FPGA as the Accelerator of Choice in Data Centric applications
Shpe conference, Chandler, arizonaMario a. bolaňos – november 1ST 2019
An analogy …
CUSTOM LOGO
ASIC
Specífic to a company
High cost from the beginning
Requires high volume
Application Specific Integrated Circuit
DESIGNS PRE-FABRICATED
ASSP
FOR SALE
Designed for specific functions
They are for general purpose use and
anyone could use and purchase them
Application Specific Standard Product
WHITE BOARD
FPGA
Flexible y Configurable
It has pre-fabricated modules that
could be added to the “White board”
Field Programmable Gate Array
WHAT IS A FPGA?
Field Programmable Gate Array, Flexible, multi-functional reprogrammable silicon with
bare-metal speed and reliability and custom parallelism
Custom hardware functionality but most of its electronic functionality could be modified
During design phase
During assembly of producto at customer
It could be modified even after the product has been released to production
Benefits of FPGA Technology
4
• Performance
• Flexibility
• Time to market
• Cost
• Integration
• Reliability
• Energy Efficiency
• Acceleration of Computing
• Long-Term Maintenance
FIRST fpga INTRODUCED IN 1985… ¿WHY IS IT
BECOMING SO CRITICAL NOW?
SOME FACTS RELATED To datA Accumulated Data up to 2013 = 4.4ZB, data estimated at 2020 = 40 ZB
90% of all data in 2018 were generated within the last 2 years
5 million tweets per day
294 billion emails per day
4 Petabytes of data in Facebook per day
65 billion messages in Whatsapp per day
5 billion searches per day
95 million photos and videos in Instagram per day
The average internet user will generate
~1.5 GB of traffic per daySmart hospitals will be generating over
3 TB per daySelf driving cars will be generating over
4,000 GB per day… each
All numbers are approximatedhttp://www.cisco.com/c/en/us/solutions/service-provider/vni-network-traffic-forecast/infographic.htmlhttp://www.cisco.com/c/en/us/solutions/collateral/service-provider/global-cloud-index-gci/Cloud_Index_White_Paper.htmlhttps://datafloq.com/read/self-driving-cars-create-2-petabytes-data-annually/172http://www.cisco.com/c/en/us/solutions/collateral/service-provider/global-cloud-index-gci/Cloud_Index_White_Paper.htmlhttp://www.cisco.com/c/en/us/solutions/collateral/service-provider/global-cloud-index-gci/Cloud_Index_White_Paper.html
Self driving cars will be generating over
4 TB per day… eachA connected plane will be generating over
40 TB per dayA connected factory will be generating over
1 PB per day
radar
~10-100
KBper
second
sonar
~10-100
KBper
second
gps ~50 KBper
second
lidar
~10-70
MBper
second
~20-40
And More Facts related To DataBy 2020
¿whathappens in
anINTERNET minute?
The Solution to Data Computing: Smarter Compute Using High Performance Computing
10
Assuming 3 cycle per multiplication operation on a 3 GHz processor. A
single threaded processor can produce 1 new multiplication product
every 1 billionth of a second.
A 2012 ImageNet classification algorithm* takes a 256x256 pixel image
and classifies it against 1000 categories that the image might map to.
This (unoptimized) algorithm takes 12.2 trillion multiplies! With a single
multiply product every 1 billionth of a second, it would take 12,200
seconds to categorize (3 hours, 24 minutes) at one multiply every one
billionth of a second.
https://vast.cs.ucla.edu/sites/default/files/publications/CNN_ICANN14.pdf
High-performance computing (HPC) is the use of parallel processing for running advanced
application programs efficiently, reliably and quickly. The term applies especially to systems that
function above a teraflop or 1012 floating-point operations per second.
THE Balancing data computing Act
11
CPU GPU FPGA ASIC & ASSP Heterogeneous
Peak
Performance
Moderate High Very High Highest Very High
Power
Consumption
High Very High Very Low Lowest Moderate
Flexibility Highest Moderate Very High Lowest Very High
Cost Moderate High Very High Highest* Very High*
Parallelism Very Low Very High Custom Custom Custom
In reality, one architecture cannot solve all the world’s compute problems.
Real-World Applications
How to Choose THE BEST DATA COMPUTING
PROCESSOR?
13
FRONT-END
• What are you trying to achieve?
• Which specifications (speed,
power, cost, time to market) are
most important?
• How many units will you need?
• Where will it be deployed?
• How often to you expect your
design to change?
• What is the expertise of your
engineering team?
• Has someone already built a
solution that’s “good enough”?
