Jim Keller-Kamfanin Kamfanin Tenstorrent ya fito da kamfaninta na gaba na gaba-deskroms, wanda da yake so ya bayar da kyakkyawan aiki a farashi mai araha.Kamfanin a halin yanzu yana ba da ƙarin katunan PCIe guda biyu waɗanda zasu iya ɗaukar ɗayan masu sarrafawa ɗaya ko biyu, da kuma tt-amo don masu haɓaka software. Dukkanin sanarwar ta yau ta yi niyya ce, ba waɗanda ke amfani da allon da ba su dace ba don ayyukan aikin kasuwanci.
"Ko da yake gamsar da samun ƙarin samfuranmu a hannun masu haɓaka. Saki Tsarin ci gaba na amfani da kayan aikin ku na iya taimakawa Software na Wormhole," in ji Jim Keller, Shugaba na Glotorrent.Baya ga wannan ƙaddamar, muna murnar ganin cigaban da muke samu tare da tef ɗin da muke fitarwa da iko na samfur na biyu, blackhole. "
Kowane mai siyar da tsutsa yana dauke da cores 72 (Biyar da ke goyon bayan Hisc-V Stores a cikin tsari na bayanai daban-daban) da 108 Mb na Sram, da 10 GHZ tare da karfin zane na 160w. Katin-guntu-guntu Wormhole N150 Katin Gdr6 Memol6 GDDR6 bidiyo GDRL6. Kuma ƙwaƙwalwar bidiyo ta bidiyo, kuma tana da bandwidth na 288 GB / s.
Masu aikin tsutsa suna ba da sassauƙa scalable don saduwa da bukatun ayyukan. A cikin tsarin aiki na daidaitawa tare da katunan NOLMHole guda hudu, za a iya haɗe da masu sarrafawa zuwa cikin software a matsayin hadewa, babban Tensix Core cibiyar sadarwa azaman hade, mai yawa Tensix Core cibiyar sadarwa azaman haɗe, m tensix Core cibiyar sadarwa. Wannan saitin yana ba da damar maimaitawa don magance aikin guda ɗaya, raba tsakanin masu haɓakawa huɗu ko gudu zuwa samfuran AI takwas daban-daban. Wani fasali na mabuɗin wannan scalability shine cewa zai iya tafiyar da gida ba tare da buƙatar chinkalization ba. A cikin yanayin cibiyar bayanai, masu aiwatar da masu sarrafawa zasuyi amfani da PCMHole don fadada a cikin injin, ko Ethernet don fadada waje.
Dangane da aikin, guntu-guntu na kati goma na goma50 (72 tonsix, cores 12 GHZ, CORESSCHED (128 MB SORES, 108 MB SORES, 108 MB SDWOLS, 108 MB SORES, 10 GHZ, 1 GHZ Fice, 192 MB Sram, tara 24 gb gdr6, 576 gb / s Bandwidth) ya gabatar da TFLops har zuwa 366 FP8 a 300w.
Don sanya 300W na 466 FP8 TFLops zuwa cikin mahallin, za mu kwatanta shi da abin da shugaban kasuwar Ai Nvidia yake bayarwa a wannan ikon kasuwa na Ai. Nvidia ta A100 A100 ba ta goyon bayan FP8, amma yana goyan bayan Int8, tare da ƙarfin aikin 624 (1 na 1,248 fi lokacin da spears). A kwatancen, H100 ya tallafawa FP8 FP8 kuma ya kai babban wasan kwaikwayo na 1,670 a 300w (3,341 TFlops da tsinkaye), wanda yake da matukar muhimmanci daga wanthumhole da tsantsa na goma.
Koyaya, akwai babbar matsala. Tenstorrent na tsutsa na N150 na $ 999, yayin da N300 Sayar da $ 1,399. By kwatanta, guda NVIDIA H100 Graphics Card Retails na $ 30,000, ya danganta da adadi. Of course, we don't know if four or eight Wormhole processors can actually deliver the performance of a single H300, but their TDPs are 600W and 1200W respectively.
Baya ga katunan, Goma ya gabatar da kayan aikin da aka riga aka gina don masu haɓaka, gami da katunan ruwa na Xeon, da kuma Eyp-tushen Xiaolong).
Lokaci: Jul-2920