- I-B200 ivelele ngenkumbulo yayo ye-HBM3e engu-180 GB kanye ne-bandwidth enkulu engu-8 TB/s.
- Sethula ukwakheka kweBlackwell ngama-Tensor cores esizukulwane sesi-5 kanye nokusekelwa kwendabuko kokunemba kwe-FP4.
- Idlula kakhulu i-H100 ekusebenzeni kokuqagela, inciphisa kakhulu izindleko ngethokheni ngayinye.
- Isebenzisa i-NVLink 5.0 interconnect ukuze ifinyelele ukuxhumana okuqondiswayo okungu-1.8 TB/s nge-GPU ngayinye.
Uma unesithakazelo ngehadiwe yanamuhla, ngokuqinisekile usuvele uzwile nge-quantum leap emelelwa uchungechunge lwe-Blackwell . I-NVIDIA B200 akuyona nje enye intuthuko; iyisilo esicubungulayo esenzelwe ngqo ukususa inkumbulo nokubala izithiyo zobuhlakani bokwenziwa.
Leli khadi livezwa njengegugu eliyinhloko lezikhungo zedatha, ligxile ekuxazululeni inkinga endala yamamodeli amakhulu olimi (i-LLM). Ngomklamo ophazamisayo namandla ahlazisa abanduleli bayo, i-B200 yenza ukuqeqeshwa nokuthunyelwa kwamamodeli kube lula uma kuqhathaniswa nalokho esasikwenza eminyakeni embalwa edlule.
Imininingwane Yobuchwepheshe kanye Nokwakhiwa Kwangaphakathi
Ngaphansi kwe-hood, i-B200 iwumsebenzi wobunjiniyela osekelwe kusakhiwo se-Blackwell. Isebenzisa umklamo we -dual-chip GB100 , lapho ama-dies amabili ehlanganiswa khona nge-10 TB/s chip-to-chip interconnect, okuvumela uhlelo lokusebenza ukuthi luyibone njengeyunithi eyodwa. Ikhiqizwa kusetshenziswa inqubo ye -TSMC ye-4NP futhi ine-transistors ebiza izigidigidi ezingama-208.000.
Ngokuphathelene nokulungiselelwa kwayo kokwenza umsebenzi, inamayunithi e-shader angu-18.944, ama-TMU angu-592, kanye nama-ROP angu-24. Isivinini sayo sewashi eliyisisekelo singama-120 MHz, kodwa singakhuphuka sifike ku-1830 MHz , kuyilapho imemori isebenza ku-2000 MHz. Konke lokhu kukhishwa kwamandla kuphumela ku-TDP engu-1000W ku-SXM form factor yayo, yize okunye ukusetshenziswa kungahluka phakathi kuka-700W no-1000W kuye ngendawo ezungezile.
Inkumbulo kanye ne-bandwidth: Inhliziyo Yokusebenza
Lapho i-B200 ikhanya khona ngempela kusekuphathweni kwedatha yayo. Inememori ye-HBM3e engu-180 GB , inani elivumela ukulayisha amamodeli amakhulu ngaphandle kokusebenzisa ukuhlukana kuma-GPU amaningi. Kodwa akuwona nje umthamo; i -bandwidth engu-8 TB/s yiyo eyenza umehluko, njengoba cishe inkulu ngokuphindwe ka-2,4 kuneye-H100.
Kalula nje, ukuqagela kwe-LLM empeleni kuyinkinga yokufunda inkumbulo. Lapho kwakhiwa ithokheni ngayinye, i-GPU kumele ifunde lonke uhlu lwesisindo semodeli. Ngale bandwidth, i-B200 ingagcina isivinini esikhulu kakhulu sokukhiqiza amathokheni , isuse ukubambezeleka okuvame ukuvela kumamodeli anamapharamitha angu-70B noma ngaphezulu.
Amandla Okusebenzisa Ikhompyutha kanye Nokugxuma Ku-FP4
Ukuqamba okusha okukhulu yi-5th generation Tensor cores, ezethula ukwesekwa kwemvelo kokunemba kwe-FP4 . Leli khono libalulekile ngoba livumela ukwehla okungu-50% kwememori uma kuqhathaniswa ne-FP8, okuphinda kabili inani lamapharamitha angacutshungulwa. Ngokwesimo esijwayelekile, i-B200 ifinyelela ama-PetaFLOPS angu-20 ku-FP4 kanye nama-PetaFLOPS angu-9 ku-FP8.
