IEEE Solid-States Circuits Magazine - Summer 2022 - 83

has a low contribution to the overall
guardband. The inductive voltage
droop is due to current transients,
and it is relatively slower and global
in nature. These inductive droops are
far deeper and larger than resistive
droops and can, therefore, have a significant
impact on fmax
in multicore
configurations. Inductive droops can
be classified as being first, second,
or third order depending on their
frequency content. However, as they
are relatively slower, it is possible to
sense their occurrence and mitigate
their impact on design margins.
Self-sensing systems monitor variations
and mitigate their impact on design
margining. They can be broadly
classified into two major categories,
namely, " sense-and-adapt " systems
and " error-resilient " systems. We discuss
them in the following sections.
Sense-and-Adapt Systems
" Sense-and-adapt " systems use explicit
sensors to measure dynamic
margins. Actuation mechanisms occur
through ensuring a relaxation
Typically, global and static variations are
relatively easier to compensate for in contrast
to fast-changing voltage-noise effects.
of critical-path timing constraints.
This is achieved by either increasing
the supply voltage through dynamic
voltage scaling and/or reducing the
operating frequency through either
momentary clock stretching or the
long-term reduction of the operating
frequency.
We show the architecture of an alldigital,
real-time voltage-noise sensor
in Figure 5 as an example. The
voltage sampler consists of a VCObased
ADC with additional signal
processing for filtering and decimation
and an SRAM buffer for the storage
of voltage waveform samples. A
triggering engine enables the setting
of programmable triggers for undershoot
and overshoot as well as the
collection of statistics on complex
software workloads, such as tide
marks and voltage-noise events. A
load circuit allows the capability of
stressing the PDN and, thus, achieving
the real-time measurement of
voltage-noise and frequency-domain
measurements of PDN impedance.
Real-time sampling enables the observation
of rare but aperiodic deep
droop events that expose PDN vulnerabilities.
Overshoot and undershoot
examples are shown in Figure 6. On
the left, a branch misprediction in a
max-power virus loop causes a sudden
overshoot. On the right, a CPU
core transitions from a wait-for-event
state to max power, causing a rapid
rise in current consumption, leading
to a deep-undershoot event.
In contrast to real-time sampling,
equivalent time voltage-noise sampling
takes advantage of the inherent
periodicity of the input signal being
measured to achieve comparatively
Pattern
Generator
Registers
Registers
Load Circuits
Tides
Trig1
A57
Cluster
Supply
Trig0
FSM
Sampler
Digital Diff.
On-Chip Oscilloscope
Decimat'n
SRAM
Buffer
(2 Kpts)
IRQ
APB
Counter1
Counter0
FIGURE 5: Real-time sampling using an on-chip digital storage oscilloscope.
IEEE SOLID-STATE CIRCUITS MAGAZINE
SUMMER 2022
83
Trigger Engine
Async Bridge
APB Slave
IEEE Solid-States Circuits Magazine - Summer 2022

Table of Contents for the Digital Edition of IEEE Solid-States Circuits Magazine - Summer 2022
