- Maia 200 ndiyo itsva Microsoft's in-house AI accelerator, yakagadzirirwa kufungidzira kwakakura ine PFLOPS dzinopfuura gumi muFP4 uye PFLOPS dzinosvika shanu muFP8.
- Chip iyi inotarisa nzvimbo dzedata rema cloud, ichipa simba remabasa akaita seCopilot, Microsoft Foundry, OpenAI's GPT-5.2 uye mapurojekiti kubva kuchikwata cheSuperintelligence.
- Yakavakirwa pamaitiro eTSMC e3 nm ane ma transistors anopfuura mabhiriyoni zana uye 216 GB yeHBM3E, Maia 200 ine chinangwa chekushanda zvakanaka uye kushanda zviri nani pamadhora ese ari maviri Maia 100 uye silicon inokwikwidzana nayo ine hyperscaler.
- Microsoft inoisa Maia 200 senzira yekuderedza kuvimba neNvidia uye kukwikwidzana zvakananga nemhuri dzeAmazon Trainium neGoogle dzeTPU.
The Microsoft Maia 200 yasvika semuchina mutsva weAI accelerator wekambani, chip yakagadzirwa kubva pasi kuti ishandise mabasa makuru ekufungidzira mugore. Pane kunanga zvishandiso zvemazuva ese zvevatengi, processor iyi yakagadzirirwa nzvimbo dzedata dzinorema uko mamodheru ane mabhiriyoni emaparameter anoshandiswa nguva dzose kupindura mibvunzo, kugadzira zvemukati uye simba remashandisirwo ebhizinesi.
Microsoft inotsanangura Maia 200 semhando ye "Silicon workhorse" yekufungidzira kweAI, yakagadzirirwa zvakananga kuti ishandise mamodheru akadzidziswa kare nekukurumidza, zvinobudirira uye pahukuru. Chip iyi iri kuiswa sedombo guru rekuvaka kweAzure yekambani, ichitsigira masevhisi akaita seMicrosoft 365 Copilot, Microsoft Foundry uye zvizvarwa zvitsva zvemamodheru makuru emitauro akagadzirwa mukati uye pamwe chete neOpenAI.
Magadzirirwo, process node uye raw performance
Pakati pemuchina mutsva wekumhanyisa zvinhu pane Maitiro ekugadzira TSMC 3-nanometer, ichibatanidza Maia 200 netekinoroji yepamusoro-soro iripo pari zvino yemakombiyuta anoshanda zvakanyanya. Ichishandisa iyi node, Microsoft yakaisa ma transistors anopfuura mabhiriyoni zana paMaia 200 die imwe chete mune zvimwe zvakaburitswa, nepo mamwe mapfupiso ehunyanzvi achinongedzera nhamba iri pedyo ne ma transistors mabhiriyoni zana nemakumi manaKunyangwe nhamba chaiyo ichisiyana zvichienderana nekwainobva, mishumo yese inoenderana nepfungwa yekuti iyi idhizaini yakakora uye yemazuva ano.
Panyaya yekushanda, Maia 200 yakanyatsogadzirwa kuti ishande zvakanaka mafomati edata asina kunyatsojeka izvo zvave zvakajairika mumabasa ekufungidzira emazuva ano. Microsoft inoshuma kuti accelerator inogona kudarika petaflops gumi (1015 floating point operations per second) kana ichishanda mu 4-bit precision (FP4), fomati yenhamba yakamanikidzwa inochinjana kurongeka kweimwe nzira kuti iwane simba rakawanda uye kushanda zvakanaka kwesimba. Mu 8-bit mode (FP8), chip inonzi inopa petaflops shanu, zvichiipa mukana wakakura wekuita mabasa anoda inference achiri kuda kunyatsorongeka kupfuura FP4.
Nhamba idzi hadzisi dzedzidzo: imwe node yeMaia 200 inotsanangurwa seinokwanisa kushandisa mamodheru makuru emazuva ano ukuwo ichisiya mukana wekuvaka kwakakura zvikuru kunogona kubuda munguva pfupi iri kutevera. Izvi zvinobatsira kutsanangura kuti sei Microsoft ichiratidza Maia 200 sesilicon ine hunyanzvi hwekutanga yekambani pakufungidzira kusvika pari zvino, yakanangana nebasa rakawanda iro ratova kusundira zvivakwa zviripo kusvika pamiganhu yaro.
