A new network-on-chip (NoC) IP aims to dramatically accelerate chip development by introducing artificial intelligence (AI)-driven automation and reducing wire length to lower power use in system-on-chip (SoC) interconnect design. Arteris, which calls its newly introduced FlexGen interconnect IP a smart NoC, claims to deliver a 10x productivity boost, shortening design iterations from weeks to days.
Modern chips—connected by billions of wires—are ever-expanding with growing complexity. Modern SoCs have 5 to 20+ unique NoC instances, and each instance can require 5-10 iterations. As a result, SoC design complexity has surpassed manual human capabilities, which calls for smarter NoC automation.
“In SoC interconnect, while technology has advanced to new levels, a lot of work is still done in manual mode,” said Michal Siwinski, CMO of Arteris. FlexGen accelerates chip design by shortening and reducing iterations from weeks to days for greater efficiency.
“While FlexGen is still using the tried-and-tested NoC IP technology as basic building blocks, it automates the existing infrastructure by employing AI technology,” said Andy Nightingale, VP of product management and marketing at Arteris. “With FlexGen, we automate the NoC IP generation to reduce the manual work while opening high-quality configurations that rival or surpass the manual designs.”
Figure 1 A FlexNoC manual interconnect (above) is shown for an ADAS chip, while an automated FlexGen interconnect (below) accelerates this chip design by up to 10x. Source: Arteris
According to Nightingale, it enhances engineering efficiency by 3x while delivering expert-quality results with optimized routing and reduced congestion. Dream Chip Technologies, a supplier of advanced driver assistance systems (ADAS) silicon solutions, acknowledges reducing design iterations from weeks to days while using FlexGen in its Zukimo 1.1 automotive ADAS chip design.
“FlexGen’s automated NoC IP generation allows us to create floorplan adaptive topologies with complex automotive traffic requirements within minutes,” said Jens Benndorf, GM at Dream Chip Technologies. “That enabled rapid experimentation to find design sweet spots and to respond quickly to floorplan changes with almost push-button timing closure.”
Shorter wire length
With AI comes a compute performance explosion, and as a result, the complexity of interconnects is going to exponential levels in SoC designs, leading to a huge explosion in the number of wires. FlexGen claims to reduce wire length by up to 30% to improve chip or chiplet power efficiency.
“We are also tackling the big problem of wire length in modern SoC designs,” said Nightingale. “As the gate count size reduces, it inevitably leads to dynamic power issues due to massive data traffic across wires.” By reducing wire length, FlexGen interconnect IP can reduce overall system power and thus help heating problems caused by the energy density of moving massive amounts of data across SoC interconnects.
Figure 2 FlexNoC manual interconnect (above) is shown with the best performance, while automated FlexGen (below) significantly reduces the interconnect wire length. Source: Arteris
Siwinski added that the number of gates doesn’t matter at smaller nodes. “Power from wire length kills you, so we reduce wire length to reduce overall power, performance, and area (PPA) in SoC designs.” That’s crucial as SoCs scale and become more powerful to meet the demands of applications like AI, autonomous driving, and cloud computing.
FlexGen is processor agnostic and supports Arm, RISC-V, and x86 processors. Moreover, its IP generation is highly repeatable to facilitate incremental design.
Related Content
- SoC Interconnect: Don’t DIY!
- What is the future for Network-on-Chip?
- Why verification matters in network-on-chip (NoC) design
- SoC design: When is a network-on-chip (NoC) not enough
- Network-on-chip (NoC) interconnect topologies explained
The post SoC interconnect automates processes, reduces wire length appeared first on EDN.