Yeqa uye kokuyinhloko
Ama-modeli kanye nobuchwepheshe

Ukuhlolwa kwemodeli ye-DeepSeek-V4-PRO AI: Ukusebenza kwe-NVIDIA Vera Rubin nge-megawatt ngayinye kufinyelela izikhathi ezingu-30 kwe-GB300

I-NVIDIA imemezele ukuthi uhlelo lwe-Vera Rubin NVL72 lukhulisa ukusebenza nge-megawatt ngayinye kumithwalo yomsebenzi ye-AI esebenza njengama-agent kuze kufike ezikhathini ezingu-30 uma kuqhathaniswa ne-GB300 NVL72, kuyilapho izindleko zama-token zehla kuze kufike ezikhathini ezingu-35. Lokhu kuchaza ukwakheka kwalo kanye nezinzuzo zalo zokusebenza: ukusebenza kwe-Vera Rubin ngeyunithi yamandla kufinyelela izikhathi ezingu-30 kwe-Blackwell

Ukuhlolwa kwemodeli ye-DeepSeek-V4-PRO AI: Ukusebenza kwe-NVIDIA Vera Rubin nge-megawatt ngayinye kufinyelela izikhathi ezingu-30 kwe-GB300

Izolo i-NVIDIA isebenzise umthwalo womsebenzi we-SemiAnalysis AgentX ukuze ihlole imodeli ye-DeepSeek-v4-PRO 1.6T, kwabonakala ukuthi ukusebenza kwe-Vera Rubin nge-megawatt ngayinye kufinyelela izikhathi ezingu-30 kwe-Blackwell.

Ukusebenza nge-megawatt ngayinye (Tokens per Megawatt, okufushaniswa ngokuthi tokens / MW) kuyinkomba eyinhloko yokusebenza kahle kwamandla ezikhungweni zedatha ze-AI, ekala ukuthi mangaki ama-AI Token angakhiqizwa ngaphansi kwesabelo samandla esinqunyiwe.

Ukuhlolwa kwemodeli ye-DeepSeek-V4-PRO AI: Ukusebenza kwe-NVIDIA Vera Rubin nge-megawatt ngayinye kufinyelela izikhathi ezingu-30 kwe-GB300

Ngokusho kwe-athikili ecashunwe yi-IT之家, i-SemiAnalysis AgentX ihlola ngokukhethekile imithwalo yomsebenzi yokucabanga ehlelwe ngama-agent kusetshenziswa izinkomba ezingu-4:

Ukusebenzisana okujwayelekile ukusuka ekugcineni kuye ekugcineni kubalwa ngesikhathi esiphelele kusukela ekuthunyelweni kwesicelo kuze kufike i-Token yokugcina, futhi kukhombisa ukusebenza kahle kokuphumayo okubonwa umsebenzisi, okuhlanganisa nesikhathi sokulinda i-Token yokuqala;

Ukusebenzisana okujwayelekile kubala kuphela isigaba sokukhiqiza esiphakathi kwe-Token yokuqala ne-Token yokugcina.

Ukubambezeleka kokuphela kuye ekupheleni kubala isikhathi esiphelele kusukela ekuthunyelweni kwesicelo esisodwa kuze kufike i-Token yokuphumayo yokugcina

Ukubambezeleka kwe-Token yokuqala (TTFT) kubala isikhathi kusukela ekuthunyelweni kwesicelo kuze kufike i-Token yokuphumayo yokuqala.

Emthwalweni womsebenzi we-DeepSeek-v4-PRO 1.6T, iseva ye-Grace Blackwell esebenzisa i-GB300 NVL72 ifinyelela ukusebenza nge-megawatt ngayinye okuphindwe izikhathi ezingu-15 kunesixazululo se-H200 NVL8 Hopper.

Ukuhlolwa kwemodeli ye-DeepSeek-V4-PRO AI: Ukusebenza kwe-NVIDIA Vera Rubin nge-megawatt ngayinye kufinyelela izikhathi ezingu-30 kwe-GB300

Izindleko ze-Blackwell zama-Token ayisigidi zicishe zibe ingxenye yeshumi yezesizukulwane esedlule. Abahlinzeki bangakwazi ukusekela ama-agent amaningi ngaphansi kwesabelomali esifanayo samandla nengqalasizinda, noma balondoloze umthamo ofanayo wama-agent ngezindleko zokusebenza eziphansi.

Ukuhlolwa kwemodeli ye-DeepSeek-V4-PRO AI: Ukusebenza kwe-NVIDIA Vera Rubin nge-megawatt ngayinye kufinyelela izikhathi ezingu-30 kwe-GB300

I-Vera Rubin iqhubekisela phambili le nzuzo. Kumodeli ye-DeepSeek V4 Pro, ukusebenza kwe-Vera Rubin NVL72 nge-megawatt ngayinye kucishe kube izikhathi ezingu-30 kune-GB300 NVL72.

Ukuhlolwa kwemodeli ye-DeepSeek-V4-PRO AI: Ukusebenza kwe-NVIDIA Vera Rubin nge-megawatt ngayinye kufinyelela izikhathi ezingu-30 kwe-GB300

Izindleko ze-Vera Rubin NVL72 zama-Token ayisigidi ziyingxenye ka-35 yezindleko ze-GB300 NVL72, okuyenza ikwazi ukusebenzisa ama-agent ngokuqhubekayo nangesilinganiso esikhulu ukuze ihlangabezane nezidingo zamakhasimende zemithwalo yomsebenzi ehlukahlukene.

Ukuhlolwa kwemodeli ye-DeepSeek-V4-PRO AI: Ukusebenza kwe-NVIDIA Vera Rubin nge-megawatt ngayinye kufinyelela izikhathi ezingu-30 kwe-GB300
Ukuhlolwa kwemodeli ye-DeepSeek-V4-PRO AI: Ukusebenza kwe-NVIDIA Vera Rubin nge-megawatt ngayinye kufinyelela izikhathi ezingu-30 kwe-GB300