Mukuenzanisa kwemukati nekweveruzhinji, Microsoft inoti Maia 200 yakawana kubudirira kweFP4 katatu kana kana yeTrainium accelerator yechizvarwa chechitatu cheAmazon, uye FP8 throughput inodarika TPU yechizvarwa chechinomwe cheGoogle mune mamwe ma metrics. Kunyange zvazvo kuenzanisa zvakananga chip-to-chip kuchigara kuchitsanangurwa, nhamba dzinoratidza kuti Microsoft ine chivimbo chakakwana chekutsanangura Maia 200 semutungamiri pakati pema accelerator akagadzirwa ne hyperscaler.
Chinonyanya kukosha ndechekuti kambani iyi haisi kungosimbisa kumhanya kwayo chete asi kushanda zvakanaka uye mariSekureva kweMicrosoft, Maia 200 inopa mashandiro ari nani ne30% padhora pane Maia 100 yechizvarwa chekutanga, kunyangwe ichishanda nesimba rekugadzira rinopisa zvakanyanya. Kune vanopa masevhisi egore nevatengi zvakafanana, rudzi irworwo rwehuwandu hwemitengo nekushanda ndicho chinhu chikuru pakusarudza kuti masevhisi eAI angawedzera sei.
Sisitimu yekurangarira uye kubata data
Chimwe chezvinhu zvinonyanya kushamisa nezveMaia 200 ndechekuti gadziriro yekurangarira, iyo yakagadzirirwa zvakananga mamodheru makuru eAI. Iyo accelerator inobatanidza 216 GB yeHBM3E high-bandwidth memory, yakabatana ne aggregate HBM bandwidth yeinenge 7 TB/s. Iyo bandwidth yakagadzirirwa kuchengetedza huremu hukuru hwemamodeli uye activations dzichiripo pasina zvipingamupinyi, chinhu chakakosha kuti pave nekufungidzira kwepamusoro-soro uye magadzirirwo emazuva ano ekuchengetedza zvinhu.
Kuwedzera kuHBM, chip iyi inosanganisira zvinhu zvakasiyana-siyana 272 MB yeSRAM iri pa-die, yakarongwa muhurongwa hwema multi-tier hunotsanangurwa neMicrosoft maererano neCluster-level SRAM (CSRAM) uye Tile-level SRAM (TSRAM). Dhizaini iyi ine layer inobvumira data rinowanika kakawanda kuti rigare pedyo nema compute units, zvichideredza latency uye zvichivandudza kushandiswa kwakanaka kwe bandwidth iripo. Chikamu chikuru chekutsvaga nekugadzira Maia 200 chinoratidzika kunge chakaiswa muhurongwa uhwu hwekurangarira, hunova hwakakosha pakuita kuti ma accelerator's compute units arambe akabatikana.
Magadzirirwo ese akagadzirirwa kuchengetedza ma parameter emuenzaniso uye data repakati remunharaunda zvakanyanya sezvinobvira, kuitira kuti basa rekufungidzira risafanirwe kugara richisvika kune imwe server kana network kuti riwane ruzivo rwunodiwa. Kutarisa kwenzvimbo kunoreva kuti ma accelerator mashoma anogona kudiwa kuti agadzirise modhi yakapihwa, zvichigona kuderedza huwandu hwehardware uye kuoma kwekushanda kwevatengi vari kushandisa mabasa makuru.
Kunze kwekuyeuka kuri package, Maia 200 inotsigira zvakakosha kuwedzera bandwidth pakati pemachipisi, yakashumwa kusvika ku2.8 TB/s. Kugona kwekubatanidza uku kwakagadzirirwa kubvumira ma accelerator akawanda kushanda sedziva rezvishandiso zvakabatana mukati me server kana rack, zvichivabatsira kugadzirisa mabasa makuru kana akaomarara ekufungidzira anogona kudarika kugona kwemudziyo mumwe chete.
