AMD stellt Vega-Architektur auf CES vor

SUNNYVALE, Calif. Jan. 5, 2017 Today AMD (NASDAQ: AMD) unvei짯led preli짯mi짯na짯ry details of its forth짯co짯ming GPU archi짯tec짯tu짯re, Vega. Con짯cei짯ved and exe짯cu짯ted over 5 years, Vega archi짯tec짯tu짯re enables new pos짯si짯bi짯li짯ties in PC gam짯ing, pro짯fes짯sio짯nal design and machi짯ne intel짯li짯gence that tra짯di짯tio짯nal GPU archi짯tec짯tures have not been able to address effec짯tively. Data-inten짯si짯ve workloads are beco짯ming the new nor짯mal, and the par짯al짯lel natu짯re of the GPU lends its짯elf ide짯al짯ly to tack짯ling them. Howe짯ver, pro짯ces짯sing the짯se huge new data짯sets requi짯res fast access to mas짯si짯ve amounts of memo짯ry. The Vega architecture셲 revo짯lu짯tio짯na짯ry memo짯ry sub짯sys짯tem enables GPUs to address very lar짯ge data sets spread across a mix of memo짯ry types. The high-band짯width cache con짯trol짯ler in Vega-based GPUs can access on-packa짯ge cache and off-packa짯ge memo짯ries in a fle짯xi짯ble, pro짯gramma짯ble fashion using fine-grai짯ned data movement.

It is incre짯di짯ble to see GPUs being used to sol짯ve giga짯byte-sca짯le data pro짯blems in gam짯ing to exa짯byte-sca짯le data pro짯blems in machi짯ne intel짯li짯gence. We desi짯gned the Vega archi짯tec짯tu짯re to build on this abili짯ty, with the fle짯xi짯bi짯li짯ty to address the extra짯or짯di짯na짯ry breadth of pro짯blems GPUs will be sol짯ving not only today but also five years from now. Our high-band짯width cache is a pivo짯tal dis짯rup짯ti짯on that has the poten짯ti짯al to impact the who짯le GPU mar짯ket, said Raja Kodu짯ri, seni짯or vice pre짯si짯dent and chief archi짯tect, Rade짯on Tech짯no짯lo짯gies Group, AMD.

High짯lights of the Vega GPU architecture셲 advance짯ments include:

