NSITEXE unveils A New Product AI accelerator "ML041", realizes high power efficiency
November 10, 2021 -- NSITEXE, Inc. (headquartered in Minato Ward, Tokyo, Japan; President and CEO: Yukihide Niimi; hereinafter “NSITEXE”) commences sales of AI accelerator “ML041”, which realizes high power efficiency.
In order to realize a mobility society that connects people and cars, a smart city that connects people and cities, and a Cyber Physical System that reproduces the real world on computers in the industrial systems, the efficient execution of AI processing by edge devices with severe power and cost constraints has become an important issue in applying AI to a wide range of fields.
In conventional neural network processing, since the intermediate data of each layer is stored in an external memory with high power consumption, power efficiency is lowered. The AI accelerator “ML041”, developed by NSITEXE, divides input data (“Tiling”), and fuses the processing of multiple layers for each divided input data (“Layer fusion”), reducing the load/store transactions of intermediate data to the external memory and improving power efficiency. This will enable neural networks such as VGG 16, MobileNet and ResNet to run at 12 TOPS/W power efficiency (when implementing 7 nm generation SoCs).
TOPS: Tera Operations Per Second
“ML041” also offers a configuration with built-in diagnostic circuitry to detect hardware random failures, enabling AI to be applied to safety-critical systems without the need for additional external diagnostic circuitry.
Details of “ML041” will be presented at Cadence CONNECT: HLS Design & Verification Seminar 2021 on Thursday, November 18, 2021.
https://www.nsitexe.com/archives/1450
Hideki Sugimoto, CTO, NSITEXE, Inc.
At NSITEXE, development efforts have been focused on three pillars – versatility, efficiency, and functional safety – with massive future trends in mind. “ML041” uses tiling & layer fusion technology to achieve high power efficiency.
“ML041” provides automobile-related customers around the world with a basic development platform for a mobility society that creates a sustainable world, including clean energy.
Building on this effort for AI accelerators as well as on our innovative processor technology, we support smart mobility and MaaS, thus contributing to changing the world in ways that enrich people’s lives.
About NSITEXE
NSITEXE is an IP vendor, established in 2017 as a spin-off from DENSO Corp., specializing in the development of advanced processors. The company develops RISC-V based processor IPs that support functional safety. Highly efficient, high-quality semiconductor IPs support a wide range of applications, contributing to the evolution of the next-generation semiconductor technology.
|
Related News
- LeapMind's Ultra Low-Power AI accelerator IP "Efficiera" Achieved industry-leading power efficiency of 107.8 TOPS/W
- LeapMind Unveils "Efficiera", the New Ultra Low Power AI Inference Accelerator IP
- Gyrfalcon's New Chip Raises Bar (12.6 TOPS/W) on High Performance Edge AI with Lower Power Use
- Cadence Launches New Tensilica DNA 100 Processor IP Delivering Industry-Leading Performance and Power Efficiency for On-Device AI Applications
- DMP Released Next-Generation AI Accelerator IP "ZIA A3000 V2" - Industry-leading PPA efficiency to propel the future of edge AI
Breaking News
- GUC Joins Arm Total Design Ecosystem to Strengthen ASIC Design Services
- QuickLogic Announces $6.575 Million Contract Award for its Strategic Radiation Hardened Program
- Micon Global and Silvaco Announce New Partnership
- Arm loses out in Qualcomm court case, wants a re-trial
- Jury is out in the Arm vs Qualcomm trial
Most Popular
- GUC Joins Arm Total Design Ecosystem to Strengthen ASIC Design Services
- Alphawave Semi Scales UCIe™ to 64 Gbps Enabling >20 Tbps/mm Bandwidth Density for Die-to-Die Chiplet Connectivity
- VeriSilicon unveils next-generation high-performance Vitality architecture GPU IP series
- SafeNet Reiterates Guidance and Clarifies Revenue Assumptions
- Micon Global and Silvaco Announce New Partnership
E-mail This Article | Printer-Friendly Page |