Kunyangwe Microsoft isati yazivisa ruzivo rwakakwana nezve marongerwo chaiwo eraki yeMaia 200-enabled kana huwandu hwese hwepetaflop yesystem yakazara nevanhu, kambani iyi inoti ma accelerator aya achaiswa pamwe chete nedzimwe mhando dzeAI hardware. Muchokwadi, izvi zvinoreva kuti Maia 200 chikamu chezvivakwa zvakasiyana-siyana zveAzure, kushanda pamwe chete nemaGPU nezvimwe zvinomhanyisa zvinhu pane kuzvitsiva zvachose.
Kutarisa kwekufungidzira uye mabasa chaiwo epasi rose
Kusiyana nema "accelerators" ekudzidzisa akagadzirirwa kuvaka mamodheru kubva pakutanga, Maia 200 yakanangwa pa fungidziro — maitiro ekumhanyisa mamodheru kana angodzidziswaChikamu ichi chekushandiswa kweAI chinowanzo tonga mari yekushanda iri kuramba ichishandiswa, sezvo mamodheru anogona kubvunzwa mamiriyoni kana mabhiriyoni enguva mushure mekunge kudzidziswa kwapera. Sezvo masangano achitora masisitimu akaomarara, kudiwa kwekugadzirisa fungidziro kwave kukurumidzirwa.
Nekutarisa paFP4 neFP8, Maia 200 ine chinangwa chekuwana mukana wakanaka pakati kurongeka, kumhanya uye kushandiswa kwesimbaFP4, kunyange isina kunyatsojeka, inobvumira bandwidth yakakwira zvikuru uye inogona kukodzera mamodheru anoshivirira quantization pasina kuora kwakanyanya mumhando yekubuda. FP8 inopa sarudzo yakanyatsojeka kune zviitiko apo kururama kunoramba kuri kunetseka kukuru, uku ichiri kupa kushanda kuri nani kupfuura mafomati ekare e16-bit kana 32-bit floating point.
Microsoft yatanga kushandisa Maia 200 kushandira Modhi yeOpenAI yeGPT-5.2 nedzimwe LLM dzepamusoro kuburikidza neAzure. Iyo accelerator iriwo muMicrosoft Foundry, chikuva chekuvaka mhinduro dzeAI dzakagadzirirwa, uye inopa simba remabasa eMicrosoft 365 Copilot, mubatsiri wekambani anotarisa kubudiriro. Mumabasa aya, Maia 200 yakagadzirirwa kubatsira kuvandudza nguva yekupindura, kuwedzera throughput uye kugonesa maficha akaomarara pasina kuwedzera kwakaenzana kwemutengo.
Mukati, Maia 200 inototsigira mapurojekiti kubva ku Chikwata cheMicrosoft Superintelligence, chikamu chine chinangwa chikuru chinoongorora masisitimu eAI epamusoro-soro. Kambani iyi iri kuvhura ongororo yekutanga yeMaia 200 software development kit kune vadyidzani vakasarudzwa, kusanganisira vadzidzi, vagadziri, maAI lab uye vanopa mipiro kumamodeli e-open-source. Kuongororwa kweSDK uku kwakagadzirirwa kubatsira zvikwata zvekunze kugadzirisa mabasa avo ku chip uye kuongorora kuti magadzirirwo ayo angashandiswa sei zvakanakisisa.
Kubva pamaonero emushandisi, kuunzwa kweMaia 200 hakuzochinji pakarepo nzira iyo vanhu vazhinji vanobata nayo masevhisi eAI zuva nezuva. Zvisinei, sezvo mabasa ekufungidzira akawanda achichinja kuenda kune accelerator itsva, Vashandisi vekupedzisira vanogona kuona mhinduro dzinokurumidza, ruzivo rwakapfava uye kugona kutsva muzvishandiso zvakaita seCopilot, maapplication akavakirwa paAzure OpenAI nemamwe masevhisi eMicrosoft anovimba nepfungwa dzakasimba dziri kumashure kwezviitiko.
Kuiswa munzvimbo dzedata dzeAzure uye kuwanikwa kwadzo
Maia 200 yatove kushandiswa mukugadzirwa mukati meMicrosoft's cloud infrastructure. Kambani iyi inosimbisa kuti Kuendeswa kwekutanga kuri kuitika munharaunda yayo yeUS Central Azure, pedyo neDes Moines, Iowa. Kuwedzerwa kwemabasa kuri kurongwa, neUS West 3 munzvimbo yePhoenix, Arizona ichinzi ndeimwe yenzvimbo dzinotevera kugamuchira Maia 200, uye nzvimbo dzakawanda dzinotarisirwa nekufamba kwenguva sezvo nzvimbo dzekupa zvinhu dzichiwedzera.