  • The world셲 most advan짯ced GPU memo짯ry archi짯tec짯tu짯re: The Vega archi짯tec짯tu짯re enables a new memo짯ry hier짯ar짯chy for GPUs. This radi짯cal new approach comes in the form of a new high-band짯width cache and its con짯trol짯ler. The cache fea짯tures lea짯ding-edge HBM2 tech짯no짯lo짯gy which is capa짯ble of trans짯fer짯ring tera짯bytes of data every second, doubling the band짯width-per-pin over the pre짯vious gene짯ra짯ti짯on HBM tech짯no짯lo짯gy. HBM2 also enables much grea짯ter capa짯ci짯ty at less than half the foot짯print of GDDR5 memo짯ry. Vega archi짯tec짯tu짯re is opti짯mi짯zed for strea짯ming very lar짯ge data짯sets and can work with a varie짯ty of memo짯ry types with up to 512TB of vir짯tu짯al address space.
  • Next-gene짯ra짯ti짯on geo짯me짯try pipe짯line: Today셲 games and pro짯fes짯sio짯nal appli짯ca짯ti짯ons make use of incre짯di짯bly com짯plex geo짯me짯try enab짯led by the extra짯or짯di짯na짯ry increase in the reso짯lu짯ti짯ons of data acqui짯si짯ti짯on devices.  The hundreds of mil짯li짯ons of poly짯gons in any given frame have mes짯hes so den짯se that the짯re are often many poly짯gons being ren짯de짯red per pixel. Vega셲 next-gene짯ra짯ti짯on geo짯me짯try pipe짯line enables the pro짯gramm짯er to extra짯ct incre짯di짯ble effi짯ci짯en짯cy in pro짯ces짯sing this com짯plex geo짯me짯try, while also deli짯ve짯ring more than 200% of the through짯put-per-clock over pre짯vious Rade짯on archi짯tec짯tures.[i] It also fea짯tures impro짯ved load-balan짯cing with an intel짯li짯gent workload dis짯tri짯bu짯tor to deli짯ver con짯sis짯tent performance.
  • Next-gene짯ra짯ti짯on com짯pu짯te engi짯ne: At the core of the Vega archi짯tec짯tu짯re is a new, next-gene짯ra짯ti짯on com짯pu짯te engi짯ne built on fle짯xi짯ble com짯pu짯te units that can natively pro짯cess 8멳it, 16-bit, 32-bit or 64-bit ope짯ra짯ti짯ons in each clock cycle.[ii] The짯se com짯pu짯te units are opti짯mi짯zed to attain signi짯fi짯cant짯ly hig짯her fre짯quen짯ci짯es than pre짯vious gene짯ra짯ti짯ons and their sup짯port of varia짯ble data짯ty짯pes makes the archi짯tec짯tu짯re high짯ly ver짯sa짯ti짯le across workloads.
  • Advan짯ced pixel engi짯ne: The new Vega pixel engi짯ne employs a Draw Stream Bin짯ning Ras짯te짯ri짯zer, desi짯gned to impro짯ve per짯for짯mance and power effi짯ci짯en짯cy. It allows for 쐄etch once, sha짯de once of pixels through the use of a smart on-chip bin cache and ear짯ly cul짯ling of pixels invi짯si짯ble in a final sce짯ne. Vega셲 pixel engi짯ne is now a cli짯ent of the onboard L2 cache, enab짯ling con짯sidera짯ble over짯head reduc짯tion for gra짯phics workloads which per짯form fre짯quent read-after-wri짯te operations.

GPU pro짯ducts based on the Vega archi짯tec짯tu짯re are expec짯ted to ship in the first half of 2017.

Sup짯port짯ing Resources

About AMD

For more than 45 years AMD has dri짯ven inno짯va짯ti짯on in high-per짯for짯mance com짯pu짯ting, gra짯phics, and visua짯liza짯ti짯on tech짯no짯lo짯gies the buil짯ding blocks for gam짯ing, immersi짯ve plat짯forms, and the dat짯a짯cen짯ter. Hundreds of mil짯li짯ons of con짯su짯mers, lea짯ding For짯tu짯ne 500 busi짯nesses, and cut짯ting-edge sci짯en짯ti짯fic rese짯arch faci짯li짯ties around the world rely on AMD tech짯no짯lo짯gy dai짯ly to impro짯ve how they live, work, and play. AMD employees around the world are focu짯sed on buil짯ding gre짯at pro짯ducts that push the boun짯da짯ries of what is pos짯si짯ble. For more infor짯ma짯ti짯on about how AMD is enab짯ling today and inspi짯ring tomor짯row, visit the AMD (NASDAQ: AMD) web짯site, blog, Face짯book and Twit짯ter pages.

AMD, the AMD Arrow logo, Rade짯on and com짯bi짯na짯ti짯ons the짯reof, are trade짯marks of Advan짯ced Micro Devices, Inc. Other names are for infor짯ma짯tio짯nal pur짯po짯ses only and may be trade짯marks of their respec짯ti짯ve owners.

