SANTA CLARA, Calif. –May 25, 2022 – Ayar Labs, the lea­der in chip-to-chip opti­cal con­nec­ti­vi­ty, is deve­lo­ping with NVIDIA ground­brea­king arti­fi­ci­al intel­li­gence (AI) infra­struc­tu­re based on opti­cal I/O tech­no­lo­gy to meet future demands of AI and high per­for­mance com­pu­ting (HPC) workloads. The col­la­bo­ra­ti­on will focus on inte­gra­ting Ayar Labs’ tech­no­lo­gy to deve­lop sca­le-out archi­tec­tures enab­led by high-band­width, low-laten­cy and ultra-low-power opti­cal-based inter­con­nects for future NVIDIA pro­ducts. Tog­e­ther, the com­pa­nies plan to acce­le­ra­te the deve­lo­p­ment and adop­ti­on of opti­cal I/O tech­no­lo­gy to sup­port the explo­si­ve growth of AI and machi­ne lear­ning (ML) appli­ca­ti­ons and data volumes.

Mee­ting Future Per­for­mance and Power Requi­re­ments with Opti­cal I/O

Opti­cal I/O uni­que­ly chan­ges the per­for­mance and power tra­jec­to­ries of sys­tem designs by enab­ling com­pu­te, memo­ry and net­wor­king ASICs to com­mu­ni­ca­te with dra­ma­ti­cal­ly increased band­width, at lower laten­cy, over lon­ger distances and at a frac­tion of the power of exis­ting elec­tri­cal I/O solu­ti­ons. The tech­no­lo­gy is also foun­da­tio­nal to enab­ling emer­ging hete­ro­ge­neous com­pu­te sys­tems, disaggregated/pooled designs, and uni­fied memo­ry archi­tec­tures that are cri­ti­cal to acce­le­ra­ting future data cen­ter innovation.

“Today’s sta­te-of-the-art AI/ML trai­ning archi­tec­tures are limi­t­ed by cur­rent cop­per-based com­pu­te-to-com­pu­te inter­con­nects to build sca­le-out sys­tems for tomorrow’s requi­re­ments,” said Charles Wuisch­pard, CEO of Ayar Labs. “Our work with NVIDIA to deve­lop next-gene­ra­ti­on solu­ti­ons based on opti­cal I/O pro­vi­des the foun­da­ti­on for the next leap in AI capa­bi­li­ties to address the world’s most sophisti­ca­ted problems.”

Deli­ve­ring the ‘Next Million‑X’ Spee­dup for AI with Opti­cal Interconnect

“Over the past deca­de, NVI­DIA-acce­le­ra­ted com­pu­ting has deli­ver­ed a million‑X spee­dup in AI,” said Rob Ober, Chief Plat­form Archi­tect for Data Cen­ter Pro­ducts at NVIDIA. “The next million‑X will requi­re new, advan­ced tech­no­lo­gies like opti­cal I/O to sup­port the band­width, power and sca­le requi­re­ments of future AI and ML workloads and sys­tem architectures.”

As AI model sizes con­ti­nue to grow, by 2023 NVIDIA belie­ves that models will have 100 tril­li­on or more con­nec­tions – a 600X increase from 2021 – excee­ding the tech­ni­cal capa­bi­li­ties of exis­ting plat­forms. Tra­di­tio­nal elec­tri­cal-based inter­con­nects will reach their band­width limits, dri­ving lower appli­ca­ti­on per­for­mance, hig­her laten­cy and increased power con­sump­ti­on. New inter­con­nect solu­ti­ons and sys­tem archi­tec­tures are nee­ded to address the sca­le, per­for­mance and power demands of the next gene­ra­ti­on of AI. Ayar Labs’ col­la­bo­ra­ti­on with NVIDIA is focu­sed on addres­sing the­se future chal­lenges by deve­lo­ping next-gene­ra­ti­on archi­tec­tures with opti­cal I/O.

To learn more about Ayar Labs’ chip-to-chip opti­cal tech­no­lo­gy, plea­se visit: https://ayarlabs.com/

About Ayar Labs

Ayar Labs is dis­rupt­ing the tra­di­tio­nal per­for­mance, cost, and effi­ci­en­cy cur­ves of the semi­con­duc­tor and com­pu­ting indus­tries by dri­ving a 1000x impro­ve­ment in inter­con­nect band­width den­si­ty at 10x lower power.  Ayar Labs’ paten­ted approach uses indus­try stan­dard cost-effec­ti­ve sili­con pro­ces­sing tech­ni­ques to deve­lop high speed, high den­si­ty, low power opti­cal-based inter­con­nect “chip­lets” and lasers to replace tra­di­tio­nal elec­tri­cal based I/O. The com­pa­ny was foun­ded in 2015 and is fun­ded by a num­ber of dome­stic and inter­na­tio­nal ven­ture capi­tal firms as well as stra­te­gic inves­tors.  For more infor­ma­ti­on, visit www.ayarlabs.com.