1 / 13

The Essential Shift Towards Specialized Systems in Technology

Specialized systems have become the norm in today's technological landscape, driven by an unquenchable demand for enhanced capabilities in fields like online services, big data, and the Internet of Things (IoT). With a significant scaling gap that no new technology can bridge alone, it falls upon hardware and software systems engineers to innovate. Specialization offers remarkable advancements in energy efficiency and performance, allowing systems to amortize instruction overhead and focus on locality. Embracing specialization is not just practical; it's inevitable for progress in computing and beyond.

lynsey
Télécharger la présentation

The Essential Shift Towards Specialized Systems in Technology

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Specialized systems are Inevitable Already the norm Practical

  2. Specialized systems are Inevitable Already the norm Practical

  3. Infinite appetite for ever increasing capabilities (online services, big data, internet of things,…) No new technology is ready to save the day Our Challenge We, the HW and SW systems engineers, will save the day The good days of technology scaling Scaling wo/ techhelp: a 30x gap per decade CMOS New Technology Chip Capability (log) 80s 90s 00s 10s 20s 30s 40s

  4. The Potential of Specialization Energyper operation Regs Data Cache ALU Instr. Fetch Pipeline Control 26% 40% 7% 26% 2-10 x by amortizing instructions overhead (e.g., SIMD) 20-100x by also specializing locality (e.g., convolution engine) 100x would take 50 years @ 10% per year 100-1,000xby specializing everything (instructions, locality, ALUs)

  5. Specialized systems are Inevitable Already the norm Practical

  6. Cooking?

  7. Transportation?

  8. Computing? ARM A5, A5, A9, A15, A50 R4, R5, R7 M0, M1, M3, M4 SC100, SC300 Mai, Mai-T, Mai-V, Mai-DP … Intel Xeon & Atom core SSE for data-parallelism AES instructions i3, i5, i7 with/without GPU E3, E5, E7 … Qualcomm Snapdragon G G S S S S AVR, NIOS, Blackfin, Tigersharc, AT90, CY8C2, 68HC, FCR4, Tricore, Mico8, Micro32, TMS320, TX19A, IP3022, MicroBlaze, Xcore, eZ80, … S S S S G S S

  9. Specialized systems are Inevitable Already the norm Practical

  10. Specialization is Practical Reusable components Gates, ALUs, cores, GPUs, accelerators, interconnects, chips Automated design & compilation tools C/OpenCL/DSLs  gates, DSLs  accelerators Domain-specific accelerators Modems, GPUs, video/ML accelerators, NPUs, … Flexible hardware FPGAs & coarse-grain FPGAs, 2.5D + 3D integration, SoCs Just make the common case efficient Use a few programmable cores for the rest

  11. Specialized systems are Inevitable Already the norm Practical

More Related