Uma kuqhathaniswa nesizukulwane se-Hopper, ijubane liyamangaza. Ngenkathi i-H100 ihluleka ekusebenzeni okunembe kakhulu, i-B200 inikeza ukusebenza okulinganiselwe okuphindwe kane . Lokhu kusho ukuthi izindleko ziphansi kakhulu ngamathokheni ayisigidi, okwenza kube nenzuzo enkulu ezinkampanini ezikhonza ama-AI API ngezinga.
Ukuqhathanisa Okuqondile: B200 vs H100 kanye ne-H200
Uma uzibuza ukuthi kuyafaneleka yini ukwenza isibuyekezo, impendulo emfushane inguyebo, uma nje izakhiwo zakho zinkulu. Uma kuqhathaniswa ne-H100 SXM5, ene-80GB ye-VRAM kuphela, i-B200 inikeza imemori engaphezu kokuphindwe kabili . Lokhu kusho ukuthi isakhiwo se-70B Llama 3.1 ku-FP16 singangena kahle ekhadini elilodwa, sigweme ubunzima be-tensor parallelism.
Ngisho noma kuqhathaniswa ne-H200, eyayivele ithuthukisa inkumbulo (141 GB), i-B200 iwina nge -VRAM eyengeziwe engu-28% kanye ne-bandwidth ephezulu ngo-67%. Ngaphezu kwalokho, i-NVLink 5.0 interconnect ikhulisa ukuxhumana okuqondiswayo ku-1.8 TB/s, okuphindwe kabili kunalokho kwe-NVLink 4.0, okuvumela ukukala okuseduze komugqa ekucushweni kwe-8-GPU.
Izidingo Zengqalasizinda Namandla
Asinakuyishaya indiva iqiniso lokuthi ukuhambisa amandla anjalo kudinga ingqalasizinda engathi sína. Uhlelo lwe-B200 olune-GPU eyisishiyagalombili lungasebenzisa amandla aphakathi kuka-7 no-8 kW ekucutshungulweni kuphela. Nakuba ukupholisa umoya kusebenza kahle kuma-rack anomoya omuhle, anomthamo ophezulu, ukupholisa uketshezi ngqo kunconywa kakhulu ukuze kulondolozwe isikhathi eside sehadiwe futhi kuvinjelwe ukugoba kokushisa.
Ngokuphathelene nokuxhumeka, isebenzisa isikhombimsebenzisi se -PCI-Express 6.0 x16 , futhi uhlelo lwesofthiwe luhambisana ngokugcwele ne-CUDA 12.x kanye ne-cuDNN 9.x. Noma iyiphi ikhodi esivele isebenza ku-H100 izosebenza ku-B200 ngaphandle kwezinguquko, yize ukuze uthole okuningi ku-FP4, kubalulekile ukusebenzisa i-TensorRT-LLM 0.17+ noma izinguqulo ezibuyekeziwe ze-vLLM.
Ukuhlaziywa Kwezindleko Zamafu Nokutholakala Kwazo
Emakethe yokuqasha ye-GPU, i-B200 izibeke njengenketho ekhangayo kakhulu. Kumapulatifomu afana ne-RunPod noma i-Spheron, amanani afunwayo ngokuvamile asukela ku-$5,89 kuya ku-$9,36 ngehora, kuyilapho amanani angasetshenziswa angancipha aze afike ku-$5,34. Nakuba izinga lehora liphakeme kune-H100, ukusebenza kahle kwethokheni ngayinye kuphezulu kakhulu kangangokuthi izindleko zokugcina zesevisi zivame ukuba phansi kakhulu.
Naphezu kwesidingo esikhulu kanye nezigidi zamayunithi asele ngemuva kokuthengwa ngqo, ifu liyindlela esheshayo yokufinyelela iBlackwell. Inketho yokukhokhisa ngomzuzwana ivumela abathuthukisi ukuthi bahlole amamodeli abo e-FP4 ngaphandle kokuzibophezela ezinkontilekeni zengqalasizinda ebonakalayo eziyizigidi zamaRandi.
I-NVIDIA B200 ichaza kabusha lokho esikulindele ku-AI accelerator, ihlanganisa inkumbulo enkulu ye-HBM3e nokusekelwa kwe-FP4 okushaqisayo kanye nokuxhumeka kwe-NVLink 5.0 okusheshayo kakhulu. Ngokudlula kakhulu imikhawulo ye-bandwidth kanye nomthamo wochungechunge lwe-Hopper, iba ithuluzi eliphelele kulabo abaphatha amamodeli olimi amakhulu, okwenza ngcono kokubili ukusebenza kokuphetha kanye nezindleko zokusebenza ngethokheni ngayinye.