1. SILICON? CHECK. HSA? CHECK.
ALL DONE? WRONG!
TONY KING-SMITH
EXECUTIVE VP MARKETING, IMAGINATION TECHNOLOGIES
2. FROM MOBILE TO DATACENTER, IT ALL STARTS WITH THE SOC…
Ensigma
RPU
Radio comms
processor
PowerVR
GPU
Graphics processor
MIPS CPU
General
processor
PowerVR VPU
Video & Vision processor
2 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
3. THEN WE NEED A WAY FOR PROCESSORS TO COMMUNICATE…
Ensigma
RPU
Cloud
technologies
MIPS CPU
General
processor
Bus Fabric
Flow
Radio comms
processor
PowerVR
GPU
Graphics processor
Unified
Memory
PowerVR VPU
Video & Vision processor
3 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
4. AND WE WRITE APPS THAT USE ALL THESE RESOURCES – THANKS TO HSA
Ensigma
VLIW
Kernel?
RPU
GPU Kernel
Radio comms
processor
MIPS HSAClever CPU
General
aware app
processor
Bus Fabric
The Cloud
PowerVR
GPU
GPU Kernel
Graphics processor
GPU Kernel
Unified
Memory
PowerVR VPU
HW Kernel?
Video & Vision processor
4 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
5. HOORAY!
ALL THE WORLD’S PROBLEMS SOLVED…?
5 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
6. DREAM ON…
Many questions to be answered
‒ Portability
‒ Scalability
‒ Simulation & modelling
‒ Power Management
This is complex stuff – so it will take time to learn how to use it
But the most important question is: relevance to developers
‒ Will developers actually be able to use HSA?
‒ Does HSA address the needs of tomorrow’s applications?
6 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
7. TARGET MARKETS
IS HSA ADDRESSING THE RIGHT ONES?
Fact:
‒ The highest volume market by unit volume is smartphones and tablets
Fact:
‒ Developers need to make money; they enjoy eating and drinking
Fact:
‒ More and more applications that are engaging consumers are on mobile
platforms from consumer to enterprise
‒ So that’s where more and more developers are focused
Fact:
‒ The SoC platforms being used in these markets are changing so rapidly,
optimising for any one platform makes little or no sense for developers
If HSA does not target these markets, it will only be relevant to high
performance developers – a vital community, but a specialised one
HSA must address mobile platforms and developers!
7 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
8. ALL PLATFORMS BY OS
3,000,000
2,500,000
Others
Windows
2,000,000
Symbian
RIM
MeeGo
1,500,000
LiMo Foundation
iOS/MacOS
1,000,000
Chrome
Bada
500,000
Android
0
2010 2011 2012 2013 2014 2015 2016 2017
8 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
Source: Imagination
9. GPUS SCALE WITHIN ONE GENERATION
Mobile GPU processing power (log)
PERFORMANCE ENVELOPE GROWING WITH EACH SOC GENERATION
256GFLOPs
Widening GPU
performance
envelope – but
power budget
remains static!
128GFLOPs
48GFLOPs
24GFLOPs
12GFLOPs
2013
9 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
2014
2015
Time
10. PORTABILITY AND SCALABILITY
CAN MORE THAN A HANDFUL OF DEVELOPERS HANDLE SCALABLE HETEROGENEOUS APPS?
APIs are intended to aid portability
‒ But HSAIL isn’t the easiest thing to code with; neither is OpenCL,
or OpenGL ES, or…
Apps must now scale over a very broad range of performance at any
given point in time
‒ Todays leading edge smartphones are tomorrow’s commodity
The reality is the numbers of developers actually skilled in using these
APIs is relatively small compared to the much broader community
‒ So middleware is essential, e.g. games engines, UI engines, function libraries
We need to stimulate the middleware community to get
traction for HSA
Writing a portable, scalable app is an emerging art form!
10 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
11. OTHER ISSUES ABOUND
HOW WILL DEVELOPERS GET TO GRIPS WITH THE COMPLEXITIES OF HETEROGENEOUS PLATFORMS?
Simulation and modeling
‒ Getting harder all the time!
‒ App understanding of the underlying SoC platform has increasingly
dramatic implications on system performance
Power management
‒ The application needs to control this, not just rely on the OS
Processing profiles are increasingly dynamic
‒ Dynamically varying voltage, clock, thermal
‒ Switching in and out of GPU and CPU cores; shared resources
If we’re not careful, we end up with the perfect solution that’s far
too clever for mere mortals to use!
11 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
12. CHALLENGING CONVENTIONAL WISDOM
OR…WHERE’S THE MONEY?
Myth: Desktop apps will move to mobile
‒ No they won’t! Look at all the successful mobile apps…
‒ Mobile apps are different - touch-based, highly interactive,
and designed to scale across a wide range of platforms
Myth: High performance apps will move to mobile
‒ Why should they? They’re best suited to cloud servers
Myth: Cloud computing will kill serious computing on mobile
‒ No, they will co-exist, and the smart money is on those that use both
HSA must not ignore the power and influence of mobile platforms
12 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
13. HOW DEVELOPER COMMUNITIES ARE PARTITIONED…
Cloud Server App
Developers
Mobile App
Developers
Big data
QoS depends on link
Medium-long timeframe
Immediate response
Always available
Instant/short timeframe
13 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
14. BUT THIS IS HOW IT WILL EVOLVE, AND QUICKLY…
Heterogeneous
App Developers
Immediate response
Always available
Instant to long timeframe
Cloud-based
analysis
14 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
Immediate
response
15. SO IF YOU HAVE A GREAT LINK…
Heterogeneous
App Developers
Immediate response
Always available
Instant to long timeframe
Cloud-based analysis
15 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
Immediate
response
16. BUT IF IT’S A BIT DODGY…
Heterogeneous
App Developers
Immediate response
Always available
Instant to long timeframe
Cloud-based analysis
16 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
Immediate
response
Local
analysis
17. OR DIES ALTOGETHER…
Heterogeneous
App Developers
Immediate response
Always available
Instant to long timeframe
17 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013
Immediate
response
Local short
and longer
timframe
analysis
18. CONCLUSIONS
HSA is great – but it must target both mobile and performance SoCs in order for it to be widely adopted
High volume mobile platforms will always be cost-sensitive
‒ So you can never assume high end capabilities will migrate onto them
Mobile platforms mean much more than smartphone app processors
‒ Data center computing will increasingly rely on low power SoC technologies derived from mobile
‒ Mobile chips will be used in ever broader markets
‒ The Internet of Things will be a key source of data essential for rich analytics – low cost, low power again
The best developers will go where the money is
‒ If HSA is to be successful, it must target volume markets, not start at the high end
Apps, not just SoCs, are going to be heterogeneous in many ways
‒ Partitioned across mobile and datacenter platforms, dynamically reconfiguring in response to environment
18 | SILICON? CHECK. HSA: CHECK. ALL DONE? WRONG! | NOVEMBER 12, 2013