AI Computer Center yawng gaw, pat da ai-Loop hka katsi ai ladat hpe jai lang ai kun?

Ga hkum kaba ni hpe sharin hkam la ai lam, generative AI inference hte high-density GPU clusters ni lawan ladan jat wa ai majaw, htunghking nbung katsi ai hte hpaw da ai nbung katsi ai nta ni gaw, dai ni na AI computing center ni a thermal management ra kadawn ai lam ni hpe n lu galaw sai.

 

Ya aten hta, madung AI supercomputing hub kaba ni, sumwi dut ai AI npawt nhpang ni htehpaga hpung shi hkum nan-gap da ai GPU hpung niyawng la aiclosed-loop liquid cooling hpe madang madi shadaw ai ladat hku nna jai lang ai, kaji ai-n-gun n law ai AI computing room ni gaw nbung katsi ai hpe galai shai ai masing hku nna naw tawn da mai ai. Closed-loop cooling gaw AI computing masa ni a matu ra kadawn ai madung npawt nhpang tai wa sai rai nna, shi a hkum tsup ai-masha law law ra sharawng ai lam gaw hpaga lam hta n mai nhtang wa ai.

 

 

 

 

Hpa majaw pat da ai-loop system ni uphkang ai a madung lam gawAI katsi ai masu gachip heat load grai rawt jat wa ai hta. NVIDIA Blackwell, AMD MI300X zawn re ai prat nnan AI chip ni hte mungdan kata na n-gun ja ai GPU ni gaw, chip langai sha hta 700W jan na n-gun atsam nga ai, n-gun langai sha gaw 80–150kW hta jan ai n-gun hpe htaw lu ai, far ex. nbung-katsi ai hka htung ni.

 

Nbung katsi ai gaw nbung n-gun hpe htawt sit na matu nbung n-gun hpe kam hpa ai, nbung htawt sit ai atsam n law ai hte fan power kaba lang ai, dai gaw data center a PUE yawng hpe 1.3–1.5 du hkra shadut ya ai.

 

Closed-loop liquid cooling gaw, chips ni hta jung da ai katsi ai plate ni de katsi ai hpe hkrak sha hkawm sa na matu, pat da ai pipeline ni hpe lang nna, katsi ai hpe 85% jan htawt sit lu ai, katsi ai n-gun lang ai lam hpe 30%–40% shayawm ya nna, PUE hpe 1.1 a lawu de shadut ya ai, hpang jahtum na wargreem}me from 45 NVIDIA gaw PUE hpe 1.08 de pyi jahkrat kau lu ai.

 

Ndai n-gun-hkye la ai akyu gaw AI npawt nhpang kaba ni a bungli galaw ai manu hpe hkrak shayawm ya ai:50MW kaba ai AI jahpancenter gaw closed-loop cooling de galai mat ai hpang shaning shagu na wan jahpu US dollar wan 4 jan hpe hkye la lu ai.

 

 

 

 

 

Hka n-gun atsam gaw AI center ni hta closed-loop system ni hpe tup hkrak hkap la na matu kaga ahkyak ai lam langai mi rai nga ai. Open cooling tower ni gaw hka n-gun hpe shapraw kau na matu hka n-gun hpe kam hpa nna, shaning shagu megawatt computing power langai hta hka san gallon wan law law hpe jai lang ai majaw, hka-n law ai ginra ni hta grai sawng ai tara masa ni hpe pat shingdang ai lam ni hpe shabyin ya ai.

 

Closed-loop katsi ai gaw, laksan pat da ai shingjawng ai lam lahkawng hpe shabyin ya ai: server katsi ai plate ni hpe matut mahkai ya ai kata na circuit hte .pat da ai katsi ai nta nishinggan grup yin kaw na tsep kawp tsan gang nga ai, coolant hpe kalang mi bang nna, prat tup bai jai lang ai, shingra tara hku galaw ai shaloi zero evaporation loss ngu na daram rai nga ai.

 

OpenAI Stargate AI sharin hpawng hte Microsoft prat nnan na computing park ni lawm ai hpaga lam ni yawng gaw, hpaw da ai -loop htingnu hpe hkap la nna, shaning shagu hka lang ai hpe lusha seng kaji langai a madang de shayawm kau ai, "AI hka hkanghkyi" ni a hka lang ai lam hpe madung hku nna hparan ya ai.

 

Dai majaw, n-gun n rawng ai hka hte shinggan na hka hpe garan nna, nbung hta nga ai nhprang wan, hka htung hte n san seng ai hka ni hpe galoi mung n hkrum hkra ai majaw, server cold plates ni a scaling hte corrosion hpe koi lu ai, server a service life hpe grau galu shangun ai.

 

 

 

Closed-loop cooling gaw AI computing a matu ahkyak ai ngang grin ai bungli masa ni hpe mung la sa ya ai. Model kaba sharin hkam la na matu 7×24h n hkring n sa ngang grin ai nbung katsi ai hpe hkang ra ai; temperature galai shai ai lam gaw GPU frequency hpe shayawm kau ya nna, sharin hkam la ai lam hpe pat shingdang kau ya nna, computing hpe grai sum mat shangun na re.

