Umkhiqizi wama-accelerator e-AI inference asezingeni le-wafer, i-Cerebras, umemezele ngomhlaka-18 ngesikhathi sendawo ukwethulwa kwe-chip yesizukulwane esisha i-WSE-3 Turbo kanye nohlelo lwe-CS-4 olwakhelwe phezu kwale chip. Uma iqhathaniswa ne-CS-3 yesizukulwane esedlule, i-CS-4 iletha umthamo wamathokheni ophindwe izikhathi eziyi-10 kanye nesivinini esiphindwe izikhathi ezi-2.
Njenge-WSE-3 yangaphambilini, i-WSE-3 Turbo ihlanganisa ama-core angu-900,000 kanye ne-cache ye-SRAM engu-44GB engaphakathi kwe-chip, kodwa amandla ayo e-FP16 sparse AI, umkhawulokudonsa wememori, umkhawulokudonsa wokuxhumana okungaphakathi kwe-chip kanye nomkhawulokudonsa we-I/O konke kuphindwe kabili. Uhlelo lwe-CS-4 lungahlanganisa ama-WSE-3 Turbo angu-3, ngaleyo ndlela lukhulise ngokwengeziwe ukuminyana kwamandla okubala.
I-Cerebras ithi i-CS-4 ifinyelela isivinini sokuletha amathokheni esingu-4465 Token/s kumodeli ye-OpenAI GPT-OSS, okuyizikhathi ezingu-30 kunesixazululo se-GPU, futhi sithuthuke ngo-93% uma siqhathaniswa ne-CS-3. Kumamodeli anamapharamitha amaningi kakhulu, ukuxhumana kwe-wafer-to-wafer okune-latency ephansi engu-2μs kwe-CS-4 kungafinyelela ekukhiqizeni amathokheni angaphezu kuka-1000 Token/s esikalini esingu-10T.
I-CS-4 isekela ngokwendabuko ukuhlukaniswa kwemithwalo ye-inference, futhi ingasetshenziswa njengeyunithi ye-decoding ehambisana nehadiwe ye-prefill eyahlukene, ukuze kusetshenziswe amandla ahlukile ezakhiwo zalezi zinhlangothi zombili.
Leli hadiwe lokubala lisekelwe epulatifomu entsha ye-Cerebras eyi-Nexus rack. Izici ezintathu ezinkulu, ukubala, ugesi nokuxhumana, manje sezakhiwe kabusha njengezinhlelo ezizimele, kanti inani lezinto ezakha uhlelo lonke lehliswe ngo-50% futhi izinga lokuzenzakalela kokukhiqiza lenyuswe kakhulu. I-CS-4 icishe yaqeda ngokuphelele ukulahleka kwamandla kagesi ekuphakeleni kwamandla ezingeni lebhodi, yahlinzeka i-WSE-3 Turbo ngamandla kagesi aphindwe kabili.