I-GIGABYTE AI TOP ATOM vs NVIDIA DGX Spark: I-Desktop AI amandla ekhanda ukuya ekhanda

Impi phakathi kwama-mini desktop supercomputers iya ngokuya ithakazelisa: i-GIGABYTE AI TOP ATOM uma iqhathaniswa ne-NVIDIA DGX Spark . Zombili ziyingxenye ebalulekile yesikhathi esisha se-NVIDIA sokubala i-AI , kodwa ziza ngezindlela ezahlukene: enye njengeplatifomu yokubhekisela kanti enye njengomkhiqizo olungele ukusetshenziswa ngokushesha. Uma ubulokhu ufuna ukusebenza kwe-AI okungathi sína ngaphandle kokwenza ukusetha kwehadiwe, kuningi ongakucabangela lapha.

Sikhuluma ngamadivayisi angange-litha elilodwa ngobukhulu, amanani okusebenza athembisayo okwakungacatshangwa kuze kube muva nje kwi-PC encane kangaka: i-petaFLOP eyodwa yokusebenza kwe-AI ku-FP4 (1.000 TOPS) , inkumbulo ehlanganisiwe yesivinini esikhulu, kanye nesistimu yonke yesofthiwe ye-NVIDIA yokukhiqiza, ukucwenga, kanye nokuqagela amamodeli esizukulwane esilandelayo. Futhi yebo, konke lokhu ngokusetshenziswa kwamandla okuphansi kanye nokuxhumeka kwenethiwekhi okukwazi ukukala uma kudingeka.

Yini ngayinye: Inkundla ye-DGX Spark kanye ne-PC encane ye-GIGABYTE

Ukunikeza umongo othile: I-NVIDIA DGX Spark iyipulatifomu yereferensi yedeskithophu ye-NVIDIA (eyayaziwa ngokuthi i-Project DIGITS) futhi ingumphumela wokubambisana ne -MediaTek . Isebenza njengesisekelo lapho abakhiqizi abahlukahlukene bangakha khona ama-PC abo e-AI amancane. Ngaphakathi kwalolu hlaka, imikhiqizo efana ne-MSI ne-ASUS isivele iveze okunikezwayo kwayo, futhi manje i-GIGABYTE ijoyina lo mncintiswano ngoshintsho lwayo oluhlukile.

Lokho kuphenduka yi- GIGABYTE AI TOP ATOM , i-PC encane engu-1 litre ethatha isisekelo sobuchwepheshe se-DGX Spark bese iyipakisha ibe umkhiqizo wokugcina onemininingwane ecacile, izinsiza zayo, kanye nokutholakala kwezentengiselwano. Ngamanye amazwi, i-Spark ibeka indinganiso, kanti i-AI TOP ATOM iyincazelo eqondile ye-GIGABYTE yokuletha lawo mandla kuma-desktop onjiniyela, izikhungo zocwaningo, kanye namakilasi.

I-Architecture: NVIDIA GB10 Grace Blackwell ukusuka phezulu kuye phansi

Isici esivamile yi- NVIDIA Grace Blackwell GB10 superchip , i-SoC ehlanganisa i-CPU ne-GPU nememori ehlanganisiwe ye-bandwidth ephezulu. Ngasohlangothini lwe-CPU, sithola iprosesa ye -Armv9 enama-core angu-20 , enama-cores angu-10 e-Cortex-X925 asebenza kahle kakhulu kanye nama-cores angu-10 e-Cortex-A725 ukuze isebenze kahle. Le nhlanganisela inikeza ukuguquguquka kokubili emisebenzini yokucubungula kanye neyokuhlela , kanye nasemipayipini ye-AI eyinkimbinkimbi.

