Centaur Releases In-Depth Analysis from The Linley Group for World’s First x86 Processor with AI Coprocessor Technology

Detail¬≠ed Tech¬≠ni¬≠cal Dis¬≠clo¬≠sure Fol¬≠lows Suc¬≠cessful Demons¬≠tra¬≠ti¬≠on at ISC East¬≠Tra¬≠de Show

Aus¬≠tin, Texas ‚ÄďDecem¬≠ber 9, 2019 ‚Äď Cen¬≠taur Tech¬≠no¬≠lo¬≠gy today reve¬≠a¬≠led in-depth infor¬≠ma¬≠ti¬≠on abou¬≠tits new pro¬≠ces¬≠sor-design tech¬≠no¬≠lo¬≠gy for inte¬≠gra¬≠ting high-per¬≠for¬≠mance x86 CPUs with a spe¬≠cia¬≠li¬≠zed copro¬≠ces¬≠sor opti¬≠mi¬≠zed for arti¬≠fi¬≠ci¬≠al intel¬≠li¬≠gence (AI) acce¬≠le¬≠ra¬≠ti¬≠on. On its website(www.centtech.com), Cen¬≠taur pro¬≠vi¬≠des a new inde¬≠pen¬≠dent¬≠re¬≠port¬≠from¬≠The Lin¬≠ley Group, the industry‚Äôs lea¬≠ding aut¬≠ho¬≠ri¬≠ty on micro¬≠pro¬≠ces¬≠sor tech¬≠no¬≠lo¬≠gy and publishers of Micro¬≠pro¬≠ces¬≠sor Report.The Lin¬≠ley Group review¬≠ed Centaur‚Äôs detail¬≠ed design docu¬≠ments andinterviewedCentaur‚Äôs CPU and AI archi¬≠tect¬≠sto¬≠sup¬≠port the ana¬≠ly¬≠sis of both Centaur‚Äôs newest x86 micro¬≠ar¬≠chi¬≠tec¬≠tu¬≠re and the AI coprocessordesign.

‚ÄúCen¬≠taur is gal¬≠lo¬≠ping back into the x86 mar¬≠ket with an inno¬≠va¬≠ti¬≠ve pro¬≠ces¬≠sor design that¬≠com¬≠bi¬≠nes eight high-per¬≠for¬≠mance CPUs with a cus¬≠tom deep-lear¬≠ning acce¬≠le¬≠ra¬≠tor (DLA). The¬≠com¬≠pa¬≠ny is the first to announ¬≠ce a ser¬≠ver-pro¬≠ces¬≠sor design that inte¬≠gra¬≠tes a DLA. The new¬≠ac¬≠ce¬≠le¬≠ra¬≠tor, cal¬≠led Nco¬≠re, deli¬≠vers bet¬≠ter neu¬≠ral-net¬≠work per¬≠for¬≠mance than even the most¬≠powerful Xeon, but wit¬≠hout the high cost of an exter¬≠nal GPU card,‚ÄĚ sta¬≠ted¬≠Lin¬≠ley Gwen¬≠nap, Edi¬≠tor-in-Chief, Micro¬≠pro¬≠ces¬≠sor Report.