Parizvino, Maia 200 iri kushandiswa chete sechikamu che Masisitimu edata anotungamirirwa neMicrosoftKusiyana nemaGPU kana mamwe ma accelerator echitatu anotengeswa zvakananga kuvatengi, chip iyi haisi kupihwa sechinhu chakasiyana chekutenga nekuisa munzvimbo dziri panzvimbo yacho. Pane kudaro, masangano achawana kugona kweMaia 200 zvisina kunanga kuburikidza nemasevhisi eAzure, maAPI nemapuratifomu anotungamirwa.
Microsoft yakaratidza kuti pachava ne kuwanikwa kwakakura kwevatengi munguva inotevera, zvichiratidza kuti vatengi vakawanda vachakwanisa kunanga zvakananga Maia 200-backed instances kana masevhisi ari muAzure sezvo ecosystem ichikura. Kana izvozvo zvichizopedzisira zvawedzera kune masisitimu akazvimirira kunze kweMicrosoft's cloud ichiri mubvunzo wakavhurika, asi kusimbiswa kuripo parizvino kwakanyanya pane zvinopihwa nehosted.
Panyaya yemabhenefiti anoshanda, vatengi veAzure OpenAI, masayendisiti, vagadziri uye makambani makuru vanofanira kuona zvishoma nezvishoma kuburitswa kwepamusoro uye kunonoka kuri nani pamabasa akaomarara ekufungidzira sezvo Maia 200 ichipararira zvakanyanya. Izvi zvinogona kuve zvakakosha kune mashandisirwo akadai semamodeli emari, kutsvaga, maautonomous agents, kana masisitimu ekugadzira enguva chaiyo anoda kumhanya uye kugara kwenguva refu.
Kunze kwevashandisi vemabhizinesi, Microsoft inoratidza kuti zvivakwa zvinoshandisa simba reMaia 200 zvinogona kutsigira zvakakura mapurojekiti ekutsvagisa munzvimbo dzakadai sekugadzirisa mamiriro ekunze, biosciences nekemesitiri. Minda iyi inonyanya kutsamira pakuenzanisa kwakadzama uye mamodheru akaomarara, ayo anogona kubatsirwa zvakanyanya nema accelerators akagadzirirwa kukurumidza kufungidzira pamaseti makuru edata.
Kushanda zvakanaka, mathermals uye nzvimbo yezvakatipoteredza
Sezvo kutariswa kuri kuwedzera pamusoro pemitengo yezvakatipoteredza yeAI infrastructure, Microsoft iri kuronga Maia 200 senzira yekudzivirira matambudziko. chinokurumidza kushanda zvakanaka pachinzvimbo pekugadzira machipisi anodya simba rose. Chip iyi ine simba rekugadzira rinosvika 750 watts (TDP), iro rakaderera zvakanyanya pane mamwe maGPU anoshandisa simba zvakanyanya pamusika, akadai seNvidia's Blackwell B300 Ultra, anonzi ane simba rinenge 1,400 watts.
Kunyangwe TDP yakakwira kudai kana tichienzanisa neMaia 100, iyo itsva accelerator inonzi inopa mashandiro ari nani ne30% pamadhora kupfuura chizvarwa chekare. Kuvandudzwa ikoko kunokonzerwa nekuchinja kwenzvimbo yekugadzira zvinhu yepamusoro, kugadziriswa kwezvivakwa uye nzira yekurangarira inoshanda zviri nani. Mishumo yapfuura yakaratidza kuti nepo Maia 100 yakagadzirwa zvine hunyanzvi kuti ive neenvelopu ye700-watts, Microsoft yakaiganhurira kuita mawatts angangoita mazana mashanu, zvichiratidza kuti Maia 200 inogonawo kushanda pasi pedanho rayo redzidziso zvichienderana nesarudzo dzekuisa.