I-GPU isebenzisa ukwakheka kwe -Blackwell ngesizukulwane sakamuva se-NVIDIA sama-cores akhethekile: ama-Tensor cores esizukulwane sesi-5 okusheshisa i-AI kanye nama-RT cores esizukulwane sesi-4 okulandelela imisebe. Ngaphezu kwalokho, i-GB10 ihlanganisa i-NVENC eyodwa kanye ne-NVDEC eyodwa , ewusizo lapho isebenza ngemisakazo yevidiyo ekuhlaziyeni noma kuzinhlelo zokusebenza eziningi. Konke lokhu kubanjwe ndawonye yi -NVLink-C2C , "iglue" yesivinini esikhulu phakathi kwe-CPU ne-GPU edlula kude lokho okunikezwa yi-PCI Express 5.0 ngaphakathi kohlelo.

Ukusebenza kwe-AI namasayizi wamamodeli

Isibalo esiphawuleka kakhulu ukubalwa okunembe kancane kwe -AI yokukhiqiza : ama-TOPS ayi-1.000 ku-FP4, okulingana ne-petaFLOP eyi-1 . Empeleni, lokhu kusho ukuthi, ku-chassis encane, ungagijima futhi uthuthukise ngokwethembeka amamodeli olimi ne-multimodal yesimanje, wenze i-prototyping esheshayo, futhi usebenzise iziphetho ezidingayo ngaphandle kokuthembela efwini.

Ngokuphathelene nosayizi wemodeli, ulwazi oluvela emithonjeni alulona oluqondile futhi kungcono ukucabangela konke: ngakolunye uhlangothi, ukwesekwa kwendawo kuboniswa kumamodeli anamapharamitha afinyelela ku-200 billion kuyunithi eyodwa; ngakolunye uhlangothi, kuphinde kwashiwo ukuthi "kuthembisa ukwesekwa kwama-LLM anamapharamitha afinyelela ku-70 trillion kanye no-200.000 trillion (US)." Lokhu kwakhiwa kwakamuva kudlala ngokungaqondakali phakathi "kwama-billones" ngesiSpanishi kanye "namabhiliyoni" ngesiNgisi, kodwa umyalezo oyinhloko uhlala: umkhawulo wokusebenza wale mishini usemazingeni aphezulu kuma-LLM amanje.

Uma amayunithi amabili exhunyiwe, umkhawulo ohlanganisiwe ufinyelela cishe kumapharamitha ayizigidigidi ezingu-405.000 , okuvula umnyango wamamodeli amakhulu nakakhulu kanye nemithwalo yemisebenzi esatshalaliswe. Kulabo abahlola izimo ezinde, ama-RAG anezinkomba ezinkulu, noma amapayipi anezigaba eziningi, lelo khanda elengeziwe lenza umehluko omkhulu.

Lolu hlobo lokusebenza alugcini nje ngezindawo zokucwaninga: intuthuko, imfundo, ucwaningo, isayensi, amadolobha ahlakaniphile, kanye namarobhothi kuvela njengezinhlelo zokusebenza zemvelo. Ukuba ne-petaFLOP ye-AI "etafuleni" kusheshisa ukuhlolwa, kunciphisa izindleko zokusebenza, futhi kuvumela ukuphindaphinda okusheshayo ngaphandle kwemigqa yamafu.

Imemori ehlanganisiwe, isitoreji kanye nomkhawulokudonsa

Inkumbulo yesistimu iyiphuzu eliqinile: i-128 GB ye-LPDDR5X ebhasini elingu-256-bit , enomkhawulokudonsa ongaba ngu-273 GB/s. Ukuhlanganiswa phakathi kwe-CPU ne-GPU kwenza kube lula ukuhlela futhi kugwema amakhophi angafuneki, okubalulekile ekuphatheni ama-tensor amakhulu nama-batch ekuqeqeshweni okulula noma ekunqumeni okunesilinganiso esiphezulu.

Ukuze kugcinwe , i-GIGABYTE inikeza ama-SSD e -PCIe Gen5 afinyelela ku- 4 TB , yize ezinye izilungiselelo zezentengiselwano ziqala ku- 1 TB ku-PCIe 4.0 . Ngale komthamo, isihluthulelo ukuthi izithiyo ziyancishiswa lapho kulayishwa amamodeli amakhulu, noma lapho kukopishwa ngokushesha amafayela amakhulu , kushumekwa amaqoqo, noma amasethi edatha omsindo nombono, okugcina uhlelo lusebenza kahle.

