Luca Benini1,2 Michela Milano1

UNIBO1, ETHZ2

Embedded Systems in UNIBO

Brain-like Energy Efficiency

2[RuchIBM11]

1012ops/J↓

1pJ/op↓

1GOPS/mW

2

Energy Proportionality

log(mW)

1GOPS/mW

log(GOPS)

0,003GOPS/mW

0,03GOPS/mW

@1MWSustainable

High-PerformanceComputing

@1mW Energy-NeutralComputationalSensing (1mW)

3

1TOPS/mW

1mW

1MW

1Gop 1Top 1Eop 1Zop

How?

Guiding Principles

Eliminate Waste

Design for power management

Scheduling andallocation

Reduce off-chip I/O Specialized

compute engines

Eliminate Waste

Design for power management

Scheduling andallocation

Reduce off-chip I/O Specialized

compute engines

Be Adaptive

Scale performance based on workload and constraints (e.g. MaxT) and available energy

Dynamic (re-) configuration

Be Adaptive

Scale performance based on workload and constraints (e.g. MaxT) and available energy

Dynamic (re-) configuration

Make Use of Technology

Advanced technology nodes

Low Power Sensors and interfaces

Use novel storage technologies and interfaces

Make Use of Technology

Advanced technology nodes

Low Power Sensors and interfaces

Use novel storage technologies and interfaces

Integrated Systems Laboratory 4

5Departement Informationstechnologie und Elektrotechnik

Compute nodeMany Core ARMv8 SoC

+ GPU Accelerator

UserSpace Energy-

Efficiency APIs

Scalable and fine-grain

power / energy monitoring

engine

High-End ES and HPC

Co-design for Energy efficiency of ARM-V8 +GP-GPU (micro)-servers

Mid-Range ES Heterogeneous SoCs

22.09.2016 Michael Gautschi6

Embedded Multi-Core + GP-GPU (e.g. Nvidia TEGRA, AMD APUs)

Predictabilty + Energy Efficiency

Mid-Range ES Heterogeneous SoCs

22.09.2016 Michael Gautschi7

Massively Parallel SpecializedLS Solver (CNN accelerator)

Sequential General purposeProcessingStandard video (camera) IFs

Embedded Multi-Core + FPGA (Xilinx Zynq SoCs)

Wearable/Implantable ES

Harvesting& PM

ApplicationCircuit

PowerSources

Audio FE

Solar Cell TEG

Multi‐Harvester

Energy consumer Power Domain

3.3 V3.3 V 1.2 V2.0 V

Power Managem.

InfiniTime fully sustainable wearalbe ES

COTS

MCU

Technology Platforms: SW

Programming models Dataflow (in the past) OpenMP, OpenVX (today) Approximate (transprecision) OpenMP (tomorrow)

Allocation and Scheduling Task graphs Stochastic graphs Data-flow graphs Cyclic Graphs

MP-OPT tool

9Departement Informationstechnologie und Elektrotechnik

Refer to Michela, Giuseppe, Alessio’s resentations

Technology Platforms: HW

10Departement Informationstechnologie und Elektrotechnik

Refer to Luca and Franceso presentations

Open-source HW and SW for near-sensor analytics

http://pulp‐platform.org