Kutumira mameseji kweMicrosoft nezve Maia 200 kwakabatana nekuedza kukuru kwemakambani kuratidza kuti ndizvo zvazviri vanoziva nezvemhedzisiro yenzvimbo huru dzedata pamusoro penharaunda dzakapoteredza nenzvimbo dzakatipoteredza. Vakuru vakuru vakataura pachena kuti budiriro yeAI inofanira kuratidza mabhenefiti chaiwo kune nzanga kuitira kuti irambe ichitsigirwa neveruzhinji uye kudzivirira izvo vamwe vanoti zvinogona kukonzera kusawirirana neAI.
Mukutaura kwakajeka, kushanda zvakanaka kwewatt imwe neimwe uye padhora rimwe nerimwe kunogona kubatsira kuderedza huwandu hwezvinomhanyisa zvinodiwa kutsigira nhanho yebasa rakatarwa, zvichideredza kushandiswa kwesimba uye hunyanzvi hwehardware. Izvi zvinogona kubatsira vanogadzira michina yekuwedzera simba kupindura kunetseka nezvekushandiswa kwesimba, kushandiswa kwemvura pakutonhodza uye zviyero zvakakura zvekuchengetedza simba.
Kunyangwe Microsoft isati yagovana ruzivo rwakazara rwakadai se kudhonza kwesimba rose uye mashandiro eraki yeMaia 200 yakazara nevanhu, kusimbiswa kwekushanda zvakanaka kwekufungidzira kunoratidza kuti chip iyi yakagadzirirwa kuita basa guru mukuchengetedza mari yekushanda kwenguva refu yemasevhisi eAI ichikwanisika, mune zvemari uye zvakatipoteredza.
Nzvimbo yemakwikwi: Trainium, TPU uye Nvidia's ecosystem
Maia 200 inosvika panguva iyo makambani makuru ekupa mabasa egore ari kumhanya-mhanya kugadzira maAI accelerator avo kuitira kuderedza kuvimba nevatengesi vekunze, kunyanya Nvidia. Google's TPU line neAmazon's Trainium series zvatova zvakakosha mune izvi, zvichipa dzimwe nzira dzakabatanidzwa zvakanyanya mumapuratifomu avo egore.
NeMaia 200, Microsoft iri kuisa chip yayo pachena se inokwikwidzana zvakananga neTrainium3 yeAmazon neTPU yechizvarwa chechinomwe cheGoogleKambani iyi inoti accelerator yayo haingopfuuri Trainium3 muFP4 throughput chete nekatatu kusvika kana, asi inopawo mashandiro eFP8 anodarika TPU v7 ukuwo Trainium3 inowedzera kaviri mune mamwe ma FP8 metrics.
Kukwanisa kurangarira uye bandwidth zvakare chikamu chenyaya iyi yemakwikwi. Maia 200's 216 GB yeHBM3E memory uye 7 TB/s yebandwidth Kusiyana nezvinodiwa neveruzhinji zveTrainium3 (144 GB uye 4.9 TB/s) uye TPU v7 (192 GB uye 7.4 TB/s). Pakukwira, bandwidth yekubatanidza ye2.8 TB/s yakakumbirwa yeMaia 200 inoenzaniswa neanosvika 2.56 TB/s yeTrainium3 uye angangoita 1.2 TB/s yeTPU v7.
Zvisinei, mamiriro emakwikwi akaoma kupfuura nhamba dzisina kugadzirwa chete. Nvidia inoramba iri mutengesi mukuru yekudzidziswa kweAI uye, kazhinji, kufungidzira zvakare, nekuda kwekukura kweCUDA-based software ecosystem yayo uye huwandu hwayo hwakakura hwehardware. Kunyangwe Microsoft, Amazon neGoogle dzichiwedzera maportfolio avo esilicon, vachiri kushanda pamwe chete neNvidia pakugadzira maarchitecture matsva uye masisitimu akabatanidzwa.
Zvechokwadi, Amazon yanga ichishanda neNvidia kuti isanganise Trainium4 iri kuuya pamwe chete nematekinoroji akaita seNVLink 6 neMGX reference designs. Saizvozvowo, Microsoft inoramba ichiisa maGPU eNvidia munzvimbo dzayo dzedata, ichiisa Maia 200 sechikamu chenzvimbo dzakasiyana pane kutsiva pasi rose.
