Co-Design 2013 Summary Exascale needs new architectures due to slowing of Dennard scaling (since...

6

Transcript of Co-Design 2013 Summary Exascale needs new architectures due to slowing of Dennard scaling (since...

Page 1: Co-Design 2013 Summary Exascale needs new architectures due to slowing of Dennard scaling (since 2004), multi/many core limits New programming models,
Page 2: Co-Design 2013 Summary Exascale needs new architectures due to slowing of Dennard scaling (since 2004), multi/many core limits New programming models,

Co-Design 2013 Summary• Exascale needs new architectures due to slowing of Dennard scaling

(since 2004), multi/many core limits• New programming models, new approaches to OS, new ways to

manage power, new ways to make reliable systems from unreliable parts

• Opportunities to integrate needs of big data as well as modeling and simulation into future Architectures

• Investments needed in basic R+D, Arch, HW and SW research, will use acquisitions to drive integrated systems

• Significant investment needed in new applications• Big challenges 20pj per (f)op, viable vendors, scale and reliability• Co-Design needed to match HW-ALG-SW• Huge challenges in data analysis and visualization need to change

our paradigm, automation, in situ, uncertainty representation and visualizaiton, UQ

Page 3: Co-Design 2013 Summary Exascale needs new architectures due to slowing of Dennard scaling (since 2004), multi/many core limits New programming models,

Worldwide Trends

• Many systems moving towards some kind of heterogeneous nodes– Combining throughput and latency computing elements

• Some well performing systems are still homogeneous, but this seems to be less likely in the future– BGQ, K, TH1A, TH-2, Tsubame, Piz Daint, Titan, etc.

• Precise tradeoffs on node balance not well understood, key bottlenecks, ,power, applications partitioning, etc.

• Networks also not yet converging, probably disruption• MPI+X, but no dominate alternative programming model• Big drivers, climate, disaster, neuroscience, materials, geoscience, energy,

fusion, fission, life sciences, drugs, aircraft, transportation, etc. shift towards applications citizens can understand and support

• Regional differences in applications focus (China, Japan, Europe, US)• Regional differences is level of strategic intent

– USA and Japan are closest to a Exascale program– China investing heavily in machines and research but needs overarching program and

more capability in applicaitons– Europe advancing on applications, coordination, building EU capability

Page 4: Co-Design 2013 Summary Exascale needs new architectures due to slowing of Dennard scaling (since 2004), multi/many core limits New programming models,

US programon decision inFY2014

Japan programon decision in 2013/2014

Page 5: Co-Design 2013 Summary Exascale needs new architectures due to slowing of Dennard scaling (since 2004), multi/many core limits New programming models,

Exascale in 2020?

100PF in 2015?

Systems with GPUs and ManyCore

“Traditional” Exascale Applications

Increasing importance of data

Overarching case for Exascale Needed

Societal Impact

Beyond Traditional Supercomputing

Connections to Industry, Innovation

Page 6: Co-Design 2013 Summary Exascale needs new architectures due to slowing of Dennard scaling (since 2004), multi/many core limits New programming models,

Common Challenges

• Can we develop a sustainable architecture suitable for HPC, Data, Internet and mobile?

• Integration of throughput accelerators, true vector pipelines, deep scalar processors, etc.

• Integrated networks (NICs, Switches, Routers)• Alg specific processing elements/logic– What should these be?

• What role is co-design playing in the development of these elements?

• Interest and basis for international collaboration!