公开(公告)号：US20220100247A1

公开(公告)日：2022-03-31

申请号：US17033753

申请日：2020-09-26

Applicant: Intel Corporation

Inventor： Vivek Garg ,  Ankush Varma ,  Krishnakanth Sistla ,  Nikhil Gupta ,  Nikethan Shivanand Baligar ,  Stephen Wang ,  Nilanjan Palit ,  Timothy Kam ,  Adwait Purandare ,  Ujjwal Gupta ,  Stanley Chen ,  Dorit Shapira ,  Shruthi Venugopal ,  Suresh Chemudupati ,  Rupal Parikh ,  Eric Dehaemer ,  Pavithra Sampath ,  Phani Kumar Kandula ,  Yogesh Bansal ,  Dean Mulla ,  Michael Tulanowski ,  Stephen Haake ,  Andrew Herdrich ,  Ripan Das

IPC: G06F1/28 ,  G06F1/30 ,  G06F13/20

Abstract: Hierarchical Power Management (HPM) architecture considers the limits of scaling on a power management controller, the autonomy at each die, and provides a unified view of the package to a platform. At a simplest level, HPM architecture has a supervisor and one or more supervisee power management units (PMUs) that communicate via at least two different communication fabrics. Each PMU can behave as a supervisor for a number of supervisee PMUs in a particular domain. HPM addresses these needs for products that comprise a collection of dice with varying levels of power and thermal management capabilities and needs. HPM serves as a unified mechanism than can span collection of dice of varying capability and function, which together form a traditional system-on-chip (SoC). HPM provides a basis for managing power and thermals across a diverse set of dice.

公开(公告)号：US12093100B2

公开(公告)日：2024-09-17

申请号：US17033753

申请日：2020-09-26

Applicant: Intel Corporation

Inventor： Vivek Garg ,  Ankush Varma ,  Krishnakanth Sistla ,  Nikhil Gupta ,  Nikethan Shivanand Baligar ,  Stephen Wang ,  Nilanjan Palit ,  Timothy Yee-Kwong Kam ,  Adwait Purandare ,  Ujjwal Gupta ,  Stanley Chen ,  Dorit Shapira ,  Shruthi Venugopal ,  Suresh Chemudupati ,  Rupal Parikh ,  Eric Dehaemer ,  Pavithra Sampath ,  Phani Kumar Kandula ,  Yogesh Bansal ,  Dean Mulla ,  Michael Tulanowski ,  Stephen Paul Haake ,  Andrew Herdrich ,  Ripan Das ,  Nazar Syed Haider ,  Aman Sewani

IPC: G06F1/28 ,  G06F1/30 ,  G06F13/20

CPC classification number: G06F1/28 ,  G06F1/30 ,  G06F13/20 ,  G06F2213/40

Abstract: Hierarchical Power Management (HPM) architecture considers the limits of scaling on a power management controller, the autonomy at each die, and provides a unified view of the package to a platform. At a simplest level, HPM architecture has a supervisor and one or more supervisee power management units (PMUs) that communicate via at least two different communication fabrics. Each PMU can behave as a supervisor for a number of supervisee PMUs in a particular domain. HPM addresses these needs for products that comprise a collection of dice with varying levels of power and thermal management capabilities and needs. HPM serves as a unified mechanism than can span collection of dice of varying capability and function, which together form a traditional system-on-chip (SoC). HPM provides a basis for managing power and thermals across a diverse set of dice.

公开(公告)号：US12298833B2

公开(公告)日：2025-05-13

申请号：US18007627

申请日：2020-06-26

Applicant: Intel Corporation

Inventor： Ujjwal Gupta ,  Ankush Varma ,  Lakshmipriya Seshan ,  Nikethan Shivanand Baligar ,  Nikhil Gupta ,  Swadesh Choudhary ,  Yogesh Bansal

IPC: G06F1/324 ,  G06F1/10 ,  G06F1/3203 ,  G06F11/34 ,  G06F13/00 ,  G06F15/16

Abstract: A single communication fabric for a data processing apparatus is provided. The fabric has an interconnection network to provide a topology of data communication channels between a plurality of data-handling functional units. The interconnection network has a first interconnection domain to provide data communication between a first subset of the data-handling functional units and a second interconnection domain to provide data communication between a second subset of the data-handling functional units. The power management circuitry is arranged to control a first performance level for the first interconnection domain independently from control of a second performance level for the second interconnection domain. Machine readable instructions and a method are provided to concurrently set performance levels of two different fabric domains to respective different operating frequencies.

公开(公告)号：US11809250B2

公开(公告)日：2023-11-07

申请号：US17074288

申请日：2020-10-19

Applicant: Intel Corporation

Inventor： Ali Mohammad ,  Asma Al-Rawi ,  Ujjwal Gupta ,  Federico Ardanaz ,  Jonathan Eastep

IPC: G06F1/28 ,  G06F1/3296 ,  G06N20/00 ,  G06F1/3206 ,  G06F1/324

CPC classification number: G06F1/28 ,  G06F1/3296 ,  G06N20/00

Abstract: A scheme to improve performance of power-constrained computers, comprising a heterogeneous mix of compute elements, by dynamically reacting to changes in the switching capacitance that present workload induces in each heterogeneous compute element and learning the coefficients of a power-frequency model for each compute element for the present workload. At each time step, the scheme forecasts a maximum frequency that the compute element can run at without exceeding an input power limit for a given workload. The scheme rapidly re-learns coefficients of the power model and rapidly adapts the frequency as the workload's characteristics shift ensuring that compute elements run at the maximum frequency they can while not exceeding the input power limit.

公开(公告)号：US20210124404A1

公开(公告)日：2021-04-29

申请号：US17074288

申请日：2020-10-19

Applicant: Intel Corporation

Inventor： Ali Mohammad ,  Asma Al-Rawi ,  Ujjwal Gupta ,  Federico Ardanaz ,  Jonathan Eastep

IPC: G06F1/28 ,  G06F1/3296 ,  G06N20/00

Abstract: A scheme to improve performance of power-constrained computers, comprising a heterogeneous mix of compute elements, by dynamically reacting to changes in the switching capacitance that present workload induces in each heterogeneous compute element and learning the coefficients of a power-frequency model for each compute element for the present workload. At each time step, the scheme forecasts a maximum frequency that the compute element can run at without exceeding an input power limit for a given workload. The scheme rapidly re-learns coefficients of the power model and rapidly adapts the frequency as the workload's characteristics shift ensuring that compute elements run at the maximum frequency they can while not exceeding the input power limit.