28 July 2025 in IEEE Microwave and Wireless Technology Letters
This article presents a 6–8-GHz fractional-N charge-pump-phase-lock-loop (CP-PLL) working down to 3.5 K temperature in 28-nm bulk CMOS for quantum computing. To accommodate the wide operating temperature and frequency tuning ranges, a self-biased current source calibration method is introduced to maintain a constant charge pump bias current, thereby ensuring adequate phase margin and minimizing phase noise. An automatic flicker noise calibration technique is also employed to mitigate flicker noise degradation, while a sampled loop filter is utilized to reduce spur degradation at cryogenic temperatures. Owing to these methods, this phase-lock loop (PLL) achieves a phase noise of −90 dBc/Hz at 10 kHz and a jitter of 445 fs at cryogenic temperature while using a power consumption of 4.93 mW. The reference spur is below −70 dB although the mismatch of charge pump degrades at least $2\times $ at cryogenic temperature.
28 October 2025 in IEEE Transactions on Circuits and Systems I: Regular Papers
This paper presents a 12-bit, two-stage, 1GS/s pipelined-SAR ADC in 28-nm CMOS. The design leverages an open-loop amplifier and an alternate-comparator SAR ADC to enhance speed and power efficiency. A dither-based background calibration scheme is proposed to mitigate gain and offset errors with minimal hardware overhead and no timing penalty. The gain error calibration reuses the sub-ADC to resolve additional bits for interference cancellation, enabling fast convergence. The offset calibration directly tunes the relative offset between alternate comparators based on LSB distribution. The prototype achieves a signal-to-noise-and-distortion ratio (SNDR) of 60.1 dB and a spurious-free dynamic range (SFDR) of 75.7 dB with a Nyquist input. With interference cancellation, the gain calibration convergence is accelerated from $4.4\times 10 {^{{6}}}$ to $1.0\times 10 {^{{6}}}$ cycles. It consumes 14.72 mW from a 1V supply. The corresponding Schreier and Walden figure-of-merit (FoM) values are 165.4 dB and 17.8 fJ/conv.-step, respectively.
27 May 2026 in IEEE Journal of Solid-State Circuits
This article presents a cryogenic quantum interface chipset at 3.5 K for superconducting transmon qubit operations. The chipset comprises a phase-detection readout and a phase-shifter-based polar-modulation controller with flexible scalability. With the proposed phase-detection readout scheme, a 9-bit time-to-digital converter (TDC)-based state detector is used to read out the qubit state by detecting the phase difference between the stimulus signal and the reflected signal. A low-IF N-path filter-based receiver and a multi-tone transmitter are co-designed to enable frequency-division multiplexing (FDM) dispersive readout of multiple qubits. A phase-shifter-based polar-modulation XY-driver is developed with support for fast phase quantum gate operation and derivative-removal-by-adiabatic-gate (DRAG) pulse generation. The Z-driver is realized using a static digital-to-analog converter (DAC) array and a 10-bit dynamic DAC for power efficiency. Fabricated in a 28-nm bulk CMOS process, the chipset occupies a die area of 4.6 mm2. The readout RX achieves a noise temperature of 40 K with a gain of 45–83 dB and a bandwidth of 1.2 GHz. The readout RX consumes 32% less power than state-of-the-art designs. The measured total readout power for simultaneous detection of two qubits is 12.2 mW, which is the lowest among readouts with a similar integration level. In addition, the controller achieves a spurious-free dynamic range (SFDR) of 49 dB for a 1-GS/s Z-driver and an SNR of 42 dB for a 1-GS/s XY-driver. The XY-driver consumes 4.3 mW per qubit under active control. Verified with practical transmon qubits, the readout fidelity exceeds 92.6%, approaching that of commercial instruments. A Rabi experiment at 5.61/6.36 MHz, a Ramsey experiment at 4.1 MHz, and a T1 experiment at $9.98~{\mu }$ s are also successfully observed using the proposed chipset at 3.5 K.
4 March 2026 in IEEE Transactions on Circuits and Systems I: Regular Papers
The autoregressive nature of Large Language Models (LLMs) has enabled remarkable performance in language generation, making them a cornerstone in natural language processing. As context windows lengthen, the per-token key–value (KV) cache grows linearly with sequence length and turns the generation stage into a memory-bound operation. However, existing works focus primarily on an isolated stage of KV cache reduction, and most efforts incur significant hardware overheads. This results in a lack of cross-stage co-optimization and leaves significant reduction potential untapped. To address these challenges, we propose CrossKV, a software-hardware co-design architecture that accelerates LLM inference to fully achieve the potential of KV caching optimization. Motivated by our observation, we identify the three stages in KV caching and present a cross-stage co-optimization algorithm, including: a derivative-enhanced dynamic progressive pruning method, a DCT-driven key-vector low-rank compressing, and a dynamic clustered hybrid run-length encoding to reduce the KV caching while maintaining high accuracy. Then, an efficient architecture featuring cross-stage dynamic co-optimization with negligible hardware overhead is proposed to fully harness the algorithm of CrossKV. Our evaluations of CrossKV over 9 popular LLMs models and various long-context tasks demonstrate an average of $6.03\times $ and $4.10\times $ improvement for energy efficiency and speedup compared to existing SoTA Accelerators architecture. Compared to the Nvidia A100 GPU, CrossKV achieves an average $23.43\times $ energy efficiency and $4.64\times $ speedup, respectively.
24 July 2025 in IEEE Journal of Solid-State Circuits
An ultra-low-power (ULP) six-class keyword spotting (KWS) ASIC is presented in this article, which can be used in always-on speech-based human–machine interface applications. The ASIC is composed of a sampling frequency and resolution adaptive (SFRA) dual-mode analog-to-digital quantizer and a power-gated KWS engine with a multiplier-less processing element (PE) array. The dual-mode quantizer mainly operates as a 1.5-bit delta quantizer (DQ), providing inherent robustness against dc drift. It adaptively switches to high-resolution successive-approximation register (SAR) quantization mode upon detecting a sound event based on the DQ output. In the KWS engine, the multiplier-less PE array is reused for both feature extraction and gated recurrent unit (GRU)-based keyword classification. To unveil the tradeoffs between power consumption and flexibility, two versions of the classifier have been implemented, with ROM-based on-chip weight memory (WM) and SRAM-based WM. Fabricated in 180-nm CMOS technology, the proposed KWS ASIC with ROM and SRAM-based WM achieves six-class classification accuracies of 87.3% and 90.1%, respectively, on the Google Speech Command dataset (GSCD) while consuming 174 and 756 nW long-term average (LTA) power with a decision latency of 14 ms at a clock frequency of 256 kHz.
1 January 2026 in IEEE Journal of Solid-State Circuits