Kune vatengi, mibvunzo mikuru ichapindurwa mutengo wese wekuva muridzi uye kubatanidzwa kwesoftware. Zviri nyore sei kuti mabasa ekushanda akwanise kufambiswa kana kugadziriswa paMaia 200, uye kuti mitengo inoenzaniswa sei neNvidia- kana TPU-based instances, zvichachinja zvakanyanya kuti ndeapi accelerator stack yavanosarudza. Kusimbisa kweMicrosoft pakushanda-pa-dollar uye pa-watt kunoratidza kuti kufunga uku ndiko kwakakosha muhurongwa hwayo.
Kubva kuMaia 100 kusvika kuMaia 200: shanduko uye mazano ekugadzira nzira
Maia 200 inovakwa pa nheyo yakaiswa neMaia 100, iyo Microsoft yakaunza muna 2023 seyekutanga AI accelerator yayo. Kunyange hazvo kambani iyi isina kupa zviyero zvakafanana pakati pezvizvarwa zviviri izvi, chip itsva iyi iri pachena kuti yakanangana nekushanda kwepamusoro, ine zvinangwa zvekugadzira zvine simba uye zviwanikwa zvekuchengetedza ndangariro zvakawedzerwa.
Mishumo inoratidza kuti Maia 200, ine zita rekodhi mukati "Braga", yakatora nguva yakareba yekugadzirwa uye kuiswa kwesoftware kupfuura zvaitarisirwa pakutanga. Tarisiro yekutanga yakaita kuti isvike muna 2025, zvichida kutopfuura mamwe maGPU echizvarwa chinotevera aitengeswa. Pakupedzisira, nguva yekuburitswa kwayo yakachinja, asi Microsoft ikozvino inoita seyakatsunga kusimbisa kuvapo kwayo semutambi wesilicon akakosha mumusika weAI wepamusoro-soro.
Kunyangwe TDP yakakwira inosvika mawatts mazana manomwe nemakumi mashanu (750 watts), kambani iyi inoti kushanda zvakanaka kweMaia 200 mune zveupfumi kuri nani zvikuru kupfuura Maia 100, nekuda kwekuvandudzwa kwemaitiro emagetsi uye kugadziriswa kwezvivakwa. Kuwedzerwa kwebhajeti remagetsi kunoonekwa sechibvumirano chinogamuchirwa chekubvisa huwandu hwakawanda hwemagetsi hunogara kwenguva refu uye kuderedzwa kwemitengo yekufungidzira pabasa rega rega.
Tichitarisa mberi, mishumo yeindasitiri inoratidza kuti Zvishandiso zveMicrosoft zvemangwana zvinogona kunge zvakagadzirwa pamaitiro eIntel 18A, zvichiratidza kuti kambani iri kutsvaga hukama hwakawanda hwekugadzira michina yekuvaka kuti igadzire magadzirirwo anotevera. Kuchinja-chinja uku kunogona kupa Microsoft mukana wekugadzirisa mashandiro, kuwanikwa kwezvinhu uye mitengo yezvinhu muzvizvarwa zvinotevera.
Parizvino, Maia 200 inomiririra danho guru murwendo rwakakura rweMicrosoft kubva kukambani inonyanya kushandisa software kuenda kukambani inogadzira zvinhu zvakakosha zvehardware zvinotsigira zvinopihwa neAI. Kuburitswa kwayo kungangove kuchachinja kuti Microsoft inogona sei kuwedzera masevhisi akadai seCopilot, Foundry neAzure OpenAI mumakore anotevera.
Tichibatanidza zvinhu izvi zvese pamwe chete, Maia 200 inobuda se chinokurumidza kufungidzira chakavakirwa pachinangwa iyo inosanganisa FP4 neFP8 throughput yakakwira, subsystem yakakura yekurangarira uye kubatanidzwa kwakatarisana negore muzvivakwa zvakasiyana-siyana zveAzure. Kunyange hazvo ruzivo rwakadzama rwekuvaka uye zviyero zvizere zvehurongwa zvichiri kuziviswa, chip iyi inoratidza zvakajeka danho rakakosha mukuedza kweMicrosoft kudzora zvakawanda zveAI stack yayo, kutarisira mari uye kukwikwidzana nevamwe vanonyanya kushandisa hyperscalers mumakwikwi ekupa simba kune mamwe maapplication makuru eAI.