Ukuxhumana kwangaphakathi kwe-NVLink-C2C kuphawula omunye umehluko obalulekile. Ngokunikeza i-bandwidth ephezulu kakhulu kune -PCIe 5.0 yokuxhumana kwe-CPU-GPU ngaphakathi kwe-SoC, kunciphisa ukubambezeleka futhi kusebenzisa inkumbulo ehlanganisiwe ezimweni ezizwela ukudluliselwa, okuyizo kanye lezo ezibusa ku-AI yokukhiqiza kanye ne-ML.

Ukuxhumana nokunwetshwa ukuze kuqhubekele phambili

Ngokuphathelene namachweba nokuxhumeka, ayinciphi. Njengokujwayelekile, idivayisi ifaka i-10GbE yokuxhumana ngezintambo okusebenzayo kakhulu, i-Wi-Fi 7 kanye ne-Bluetooth 5.3 yokuxhumana okungenantambo kwesimanje, kanye ne -HDMI 2.1a kanye namachweba amaningana e-USB 3.2 Gen 2x2 Type-C (kufaka phakathi eyodwa ene-PD IN yamandla). Umsindo weziteshi eziningi ukhishwa nge-HDMI, ngakho-ke uhlanganisa nezidingo ze-multimedia ngaphandle kwezesekeli ezengeziwe.

Ingxenye evumela lokhu kukhula yi- NVIDIA ConnectX-7 SmartNIC . Ixhumanisa izinhlelo ezimbili nezixhumanisi ezifika ku- 200 Gbps , okuvumela ukubalwa kwe-cluster ye-low-latency. Imithombo iphinde ichaze lesi sixhumi njenge- ConnectX-7 InfiniBand NIC , igcizelela ukugxila kwayo ekubambezelekeni okuncane kanye nomkhawulokudonsa oqhubekayo. Empeleni, lokhu kuvumela ukwanda komkhawulokudonsa kusuka ku-200 Mbps kuya cishe ku-405 Mbps lapho kuxhuma ama-node amabili.

Ngaphezu kokuhlanganisa amayunithi amabili, ngokombono wokusebenza, kungenzeka ukuhlanganisa amayunithi amaningi e-AI TOP ATOM bese uwaxhumanisa nge-ConnectX-7 ukuze uwasebenzise ngesikhathi esisodwa futhi wenze kube lula ukuvumelanisa amafolda enethiwekhi . Kuma-workshop ocwaningo noma amakilasi athuthukile, ukuba "namabhulokhi" e-modular kuwusizo kakhulu.

Isoftware, izinhlelo zokusebenza nezinsiza

Akukho ukuyekethisa maqondana nesofthiwe: idivayisi isekela i-NVIDIA DGX OS kanye ne-Ubuntu Linux , futhi ihlangana ne- NVIDIA AI software stack — isethi yamathuluzi, izinhlaka, kanye nemitapo yolwazi esetshenziselwa ukwakha nokusabalalisa amaphrojekthi e-AI asezingeni eliphezulu. Lokhu kunciphisa isithiyo sokungena kulabo asebejwayelene ne-CUDA ecosystem kanye namaqembu ezemfundo, futhi kuvumela, isibonelo, izibuyekezo ezijwayelekile ze-firmware ku-Linux .

I-GIGABYTE ifaka ne -AI TOP Utility yayo , ungqimba olwenza kube lula ukuhamba komsebenzi okuvamile: ukulandwa kwamamodeli, ukuqagela, i-RAG, kanye nokufunda komshini, konke kusuka ku-interface eqondile. Lo mnikelo ugcizelela ubumfihlo nokuphepha kwedatha yendawo , okwenza kube kuhle ezinhlanganweni okumele zilandele imithetho noma zikhethe ukugcina ulwazi lwazo olubucayi endlini.