CPU GPU ASSP
ASIC FPGA
14
FPGA FOCUS MARKETS
Cloud Computing Autonomous DrivingSmart Cities
Networking5G Wireless Aerospace
Transforming Data Centers and cloud computing
15
CPU GPU ASSP
ASIC FPGA
Artificial Intelligence
Big Data Analytics (Hadoop, SPARK, SQL, NoSQL)
Video Transcoding
Network functions virtualization
Storage Acceleration
Security and DPI (Deep Packet Inspection)
16
CASE STUDY: Microsoft
125%Gain in
Throughput
29%DECREASE in Latency
8XIncrease in speed
With 15% less power
Accelerating the Critical tasks of autonomous
driving
17
Sensor Fusion
AI/Machine Learning
Functional Safety
5g connectivity
18
Case Study: Embedded Systems
Video: How Intel FPGAs Enable the Industrial Internet of Things
19
Case Study: Video and Vision
Real-Time Analytics
Simultaneous motion
detection, facial
recognition, and object
detection
Multiple input feeds
Flexible Sensor Interfaces
Adapt to changes in
proprietary interfaces
without changing the
rest of the design
Signal Processing
Pick and choose video
processing functionality
using IP cores
Accelerate pre-processing
of high-res videos
Video Compression
Integrate CODECs with
other processing functions
on a single FPGA
20
What is ARTIFICIAL INTELLIGENCE (AI)?Artificial Intelligence
Data Analytics
Build a representation, query, or model that enables descriptive,
interactive, or predictive analysis over any amount of diverse data
Sense, learn, reason, act, and adapt to the real world without
explicit programming
Perceptual Understanding Detect patterns in audio or visual data
Machine LearningComputational methods that use learning algorithms to build a model from data (in supervised,
unsupervised, semi-supervised, or reinforcement mode)
Deep LearningAlgorithms inspired by neural networks with multiple
layers of neurons that learn successively complex
representations
Convolutional Neural Networks
(CNN)DL topology particularly effective at
image classification
AI is Transforming Industries
21
Smart
Assistants
Chatbots
Search
Personalization
Augmented
Reality
Robots
Enhanced
Diagnostics
Drug
Discovery
Patient Care
Research
Sensory Aids
Algorithmic
Trading
Fraud
Detection
Research
Personal
Finance
Risk Mitigation
Support
Experience
Marketing
Merchandising
Loyalty
Supply Chain
Security
Defense
Data Insights
Safety &
Security
Resident
Engagement
Smarter Cities
Smart Grid
Conservation
Operational
Improvement
Oil & Gas
Exploration
Automated Cars
Automated
Trucking
Aerospace
Shipping
Field
Automation
Factory
Automation
Predictive
maintenance
Precision
Agriculture
Field
Automation
CONSUME
R
HEalth FINANCE RETAIL GOVERNME
NT
ENERGY TRANSPOR
T
INDUSTRIA
L
22
BENEFITS OF Intel FPGAS fOR AI
Power Efficiency
Reduced total cost of ownership
Speed
Real-time decision making
Throughput
Do more with less
Deployment Flexibility
Offloaded or in-line processing
I/O Types
Direct interface to data source
Power envelope
Only use as much as you need
Precision
Customizable precision and data types
Future Algorithms
Adaptable to architectures of the future
Multi-functionality
AI and more, all on one chip
PERFORMANCE HARDWARE FLEXIBILITY WORKLOAD FLEXIBILITY
Management,
Sensors and
edge devices
Vision systems,
and purpose-built,
application-specific hardware
Scalable and efficient
computing performance
Cloud, datacenter,
and HPC
Copyright © 2017, Intel Corporation. All rights reserved.
Board Management
Edge Compute
I/O Expansion
Automobile sensors, traffic sensors
Machine Vision
Embedded Vision
Robotics
Infotainment
Datacenter
Networking
Military / Defense
ADAS
Datacenter / CSP Acceleration
5G Wireless Infrastructure
Network Communications
Military / Defense
The right performance and features for the right application
Intel® FPGA : application specific performance
High-Level Design (HLD) Portfolio
24
Algorithm
Designer
Software
Programmer
Embedded
Designer
Hardware
Designer
Intel® FPGA SDK
for OpenCL™
DSP Builder for
Intel® FPGAs
Intel® HLS
Compiler
HDL Code,
Qsys (schematic)
PE
RS
ON
A
FRONT-END
Generated RTL
Generated RTL
Quartus®
PrimeGenerated RTL
Generated RTL
READY TO LEARN MORE?
25
Online Training
Deepen your expertise with Intel
FPGA training courses.
Intel FPGA YouTube
Short videos on tools and
technology
Design Examples
Get started with Intel FPGA products
with ready-to-use design examples
Community Forum
Get your questions answered by
Intel® FPGA technical experts
Intel® FPGA Academic Program
Online tutorials, labs, curriculum,
software and hardware.
Click the images to be taken to our website
Intel AI Academy
Videos and classes on AI, Machine
Learning and Deep Learning
THANKS