Ngefomethi efana ne-PC encane, uNvidia wethule I-DGX Spark, ithimba elichazwa umkhiqizo njengelokuqala I-supercomputer yomuntu siqu yobuhlakani bokwenziwa. Inhloso yayo icacile: ukuletha etafuleni lonjiniyela amandla abewadinga ngaphambilini izikhungo zedatha noma ifu.
Isiphakamiso sigqame ngokusivumela okucatshangwayo kanye nokulungiswa kwamamodeli wendawo, enezibalo ezingavamile zedeskithophu exhunywe endaweni evamile. NgokukaNvidia, ngokunemba okuncishisiwe kwe-AI (FP4) inikeza kuze kufike ku-1 petaFLOP ukusebenza, okwanele ukukhonza amamodeli ngawo 200.000 billion amapharamitha ngaphandle kokushiya ikhaya, ilabhorethri noma ihhovisi.
Iyini i-DGX Spark futhi ngekabani?
I-DGX Spark akuyona i-PC yabathengi noma i-rig yokudlala; Ikona ithuluzi lokuthuthukisa. Isebenza nge I-DGX OS (okuhlukile kwe-Ubuntu ngesoftware ye-Nvidia's AI efakwe ngaphambili) futhi kuhloswe ngayo abacwaningi, abathuthukisi nabafundi abadinga ukuzama, ukucwenga kanye nokuqinisekisa amaphrojekthi endaweni yabo.
Ifilosofi yokusebenzisa ilula: kusuka ezintweni zakho ezijwayelekile usebenzisa izinhlelo zokusebenza zendabuko futhi, lapho i-AI idingeka, unikeza umthwalo ku-DGX Spark ngenethiwekhi yendawo. Ngale ndlela ugwema ukuncika kwamafu, ukubambezeleka kanye nezindleko eziguquguqukayo, ukugcina ukulawula idatha namamodeli.
Izakhiwo kanye nokucaciswa okubalulekile
Inhliziyo yeqembu yi- GB10 Grace Blackwell, i-superchip ehlanganisa i-CPU ne-GPU ngaphansi kophahla olufanayo ukuze kwandiswe ukuhambisana kwenkumbulo futhi unciphise izingqinamba. I-CPU ihlanganisa 20 Arm cores (I-Cortex-X925 esebenza kahle engu-10 kanye ne-10 Cortex-A725 esebenza kahle), 128GB yememori ehlanganisiwe ye-LPDDR5X okuvimbela ukudluliswa kwedatha phakathi kwe-RAM ne-VRAM.
Kusitoreji, isistimu isekela i-a I-NVMe M.2 SSD kufika ku-4TB ngokuzibethela ngokwakho, futhi ukuxhumana kuhlanganisa I-WiFi 7, I-Bluetooth 5.4, obomvu I-10 GbE, I-HDMI y ezine USB-C. Amandla aphezulu azungezile 240 W, esiza ukugcina i- ifomethi ehlangene nomsindo oqukethwe, okuwubuqili kancane eziteshini ezinama-GPU ahlukene asezingeni eliphezulu.
Inkundla isilungele izixhumanisi ezinesivinini esikhulu kanye namakhono enethiwekhi njenge I-ConnectX-7, eyenzelwe izimo ezithuthukile zokuthuthukisa. Empeleni, i-chassis ingafakwa kunoma iyiphi ideskithophu futhi, ngesisindo sokuzungeza 1,2 kg, kulula ukuthutha phakathi kwamakamelo noma izindawo.
Ukusebenza kwe-AI yendawo
Isebenza ku-FP4, i-DGX Spark ifinyelela cishe 1 petaFLOP, isibalo esisibeka phakathi nendawo yokusebenza ehlangene nesiphakeli selabhorethri encane. Ngalolu siling, iyakwazi ukukhonza endaweni LLM kuze kufike ku-200B yamapharamitha, into kuze kube manje ebingavunyelwe emafini noma amathuluzi amakhulu kakhulu.
Kumaphrojekthi amakhulu kunalowo sayizi, uNvidia uphawula ukuthi kungenzeka xhuma izinhlelo ezimbili ze-Spark futhi ubhekane namamodeli amakhulu nakakhulu, anezinkomba zokufika ku 405.000 billion amapharamitha. Ngaphezu kwalokho, inkampani iqinisekisa ukuhambisana namamodeli we DeepSeek, Meta, NVIDIA, Google, Qwen futhi ngokujwayelekile nge ezinye izindlela zomthombo ovulekile ezingavumelana nemvelo.
I-ecosystem nezinguqulo zomkhiqizi
Ngaphandle kwemodeli esemthethweni ye-Nvidia, isiteji sizonikeza imishini evela kumikhiqizo efana ne I-Acer, i-Asus, i-Dell, i-Gigabyte, i-HP, i-Lenovo ne-MSI, ezokwethulwa izinguqulo zabo siqu nalesi sisekelo esifanayo. Le nkampani iphinde yamemezela ukuthi kukhona uhlelo olukhulu lwedeskithophu kulabo abaludingayo inkumbulo eyengeziwe nokwandiswa, futhi igcina izinketho zayo zekhathalogi njenge Isiteshi se-DGX ngezimo zesilinganiso esikhulu.
Kufanelekile ukugcizelela: Akuyona i-Windows PC futhi ayigxile emdlalweni. Inani lingaphakathi i-prototype, yenza ngcono, futhi infer nge-AI yendawo, nokukwazi ukuyisa lowo msebenzi efwini noma esikhungweni sedatha lapho iphrojekthi ikhula.
Intengo nokutholakala
I-Nvidia iqinisekisa inani le Amadola ka-3.999 (cishe I-3.451 euro ngaphambi kwezintela) kanye nokutholakala kusuka I-15 ngo-Okthoba kuwebhusayithi yabo. Ngeke kushibhile emaphaketheni amaningi, kodwa kumaphrofayili athile ochwepheshe kungase kukufanele. ukonga kokubhaliselwe kanye nentuthuko ku izikhathi zobumfihlo nezokuphendula.
Amayunithi okuqala kanye nokutholwa kokuqala
Idivayisi isivele iqalile ukujikeleza phakathi kozakwethu nochwepheshe bezimboni. Jensen Huang uzihambisele mathupha iyunithi ku-Elon Musk ezikhungweni ze-SpaceX, isenzo esikhumbuza ukulethwa kokuqala I-DGX-1 ngo-2016 ku-OpenAI. Ngokuhambisana, izinhlangano ezifana Anaconda, Cadence, Google, Hugging Face, LM Studio, Meta o Microsoft Bahlola uhlelo futhi bazivumelanise nezimo amamodeli namathuluzi ngokusatshalaliswa kwendawo.
I-DGX Spark mayelana amandla esikhungo sedatha kufomethi yedeskithophu, ene-ecosystem yesofthiwe ekhulile kanye nozakwethu behadiwe ngemuva kwayo. Kulabo okufanele basebenze ngamamodeli amakhulu, amathuba okusebenza endaweni ngaphandle kokuncika kwamafu kungenza umehluko ezindlekweni, ekulawuleni nasekusebenziseni kalula.