Cau짯tio짯na짯ry Statement

This press release con짯ta짯ins for짯ward-loo짯king state짯ments con짯cer짯ning Advan짯ced Micro Devices, Inc. (AMD) inclu짯ding the fea짯tures, func짯tion짯a짯li짯ty and expec짯ta짯ti짯ons regar짯ding AMD셲 Vega GPU archi짯tec짯tu짯re and the timing and avai짯la짯bi짯li짯ty of GPU pro짯ducts based on AMD셲 Vega archi짯tec짯tu짯re, which are made pur짯su짯ant to the Safe Har짯bor pro짯vi짯si짯ons of the Pri짯va짯te Secu짯ri짯ties Liti짯ga짯ti짯on Reform Act of 1995. For짯ward-loo짯king state짯ments are com짯mon짯ly iden짯ti짯fied by words such as 쐗ould, 쐌ay, 쐃xpects, 쐀elie짯ves, 쐏lans, 쐇ntends, 쐏ro짯jects and other terms with simi짯lar mea짯ning. Inves짯tors are cau짯tio짯ned that the for짯ward-loo짯king state짯ments in this docu짯ment are based on cur짯rent beliefs, assump짯ti짯ons and expec짯ta짯ti짯ons, speak only as of the date of this press release and invol짯ve risks and uncer짯tain짯ties that could cau짯se actu짯al results to dif짯fer mate짯ri짯al짯ly from cur짯rent expec짯ta짯ti짯ons. Such state짯ments are sub짯ject to cer짯tain known and unknown risks and uncer짯tain짯ties, many of which are dif짯fi짯cult to pre짯dict and gene짯ral짯ly bey짯ond AMD셲 con짯trol, that could cau짯se actu짯al results and other future events to dif짯fer mate짯ri짯al짯ly from tho짯se expres짯sed in, or impli짯ed or pro짯jec짯ted by, the for짯ward-loo짯king infor짯ma짯ti짯on and state짯ments. Mate짯ri짯al fac짯tors that could cau짯se actu짯al results to dif짯fer mate짯ri짯al짯ly from cur짯rent expec짯ta짯ti짯ons include, wit짯hout limi짯ta짯ti짯on, the fol짯lo짯wing: Intel Corporation셲 domi짯nan짯ce of the micro짯pro짯ces짯sor mar짯ket and its aggres짯si짯ve busi짯ness prac짯ti짯ces may limit AMD셲 abili짯ty to com짯pe짯te effec짯tively; AMD is par짯ty to a wafer sup짯p짯ly agree짯ment with GLOBALFOUNDRIES Inc. (GF) with obli짯ga짯ti짯ons to manu짯fac짯tu짯re pro짯ducts at GF with cer짯tain excep짯ti짯ons. If GF is not able to satis짯fy AMD셲 manu짯fac짯tu짯ring requi짯re짯ments, its busi짯ness could be adver짯se짯ly impac짯ted; AMD reli짯es on third par짯ties to manu짯fac짯tu짯re its pro짯ducts, and if they are unable to do so on a time짯ly basis in suf짯fi짯ci짯ent quan짯ti짯ties and using com짯pe짯ti짯ti짯ve tech짯no짯lo짯gies, AMD셲 busi짯ness could be mate짯ri짯al짯ly adver짯se짯ly affec짯ted; fail짯ure to achie짯ve expec짯ted manu짯fac짯tu짯ring yields for AMD셲 pro짯ducts could nega짯tively impact its finan짯cial results; the suc짯cess of AMD셲 busi짯ness is depen짯dent upon its abili짯ty to intro짯du짯ce pro짯ducts on a time짯ly basis with fea짯tures and per짯for짯mance levels that pro짯vi짯de value to its cus짯to짯mers while sup짯port짯ing and coin짯ci짯ding with signi짯fi짯cant indus짯try tran짯si짯ti짯ons; if AMD can짯not gene짯ra짯te suf짯fi짯ci짯ent reve짯nue and ope짯ra짯ting cash flow or obtain exter짯nal finan짯cing, it may face a cash short짯fall and be unable to make all of its plan짯ned invest짯ments in rese짯arch and deve짯lo짯p짯ment or other stra짯te짯gic invest짯ments; the loss of a