Ukuklama ngokomzimba, ukusetshenziswa kanye nefomethi

Lonke uhlelo lungena ku-chassis engaba yilitha eli-1 . Luyikhava elincane ngokumangalisayo elinobukhulu obusemthethweni obungu- 150 × 150 × 50,5 mm kanye nesisindo esingaba ngu -1,2 kg . Amandla ahlinzekwa yi-adaptha yangaphandle engu-240W , eyanele ukusekela ukusebenza okuphezulu ngaphandle kwenkinga, futhi, njengoba imithombo ichaza, uhlelo luhambisana nezindawo zokungenisa ezijwayelekile zasekhaya ngaphandle kokudinga noma iyiphi ingqalasizinda ekhethekile.

Ngale kosayizi, i-GIGABYTE icabangele ukuthunyelwa okusheshayo nokulula: amayunithi amaningi angafakwa ngokoqobo futhi axhunywe ne-SmartNIC efanele, okusiza ekwakheni iqoqo elincane endaweni encane kakhulu. Kubasebenzisi abasekelwe kuphrojekthi, ikhono lokwengeza nokususa ama-node njengoba kudingeka libaluleke kakhulu.

Ukucaciswa okubalulekile kwe-GIGABYTE AI TOP ATOM

• I-SoC - NVIDIA GB10
  • CPU - 20 Armv9 cores ene-10× Cortex‑X925 kanye ne-10× Cortex-A725
  • Izakhiwo – NVIDIA Grace Blackwell
  • GPU - I-Blackwell Architecture
  • Izinhlobo ze-CUDA – Blackwell Generation
  • Ama-Tensor Cores - Isizukulwane sesi-5
  • Ama-RT cores - Isizukulwane sesi-4
  • Ukusebenza kwe-Tensor - 1.000 AI TOPS (FP4)
  • I-VPU - 1 × VENNC, 1 × NNVDEC
• Inkumbulo yesistimu – 128 GB 256-bit LPDDR5X (≈273 GB/s)
• Isitoreji - Kufika ku-4 TB PCIe Gen5 SSD
• Ividiyo – HDMI 2.1a okukhiphayo
• Audio - Umsindo weziteshi eziningi nge-HDMI
• Red
  • 1× 10GbE RJ45
  • I-SmartNIC I-ConnectX‑7 ukuze ujoyine amasistimu amabili afika ku-200 Gbps
  • I-Wi-Fi 7 y I-Bluetooth 5.3
• USB
  • 1× USB 3.2 Gen 2×2 Type‑C (PD IN)
  • 3× USB 3.2 Gen 2×2 Type‑C (kufika ku-20 Gbps)
• Ukusetshenziswa – 240 W i-adaptha yangaphandle
• Ubukhulu - 150 × 150 × 50,5 mm
• Isisindo - 1,2 kg

Intengo nokutholakala

I-GIGABYTE AI TOP ATOM manje isifakwe ohlwini lwabathengisi abanjengoNewegg , ngentengo yokuqala engaba ngu- $3.499,99 yenguqulo ye-1TB PCIe 4.0 SSD. Ngaphezulu kwalokho kukhona inketho ye -4TB PCIe 4.0 (cishe u-$3.899,99) kanye nenguqulo ephezulu ye- 4TB PCIe 5.0 (cishe u-$3.999,99). Njengokujwayelekile, kuyalulekwa ukuqinisekisa ukutholakala kubathengisi bendawo bese ubheka ikhasi lomkhiqizo elisemthethweni kanye nesitatimende sabezindaba ukuze uthole ukucushwa kanye nezikhathi zokulethwa.

I-ecosystem nezinye izindlela ngesisekelo esifanayo

Ukufika kwale pulatifomu akugcini nje ku-GIGABYTE. I-MSI ne-ASUS sezivele zibonise ama-PC azo amancane asekelwe ku-chip efanayo ye-NVIDIA DGX Spark, okuqinisekisa ukuthi sibheke indinganiso yangempela ku-AI yedeskithophu. Kumsebenzisi wokugcina, lokhu kuholela ekukhetheni umenzi ofanelana kahle nezidingo zakhe ngokwedizayini, amachweba, isevisi, kanye nezici ezengeziwe, bazi ukuthi ubuchwepheshe obuyisisekelo (GB10) buyafana.

