Rethinking Boundaries through HardwareSoftware Codesign for Productive PostMoore

Rethinking Boundaries through Hardware-Software Co-design for Productive Post-Moore Computing Samira Khan

My Background PMTest 2010 -2015 My core is in architecture 2015 -now XFTest Debugging and testing tools for persistent memory applications Software directed pre-execution for persistent memory Main Memory System DRAM, NVM Persistent Memory Processing-in-Memory Janus Profiling tool for PIM applications PIMProfiler

My Background PMTest 2010 -2015 -now XFTest Debugging and testing tools for persistent memory applications My core is in architecture Software directed pre-execution Main Memory System DRAM, NVM Persistent Memory Processing-in-Memory Janus for persistent memory tool Profiling for PIM applications PIMProfiler I design better software because I know hardware works

Looking Backward: Lessons From Past Hardware Transactional Has captivated. Memory architects for more than 20 years Intel i. APX 432 My favorite coolest chip design • MYTH: Cool hardware design is enough, everything will follow • REALITY: Has to be ultra fast to justify new programming models (5 X-10 X) • MYTH: Programmers want transparent hardware support • REALITY: What programmers really want • Freedom and flexibility • Deterministic operations • Debugging support

Systems In The Future SPECIALIZED CORES STORAGE BILLIONS OF EDGE DEVICES MEMORY CPU Architects’ Bubble Ld/St DRAM FILE I/O MEMORY WITH LOGIC PERSISTENT MEM SMART SSD SMART NIC 5

Systems In The Future SPECIALIZED CORES PROBLEM APPLICATION MEMORY WITH LOGIC PERSISTENT MEM Software OS RUNTIME SMART SSD ARCHITECTURE Hardware CIRCUITS Which functions to offload? What is the interface? Who provides data consistency? How to map and schedule data? How to enable computation? SMART NIC BILLIONS OF EDGE DEVICES Need to build the proper software stack 6

Systems In The Future PROBLEM APPLICATION oftware OS RUNTIME ARCHITECTURE ardware CIRCUITS • MYTH: Cool hardware design is enough, everything will follow Which functions • REALITY: Has to be ultra fast to justify a to offload? new programming model (5 X-10 X) What is the interface? Who provides data consistency? • How to map and schedule data? • How to enable computation? MYTH: Programmers want transparent hardware support REALITY: What programmers really want • Freedom and flexibility • Deterministic operations • Debugging support Absolutely no way to make it programmable and easy to use without software-hardware collaborative design 7

Rethinking Boundaries through Hardware-Software Co-design for Productive Post-Moore Computing Samira Khan