NOCs: Past, Present and Future Ahmed Hemani Professor, Dept. Of Electronics, School of ICT, KTH, Stockholm Sweden Email: hemanikth.se 1 2 The Promise of NOC Enable Billion Gate ASICs Make them affordable 3 Exponentially Reduce the VLSI
"NOCs: Past, Present and Future Ahmed Hemani" is the property of its rightful owner. Permission is granted to
download and print the materials on this website for personal, non-commercial use only, and to display it
on your personal computer provided you do not modify the materials and that you retain all copyright
notices contained in the materials. By downloading content from our website, you accept the terms of this
agreement.
Presentation Transcript
01
NOCs: Past, Present and Future Ahmed Hemani
Professor, Dept. Of Electronics, School of ICT, KTH, Stockholm Sweden
Email: hemani@kth.se 1<br>
02
2 The Promise of NOC Enable Billion Gate ASICs Make them affordable<br>
03
3 Exponentially Reduce the VLSI Design Space Automation RTL Physical Design Platform Boolean Level Physical Design Platform Auotmated
One Time<br>
04
4 Synchronicity Synchoricity Syn + Choros (space in Greek)<br>
6 Composition by Abutment Regional / Local NOCs Without NOCs and Synchoricity
The compostion by abutment scheme
Would be impractical This policy is applied to
Clock, Reset and Power Grid as well<br>
07
7 NOC/CGRAs + Synchoricity Enables Composition By Abutment No further VLSI Engineering Needed<br>
08
8 SiLago Design Instances = Region Instances System
Controller Program
Memory Ethernet PLL/CGU
PMC Data
Memory Interrupt Ctrl Dynamically Reconfigurable
Resource Array
DRRA<br>
09
9 NOCs are key enablers of Composition by Abutment for DSE SiLago Physical Design Platform Design
Instance 1 Design
Instance 2 . . . VLSI Designs
Composition By Abutment SiLago Physical Design Platform Design Space Exploration<br>
10
10 One VLSI Design Instance generated by Abutting SiLago Blocks<br>
11
11 NOCs are multi-layer clustering agents for DSE SiLago Physical Design Platform Design
Instance 1 Design
Instance 2 Implementation
Alternative 1 Implementation
Alternative 2 . . . VLSI Designs
Composition By Abutment Coarse Grain
Reconfiguration SiLago Physical Design Platform Design Space Exploration . . . LSTM Model in MATLAB<br>
12
12 SiLago vs. Standard Cell Based ASIC Design Flow Area Energy 30-50% Area Overhead
10-30% Energy Overhead
In some cases SiLago beats ASIC in energy<br>
13
SiLago Enables Automation of Highly Efficient Custom Design 3 kW 12 Watts eBrain II@KTH 13 SiLago
Enables design automation of such high performance custom computing machines Biological Plausible Model of Cortex: ~1 PetaFlops,
40 TBs Synaptic Weights at 140 TBs/s bandwidth<br>
14
The Impact Software Centric / GPU + Processor
Based Designs Hardware Centric
SiLago Based Designs 1000 X 1000 X<br>
