFF offers fully static and contention-free operation without redundant internal clock transitions with footed differential latches and the same area as a conventional transmission-gate flip-flop (TGFF). The fully static and contention-free operation allows high variation tolerance at a low supply voltage regime, achieving wide-range voltage scalability (1–0.3 V). The measurement results with a test chip fabricated in 28-nm CMOS technology show that power consumption is reduced by 64%/56% with a 0%/10% activity ratio at 1 V compared to that of a TGFF. All 100 dies from five process corners were functional with supply voltage as low as 0.28 V.
28 June 2023 in IEEE Journal of Solid-State Circuits
A pressure-sensing system with a half-Wheatstone-bridge-to-digital converter (HBDC) is proposed for implantable intraocular pressure (IOP) monitoring systems. The half-Wheatstone-bridge (H-WhB) sensor uses an RC-delay comparison instead of direct bias of the pressure transducer, allowing it to self-limit the current for efficient operation. To overcome the limited sensitivity of the H-WhB, energy-efficient bit-level oversampling (OS) is introduced. The system achieves 0.24-mmHg (1 $\sigma )$ resolution with an 8.58-nJ $\cdot$ mmHg $^{2}$ FOM and 12.79- $\mu $ W power consumption with a 11.52-ms conversion time. This marks a 2.8 $\times$ improvement in measured system power, 56.24 $\times$ improvement in energy consumption, and 2.8 $\times$ improvement in resolution compared to the prior H-WhB-based pressure sensor. In addition, the HBDC overcomes the low sensitivity limitations of the H-WhB and achieves a resolution FOM comparable to that of Wheatstone-bridge (WhB) sensor-based pressure sensors.
21 September 2020 in IEEE Solid-State Circuits Letters
A Redundancy Eliminated Flip-Flop (REFF) is presented in 28nm LP process, targeting wide-range voltage scalability (1V to 0.3V). The REFF removes redundant clock transitions to reduce dynamic power consumption and further eliminates redundant transistors with topological and logical methods while keeping it fully static and contention-free. The measured power is reduced by 70%/59% with 0%/10% activity at 1V compared to the TGFF, and 100 dies from 5 corners were functional down to 0.28V.