The Lin¬≠ley Group refe¬≠ren¬≠ced cer¬≠ti¬≠fied MLPerf1benchmark (Pre¬≠view) scores¬≠to compa¬≠re Centaur‚Äôs AI per¬≠for¬≠mance to high-end x86 CPU¬≠co¬≠re¬≠s¬≠from theleadingx86 ven¬≠dor. Based on MLPerfs¬≠cores, Centaur‚Äôs AI-copro¬≠ces¬≠sor¬≠in¬≠fe¬≠rence per¬≠for¬≠mance¬≠is¬≠com¬≠pa¬≠ra¬≠ble to 23of Intel‚Äôs world-class x86 cores that now sup¬≠port¬≠512-bit vec¬≠tor neu¬≠ral net¬≠work ins¬≠truc¬≠tions (VNNI)2.Centaur‚Äôs AI copro¬≠ces¬≠sor uses an archi¬≠tec¬≠tu¬≠ral¬≠ly¬≠si¬≠mi¬≠lar sin¬≠gle-ins¬≠truc¬≠tion-mul¬≠ti¬≠ple-data (SIMD) approach¬≠as VNNI, but crun¬≠ches 32,768bitsin a sin¬≠gle clock cycleu¬≠sing a 16MB memo¬≠ry with 20 terabytes/sec of band¬≠width. Moreo¬≠ver, by off¬≠loa¬≠ding infe¬≠rence pro¬≠ces¬≠sing to a spe¬≠cia¬≠li¬≠zed copro¬≠ces¬≠sor, the x86 CPU cores remain available for other, more gene¬≠ral-pur¬≠po¬≠se tasks. Appli¬≠ca¬≠ti¬≠on deve¬≠lo¬≠pers can inno¬≠va¬≠te new algo¬≠rith¬≠ms that take advan¬≠ta¬≠ge of the unpar¬≠al¬≠le¬≠led infe¬≠rence laten¬≠cy enab¬≠led by Centaur‚Äôs AI per¬≠for¬≠mance and tight inte¬≠gra¬≠ti¬≠on with x86 CPUs.

Atten¬≠de¬≠es at the ISC East trade show in NYC¬≠saw Centaur‚Äôs new tech¬≠no¬≠lo¬≠gy up clo¬≠se¬≠for the first time. The demo¬≠show¬≠ca¬≠sedvi¬≠deo ana¬≠ly¬≠tic¬≠su¬≠s¬≠ing Centaur‚Äôs refe¬≠rence sys¬≠tem¬≠wi¬≠thx86-based¬≠net¬≠work-video-recor¬≠ding (NVR) soft¬≠ware from Qvis Labs. In addi¬≠ti¬≠on to conventional,real-time object detection/classification, Cen¬≠taur¬≠was the only ven¬≠dor at the show tohigh¬≠light lea¬≠ding-edge appli¬≠ca¬≠ti¬≠ons¬≠such as seman¬≠tic seg¬≠men¬≠ta¬≠ti¬≠on (pixel-level image clas¬≠si¬≠fi¬≠ca¬≠ti¬≠on) and a new tech¬≠ni¬≠que for human pose estimation(‚Äústickfigures‚ÄĚ).Centaurisfocusedon impro¬≠ving the hard¬≠ware price/performance and soft¬≠ware pro¬≠duc¬≠ti¬≠vi¬≠ty for plat¬≠forms to sup¬≠port this next wave ofre¬≠se¬≠arch appli¬≠ca¬≠ti¬≠ons and¬≠speed deploy¬≠ment into new ser¬≠ver-class products.About Cen¬≠taur Tech¬≠no¬≠lo¬≠gy Aus¬≠tin, Texas-based Cen¬≠taur Tech¬≠no¬≠lo¬≠gy is a small group of very talen¬≠ted engi¬≠neers that have been desig¬≠ning AI acce¬≠le¬≠ra¬≠tor tech¬≠no¬≠lo¬≠gy for high-per¬≠for¬≠mance, low-cost x86-com¬≠pa¬≠ti¬≠ble micro¬≠pro¬≠ces¬≠sors. Over the past 24 years, Cen¬≠taur Tech¬≠no¬≠lo¬≠gy has ship¬≠ped 26 dif¬≠fe¬≠rent x86-based designs with mil¬≠li¬≠ons of units sold. More infor¬≠ma¬≠ti¬≠on is at www.centtech.com.

[1] MLPerf v0.5 Infe­rence Closed/Preview audi­ted sub­mis­si­on, Sept. 2019. MLPerf name and logo are trade­marks. See www.mlperf.org for more information.

[2] MLPerf Inf‚ÄĎ0.5‚Äď23. Dual Intel¬ģ Xeon¬ģ Pla¬≠ti¬≠num 9282 (112 total cores).MobileNet-V1 and Res¬≠Net-50 on Intel x86coreswith VNNI. Intel and Xeon are trade¬≠marks of Intel Corporation.