Ukuqhathanisa okuqondile: I-AI TOP ATOM vs DGX Spark

Uma ucabanga ngakho konke okungenhla, ukuqhathanisa kuyacaca. I-DGX Spark ingumqondo kanye neplatifomu yokulinganisa: ichaza isici sefomu esingu-1-litre, ukusetshenziswa kwe -chipset ye-GB10 Grace Blackwell , kanye nomgomo wokuletha i-petaFLOP FP4 etafuleni. Yakhelwe phezu kwaleso sisekelo, i-AI TOP ATOM ye-GIGABYTE iwumkhiqizo wokugcina ohlanganisa lo mbono nge-128GB yememori ehlanganisiwe, isitoreji esisheshayo, amachweba azinikele, kanye nokusetshenziswa kwayo kokwenza lula ukuhamba komsebenzi kwe-AI.

Ngakho-ke, maqondana namandla aluhlaza kanye nobubanzi bemodeli, sibheka izibalo ezilingana cishe (amapharamitha angu-1.000 e-TOPS FP4 kanye namapharamitha afinyelela ku- 200B ku-node eyodwa, ene- ~405B kuma-node amabili axhunyiwe). Umehluko osebenzayo usekusetshenzisweni okuthile : izinsiza ezifakiwe (i-AI TOP Utility), ukucushwa kwetheku eliqondile, izinketho zokugcina, kanye, nentengo kanye nokutholakala . Uma ufuna ulwazi "olungaphandle kwebhokisi" oluvela ku-OEM, i-Atom ifanelana kahle kakhulu.

Ukuze siqedele lesi sithombe, kubalulekile ukukhumbula lokho okushiwo imithombo yolimi lwesiSpanishi: I-GIGABYTE ikhuluma ngokusobala ngokuqoqa amayunithi amaningi kanye nezinhlelo zokusebenza kusukela kwezemfundo nesayensi kuya emadolobheni ahlakaniphile kanye namarobhothi. Umoya we-Spark uhlala unjalo: amandla e-AI angempela ngesimo esincane kakhulu futhi akhula ngokushesha lapho kudingeka amandla amaningi okucubungula.

Kuyacaca ukuthi sikhuluma ngokugxuma kwesizukulwane ku-"desktop AI": i-PC encane efana ne- GIGABYTE AI TOP ATOM izuza ifa le- NVIDIA DGX Spark (eyayikade i-Project DIGITS), ifaka i-petaFLOP eyi-1 ye-FP4 ilitha, inikeza i- 128 GB ehlanganiswe ne-NVLink-C2C, ikhuphula kufika kumapharamitha angu-~405B uma ihlanganisa ama-node amabili nge-ConnectX-7, futhi iletha i- NVIDIA stack ene-DGX OS/Ubuntu kanye ne-AI TOP utility ukuze kube lula ukulanda, ukuphetha, i-RAG, kanye ne-ML; uma sengeza i-10GbE , i-HDMI 2.1a , i-USB 3.2 Gen 2x2 , kanye namanani asevele ashicilelwe ku-Newegg, sinesiphakamiso esiqinile socwaningo, amakilasi, kanye namaqembu afuna ukusheshisa imijikelezo yawo ye-AI ngaphandle kokusetha iseva ezinikele.

I-athikili ehlobene:

Ungayithola kanjani i-NPU ku-Copilot+ PC: Umhlahlandlela Ophelele Wokuthola Okuningi Ku-AI

Isaka

Umbhali oshisekayo ngomhlaba wamabhayithi nobuchwepheshe ngokujwayelekile. Ngiyathanda ukwabelana ngolwazi lwami ngokubhala, futhi yilokho engizokwenza kule bhulogi, ngikubonise zonke izinto ezithakazelisayo kakhulu ngamagajethi, isofthiwe, ihadiwe, izitayela zobuchwepheshe, nokuningi. Inhloso yami ukukusiza ukuthi uzulazule emhlabeni wedijithali ngendlela elula nejabulisayo.