ARM Architecture

aiMotive ships first aiWare4 NPU manufacturing RTL


Budapest, Hungary — June 2, 2022 – aiMotive, one of many world’s main modular automated driving expertise suppliers, introduced at present that it has shipped remaining production-validated RTL of the newest era of its ultra-high effectivity NPU aiWare4 to guide prospects. The newest aiWare4 RTL shipped delivers as much as 5x the efficiency of the earlier era aiWare3 NPUs, whereas utilizing lower than 2x the silicon space. This demonstrates the distinctive scalability, PPA and efficiency per mm2 of aiWare4, whereas extending the function set and operational “candy spot” for high-efficiency CNN acceleration.

“Our aiWare group has relentlessly refined our manufacturing validation processes to allow us to ship buyer configurations at report velocity to full automotive high quality for aiWare4,” says Márton Fehér, SVP {hardware} engineering at aiMotive. “Because of our subtle wavefront processing making full use of our new WFRAM expertise, plus many different architectural advances from aiWare3, we now have been in a position to obtain distinctive PPA for our lead prospects with out compromising our management in high-efficiency execution as much as 95% of probably the most demanding automotive inference CNN workloads”.

To realize extraordinarily demanding PPA constraints from prospects, aiMotive was in a position to fine-tune the precise function set of the aiWare4 manufacturing RTL to greatest meet prospects’ necessities. Making full use of the bodily tile-based structure and dataflow methodologies, the aiWare group demonstrated clock speeds for the manufacturing RTL of as much as 1.3GHz over the total automotive AEC-Q100 Grade 2 temperature vary on a 14nm course of. The aiWare4 {hardware} IP has been assessed externally as appropriate for certification to ASIL-B or larger as an SEooC.

The aiWare4 NPU scales from 1 to 256 TOPS and is supported by an exceptionally complete SDK that includes extremely correct offline efficiency estimation, enabling prospects to precisely estimate and fine-tune their CNN workloads to inside 5% of ultimate silicon efficiency previous to first silicon.

aiWare4 {hardware} IP is out there now for licensing.

For extra particulars, click on right here.

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button