Kamfanin Tenstorrent wanda Jim Keller ke jagoranta ya fitar da na'urar sarrafa Wormhole ta zamani don ayyukan AI, wanda yake tsammanin zai samar da kyakkyawan aiki a farashi mai araha.Kamfanin a halin yanzu yana bayar da ƙarin katunan PCIe guda biyu waɗanda zasu iya ɗaukar na'urori masu sarrafawa guda ɗaya ko biyu na Wormhole, da kuma wuraren aiki na TT-LoudBox da TT-QuietBox ga masu haɓaka software. Duk sanarwar yau an yi ta ne ga masu haɓakawa, ba waɗanda ke amfani da allon Wormhole don ayyukan kasuwanci ba.
"A koyaushe abin farin ciki ne a sami ƙarin samfuranmu a hannun masu haɓaka software. Tsarin haɓaka software ta amfani da katunan Wormhole™ ɗinmu na iya taimaka wa masu haɓaka software na AI mai guntu da yawa," in ji Jim Keller, Shugaba na Tenstorrent.Baya ga wannan ƙaddamar, muna farin cikin ganin ci gaban da muke samu tare da fitar da tef ɗin samfurinmu na ƙarni na biyu, Blackhole.
Kowace na'urar sarrafa Wormhole tana ɗauke da cores 72 na Tensix (biyar daga cikinsu tana goyon bayan cores na RISC-V a cikin nau'ikan bayanai daban-daban) da kuma 108 MB na SRAM, suna isar da 262 FP8 TFLOPS a 1 GHz tare da ƙarfin ƙirar zafi na 160W. Katin Wormhole n150 mai guntu ɗaya yana da ƙwaƙwalwar bidiyo ta GDDR6 12 GB kuma yana da bandwidth na 288 GB/s.
Masu sarrafa Wormhole suna ba da sassaucin daidaitawa don biyan buƙatun ayyuka daban-daban. A cikin saitin wurin aiki na yau da kullun tare da katunan Wormhole n300 guda huɗu, ana iya haɗa masu sarrafa su zuwa naúra ɗaya da ke bayyana a cikin software a matsayin hanyar sadarwa mai faɗi da faɗaɗɗiya ta Tensix. Wannan tsari yana ba wa mai haɓaka damar sarrafa nauyin aiki iri ɗaya, raba tsakanin masu haɓakawa huɗu ko gudanar da samfuran AI daban-daban har zuwa takwas a lokaci guda. Babban fasalin wannan daidaitawa shine cewa yana iya gudana a cikin gida ba tare da buƙatar virtualization ba. A cikin yanayin cibiyar bayanai, masu sarrafa Wormhole za su yi amfani da PCIe don faɗaɗawa a cikin injin, ko Ethernet don faɗaɗawa waje.
Dangane da aiki, katin Wormhole n150 na Tenstorrent mai kwakwalwa ɗaya (ƙwayoyin Tensix 72, mita 1 GHz, 108 MB SRAM, 12 GB GDDR6, bandwidth 288 GB/s) ya sami 262 FP8 TFLOPS a 160W, yayin da allon Wormhole n300 mai kwakwalwa biyu (ƙwayoyin Tensix 128, mita 1 GHz, 192 MB SRAM, jimlar 24 GB GDDR6, bandwidth 576 GB/s) yana isar da har zuwa 466 FP8 TFLOPS a 300W.
Domin sanya 300W na 466 FP8 TFLOPS a cikin mahallin, za mu kwatanta shi da abin da Nvidia, babbar kasuwar AI, ke bayarwa a wannan ƙarfin ƙira na zafi. A100 na Nvidia ba ya goyan bayan FP8, amma yana goyan bayan INT8, tare da mafi girman aikin TOPS 624 (TOPS 1,248 lokacin da ya yi ƙaranci). Idan aka kwatanta, H100 na Nvidia yana goyan bayan FP8 kuma ya kai mafi girman aikin TFLOPS 1,670 a 300W (TFLOPS 3,341 a ƙarancin), wanda ya bambanta sosai da Wormhole n300 na Tenstorrent.
Duk da haka, akwai babbar matsala guda ɗaya. Tenstorrent's Wormhole n150 yana sayarwa akan $999, yayin da n300 ke sayarwa akan $1,399. Idan aka kwatanta, katin zane guda ɗaya na Nvidia H100 yana siyarwa akan $30,000, ya danganta da adadin. Tabbas, ba mu san ko masu sarrafa Wormhole huɗu ko takwas za su iya isar da aikin H300 guda ɗaya ba, amma TDPs ɗinsu sune 600W da 1200W bi da bi.
Baya ga katunan, Tenstorrent yana ba da wuraren aiki da aka riga aka gina ga masu haɓaka, gami da katunan n300 guda 4 a cikin TT-LoudBox mai araha mai tushen Xeon tare da sanyaya mai aiki, da kuma TT-QuietBox mai ci gaba tare da aikin sanyaya ruwa na Xiaolong) wanda ke tushen EPYC).
Lokacin Saƙo: Yuli-29-2024