signi짯fi짯cant cus짯to짯mer may have a mate짯ri짯al adver짯se effect on AMD; AMD셲 receipt of reve짯nue from its semi-cus짯tom SoC pro짯ducts is depen짯dent upon its tech짯no짯lo짯gy being desi짯gned into third-par짯ty pro짯ducts and the suc짯cess of tho짯se pro짯ducts; glo짯bal eco짯no짯mic uncer짯tain짯ty may adver짯se짯ly impact AMD셲 busi짯ness and ope짯ra짯ting results; the mar짯kets in which AMD셲 pro짯ducts are sold are high짯ly com짯pe짯ti짯ti짯ve; AMD may not be able to gene짯ra짯te suf짯fi짯ci짯ent cash to ser짯vice its debt obli짯ga짯ti짯ons or meet its working capi짯tal requi짯re짯ments; AMD has a sub짯stan짯ti짯al amount of indeb짯ted짯ness which could adver짯se짯ly affect its finan짯cial posi짯ti짯on and pre짯vent it from imple짯men짯ting its stra짯tegy or ful짯fil짯ling its con짯trac짯tu짯al obli짯ga짯ti짯ons; the agree짯ments gover짯ning AMD셲 notes and the secu짯red revol짯ving line of cre짯dit (Secu짯red Revol짯ving Line of Cre짯dit) impo짯se rest짯ric짯tions on AMD that may adver짯se짯ly affect its abili짯ty to ope짯ra짯te its busi짯ness; uncer짯tain짯ties invol짯ving the orde짯ring and ship짯ment of AMD셲 pro짯ducts could mate짯ri짯al짯ly adver짯se짯ly affect it; the demand for AMD셲 pro짯ducts depends in part on the mar짯ket con짯di짯ti짯ons in the indus짯tries into which they are sold. Fluc짯tua짯tions in demand for AMD셲 pro짯ducts or a mar짯ket decli짯ne in any of the짯se indus짯tries could have a mate짯ri짯al adver짯se effect on its results of ope짯ra짯ti짯ons; AMD셲 abili짯ty to design and intro짯du짯ce new pro짯ducts in a time짯ly man짯ner is depen짯dent upon third-par짯ty intellec짯tu짯al pro짯per짯ty; AMD depends on third-par짯ty com짯pa짯nies for the design, manu짯fac짯tu짯re and sup짯p짯ly of mother짯boards, soft짯ware and other com짯pu짯ter plat짯form com짯pon짯ents to sup짯port its busi짯ness; if AMD loses Micro짯soft Corporation셲 sup짯port for its pro짯ducts or other soft짯ware ven짯dors do not design and deve짯lop soft짯ware to run on AMD셲 pro짯ducts, its abili짯ty to sell its pro짯ducts could be mate짯ri짯al짯ly adver짯se짯ly affec짯ted; and AMD셲 reli짯ance on third-par짯ty dis짯tri짯bu짯tors and Add-in-Board part짯ners sub짯jects it to cer짯tain risks. Inves짯tors are urged to review in detail the risks and uncer짯tain짯ties in AMD셲 Secu짯ri짯ties and Exch짯an짯ge Com짯mis짯si짯on filings, inclu짯ding but not limi짯t짯ed to AMD셲 Quar짯ter짯ly Report on Form 10멡 for the quar짯ter ended Sep짯tem짯ber 24, 2016.

[i] Data based on AMD Engi짯nee짯ring design of Vega.  Rade짯on R9 Fury X has 4 geo짯me짯try engi짯nes and a peak of 4 poly짯gons per clock. Vega is desi짯gned to hand짯le up to 11 poly짯gons per clock with 4 geo짯me짯try engi짯nes.  This repres짯ents an increase of 2.6x.  VG3

[ii] Dis짯crete AMD Rade짯on꽓 and Fire짯Pro꽓 GPUs based on the Gra짯phics Core Next archi짯tec짯tu짯re con짯sist of mul짯ti짯ple dis짯crete exe짯cu짯ti짯on engi짯nes known as a Com짯pu짯te Unit (CU). Each CU con짯ta짯ins 64 shaders (쏶tream Pro짯ces짯sors) working tog짯e짯ther. GD-78