 

Closed-loop system ni a hkum tsup ai hku pat da ai loop hpe hpaji rawng ai frequency galai shai ai pump ni, nbung katsi ai hpe chye na ai module ni hte automatic bypass control ni hte hpaw da ai majaw, shaning ting ±0.5℃lapran hta nbung katsi ai hpe galai shai shangun lu ai.

 

Spring, Autumn hte n-gun n law ai shana ni hta, system gaw shi hkum shi 1 de galai mat wa ai.shingra tara manu jahpu n jaw ra ai katsi ai ladat, n-gun lang ai lam hpe grau shayawm lu na matu chiller ni hpe pat kau ai; n-gun ja ai-katsi ai hka lum aten hta, ga nhpu n hkra hkra, ngang grin ai katsi ai hpe shapraw kau lu na matu, karum shingtau hka lum ai jak hte rau jawm galaw ai,nhpu hte hka lim aishinggan grup yin hta byin nga ai.

 

Dai hte nhtan shai ai hku nna, hpaw da ai katsi ai nta ni gaw, n san seng ai pipeline ni hpe pat shingdang chye ai majaw, aten tup chemical dosing hte hka hpe galai ra ai, nbung katsi ai lam n kaja ai majaw, aten galu-n ​​hkring n sa ai AI sharin hpawng bungli ni hpe madi shadaw ya lu na n re.

 

 

 

Raitim, AI- hte seng ai computer room yawng gaw, pat da ai-loop katsi ai ladat ni hpe tup hkrak n jai lang ai. Kaji ai-madang AI computer nodes ni, offline n-gun n law ai sawk sagawn jak ni hten law ai-hkum hkrang chyam dinglik nta ni20kW npu na rack power hte nbung katsi ai (sh) asan sha hpaw da ai katsi ai nta ni hpe naw lang nna, shawng nnan na gumhpraw bang ai hpe hkang lu ai.

 

Ndai zawn re ai mabyin masa ni hta n-gun n law ai, matut manoi bungli galaw ai aten kadun ai hte hka jai lang ai lam hpe n-gun n law ai majaw, closed-loop system ni gaw grai law ai gawgap ai hte gram lajang ai manu ni hpe shabyin ya na re.

 

Raitim, GPU cabinet 10 jan nga ai lapran tsang hte kaba ai AI computer center ni a matu, hpaji rawng ai computer center ni a matu mungdan ting na hpaga lam madang ni gaw, closed-loop liquid cooling hpe ra sharawng ai cold source scheme hku nna asan sha hkam la da sai, supercomputing masing nnan ni hpe configure galaw na matu ra nga ai.pat da ai-loop hpe katsi shangun aitowers ni hpe madang madi shadaw ai arung arai ni hku nna jai lang ai.

 

 

 

 

 

Industrial iteration a ningmu hku yu yang, closed-loop cooling gaw mungkan ting na AI npawt nhpang gawgap ai lam a langai sha re ai hkrang shapraw ai madang tai wa sai. NVIDIA a DSX AI factory reference design, sumwi dut ai ni a madung htingnu gawgap ai lam gaw, GPU computing platform yawng a matu 100% tup hpaw da ai-loop liquid cooling hpe masat da ya ai, server ni a katsi ai fan ni hpe tsep kawp jahkring kau ai.

 

Dinghku katasinpraw-data-sinpraw-hpaji jaw ai intelligent computing hubs, n-gun AI sharin hpawng nnan ni hte tsi mawan ni hpe hti ai shara ni yawng gaw, n-gun hpe hkye la ai, hka hpe hkye la ai hte n-gun kaba ai -density computing ngu ai yaw shada lam lahkawng hpe shadik shatup ai hku nna, hpaw da ai hka n-gun hpe jai lang ai.

 

AI chip power hpe matut manoi jat wa ai hte maren, 2019 ning hpe n-gun jaw ra ai hku jai lang na matu threshold hpe mu lu ai.pat da ai-loop masa nigrau nna jahkrat kau na rai nna, lapran tsang hte kaba ai computing masa ni hta galai shai ai nbung katsi ai masing ni hpe loi loi hte yeng seng kau na re.

 

 

 

Tsun ga nga yang, AI jak kaji ni gaw nbung hpe katsi shangun ai hpe aten kadun laman lang mai ai, raitim, n-gun kaba ai AI computer center ni yawng, super training clusters ni hte hyperscale hpaji rawng ai computer hub ni yawng hpe lang ra ai.pat da ai-loop ntsin hpe katsi shangun aisystem ni hpe pat da ai katsi ai nta ni hte madi shadaw ya ai.

 

 

Ndai hparan ladat gaw, nbung n-gun kaba ai, hka grai law hkra jai lang ai hte AI computing a n-gun n-gun n-rawng ai nbung hpe hkang lu ai core pain point masum hpe hparan ya lu nna, artificial intelligence industry kaba-scale rawt jat galu kaba wa na matu, galu kaba ai nbung hpe hkang lu ai lam langai sha tai wa sai.