Latch Based Design for Fast Voltage Droop Response
1 st Shreyas Srinivas, 2 nd Ian W. Jones, 3 rd Goran Panic, 4 th Christoph Lenzen · 2025
Read Latch Based Design for Fast Voltage Droop Response by 1 st Shreyas Srinivas, 2 nd Ian W. Jones, 3 rd Goran Panic, 4 th Christoph Lenzen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-18843.txt.
Introduction
Shreyas Srinivas, Ian Jones, Goran Panic, and Christoph Lenzen redesign a circuit that protects computer chips from sudden voltage dips, in a 2025 paper. When voltage droops, a chip must slow its clock at once or compute garbage. The circuit they rework senses the dip and retunes the clock frequency on the fly.
Their version uses latches instead of flip-flops and drops a timing component the old design needed. It reacts within two clock cycles and halves the synchronizer chain that guards the changeover. A new differential sensor design covers a masking part the earlier work left open. The team also reworked the correctness proofs for the new timing and built the circuit on a real chip process.
Catalog record
| CREATOR | 1 st Shreyas Srinivas, 2 nd Ian W. Jones, 3 rd Goran Panic, 4 th Christoph Lenzen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 59 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-18843
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Latch Based Design for Fast Voltage Droop Response Thanks: This work was supported in part by the European Research Council (ERC) through the European Union’s Horizon 2020 Research and Innovation Programme under Grants 716562 and 101123525.This work was supported in part by the European Research Council (ERC) through the European Union’s Horizon 2020 Research and Innovation Programme under Grants 716562 and 101123525. 1 st Shreyas Srinivas, 2 nd Ian W. Jones, 3 rd Goran Panic, 4 th Christoph Lenzen Abstract We present a latch-based and PLL-free design of the voltage droop correction circuit of Lenzen, Fuegger, Kinali, and Wiederhake [ 1 ] . Such a circuit dynamically modifies the clock frequency of a digital clock for VLSI systems. Our circuit responds within two clock cycles and halves the length of the synchroniser chain compared to the previous design. Further, we introduce a differential sensor based design for masking latches as a replacement for masking flip flops that the design of [ 1 ] requires, but leaves unspecified. The use of latches instead of threshold-altered flip flops alters the timing properties of our design and thus the proofs of correctness that accompanied their design require modifications which we present here. This design has been successfully implemented on the IHP 130 nm process technology. The results of the experimental measurements will be discussed in a subsequent publication. I Introduction Synchronous VLSI circuits are state machines which perform state transitions at each rising clock edge. A critical requirement for maintaining this abstraction is that the combinational circuits computing the transitions finish before the next rising edge of the clock arrives. A droop in the supply voltage VDD interferes with this abstraction by slowing down all transistors. A number of solutions exist to deal with this problem, cf. [ 1 ] for a discussion. One line of work focuses on adapting the clock frequency on the fly to allow for the extra computation time needed. A key challenge in this approach lies in the metastability that might arise when measuring the analog property of a voltage droop in the VDD signal and converting it into a digital droop_detected value. As noted by Marino [ 2 ] , metastability cannot be deterministically detected, resolved, or avoided. It arises from the mismatch between the response of physical circuits to continuous time-varying signals and the abstraction of discrete digital values. Previous solutions either rely on analog properties of the circuit, rendering them fragile to process variations and changes in technology, or funnel the droop detection signal through a synchronizer chain first. The latter means that any response to a droop happens several clock cycles after detection of its onset, which is acceptable for slow droops only. I-A Previous Work Függer, Kinali, and Wiederhake [ 1 ] introduce a fast voltage droop response circuit, which responds to the detection of a droop within a clock cycle. They do this by modifying the synchronizer chain to delay the clock signal if the sampled value indicates a droop, while defaulting to fast propagation if metastability resolution is still ongoing. If metastability resolution results in a “late” decision that a droop was perceived, the clock signal delay is inserted late as well; as the measurement by the droop detector was inconclusive regarding whether the supply voltage had fallen below acceptable levels, this behavior is safe. The reason why this design can circumvent Marino’s result is that the timing delay can be continuously split between the clock pulse at which the resolution happens and the next one, i.e., no discrete decision between delaying and not delaying clock pulses is forced. The key circuit element enabling this is a masking flip flop (hereon referred to as MFF ), which masks internal metastability until resolution takes place, translating it to a late output transition. In [ 1 ] , the implementation of this device is left unspecified, but the following behaviour is required: For b\in\left\{0,1\right\} , a b-MFF is a flip flop that behaves like an ordinary flip flop when there is no metastability. When it is metastable internally, then for the duration of this metastability, it outputs b . Once the internal metastability resolves, the output becomes the respective value. Note that if metastability resolves to b , metastability is fully masked, whereas resolution to 1-b is observed as a late output transition from b to 1-b . Another critical assumption of their circuit design are stable delays across all used delay lines, which they achieve through the use of PLLs. Fig. 1 : Frequency Adaptation Module: The high-level schematics are identical to [ 1 ] , with the exception that the input frequency is twice the output frequency. This allows to add phase delay to the output clock of default period T=50 ns in increments of T/4 without needing accurate delays within the phase accumulator implementation. The phase accumulator decides in each clock cycle whether to delay the next rising clock transition by T/4 based on the binary input G_IN . The delay elements form a synchronizer chain ensuring that the droop_detected signal, which is sampled with each element’s output clock, exposes the phase accumulator to a negligible probability of setup/hold time violations only. However, the delay elements simultaneously serve the purpose of immediately adding a phase shift of T/4 whenever the droop_detected signal is stable 1 upon being sampled. This requires a careful design of the delay element that guarantees a glitch-free clock signal with feasible timing even when the delay element suffers from metastability due to sampling a non-stable value at E_IN . I-B Our Contributions We introduce a latch based design of this circuit which does not use PLLs. Further, we specify and implement masking latches, which replace the flip flops in [ 1 ] . 1 1 1 Alternatively, one might take the view that we implement masking registers in which a pair of registers shares the first latch. This surfaces challenges for the timing properties of the circuit that were hidden by the abstraction of masking registers. Further, for the purposes of completeness, we provide a rudimentary droop detector design, which we used in our chip. We elaborate on aspects of the design motivated by practical implementation issues, such as the asymmetry of buffer delays, the debilitating issues that arise when a proper PLL is unavailable as a cell to stabilise delay lines against process, temperature, and voltage variations, and the avoidance of a glitch in the clock signal that arises from violating certain timing constraints. We successfully implemented our design in the IHP 130 nm technology. I-C Paper Organisation In Section II , we introduce our design of our circuit. This section includes a subsection for each component that describes their functionality as well as sketches of the proofs of correctness. In particular, we introduce our masking latches in this section (see Section II-A ). In Section III , we briefly describe the design choices for the RTL synthesis and layout of our test chip, including parameters like clock frequency and choice of IO ports. II The Design In this section we elaborate on the design changes in the components of our circuit as compared to [ 1 ] . At a high level, the schematic