I-AMD CTO ikhuluma nge-Chiplet: Isikhathi sokuhlanganisa isithombe sikagesi siyeza

Abaphathi benkampani ye-chip ye-AMD bathi amaphrosesa e-AMD azayo angase ahlonyiswe ngama-accelerator aqondene nesizinda, futhi amanye ama-accelerator enziwa abantu besithathu.

IPhini likaMongameli Omkhulu uSam Naffziger ukhulume neSikhulu Esiphezulu Sezobuchwepheshe be-AMD uMark Papermaster kuvidiyo ekhishwe ngoLwesithathu, egcizelela ukubaluleka kokumiswa kwe-chip encane.

“Ama-accelerator aqondene nesizinda, leyo yindlela engcono kakhulu yokuthola ukusebenza okuhle kakhulu ngedola ngewatt ngayinye.Ngakho-ke, kuyadingeka ngempela ukuze kuthuthuke.Ngeke ukwazi ukwenza imikhiqizo ethile yendawo ngayinye, ngakho-ke esingakwenza ukuba ne-chip ecosystem encane – empeleni umtapo wolwazi, “kuchaza uNaffziger.

Ubekhuluma nge-Universal Chiplet Interconnect Express (UCIe), okuyindinganiso evulekile yokuxhumana kwe-Chiplet ebilokhu ikhona kusukela yasungulwa ekuqaleni kuka-2022. Izuze ukwesekwa okugcwele okuvela kubadlali bezimboni ezinkulu abafana ne-AMD, i-Arm, i-Intel ne-Nvidia, futhi. njengamanye amabhrendi amaningi amancane.

Kusukela yethula isizukulwane sokuqala se-Ryzen ne-Epyc processors ngo-2017, i-AMD ibihamba phambili ekwakhiweni kwe-chip encane.Kusukela lapho, umtapo wezincwadi we-House of Zen wama-chips amancane usukhulile wafaka ama-compute amaningi, i-I/O, nama-graphics chips, ukuhlanganisa nokuwahlanganisa ku-customer and data center processors.

Isibonelo sale ndlela singatholakala ku-AMD's Instinct MI300A APU, eyethulwe ngoDisemba 2023, Ipakishwe ngama-chips amancane ayi-13 (ama-I/O ama-chips amane, ama-GPU ayisithupha, nama-CPU ama-CPU amathathu) nezitaki eziyisishiyagalombili zememori ye-HBM3.

UNaffziger uthe ngokuzayo, amazinga afana ne-UCIe angavumela ama-chips amancane akhiwe abantu besithathu ukuthi athole indlela yawo kumaphakheji e-AMD.Ubalule i-silicon photonic interconnect - ubuchwepheshe obungadambisa amabhodlela omkhawulokudonsa - njengamandla okuletha ama-chips amancane enkampani yangaphandle emikhiqizweni ye-AMD.

U-Naffziger ukholelwa ukuthi ngaphandle kokuxhumeka kwe-chip enamandla aphansi, ubuchwepheshe abunakwenzeka.

"Isizathu sokuthi ukhethe ukuxhumeka kwe-optical kungenxa yokuthi ufuna umkhawulokudonsa omkhulu," uyachaza.Ngakho udinga amandla aphansi kancane ukuze ufinyelele lokho, futhi i-chip encane ephaketheni iyindlela yokuthola isixhumi esibonakalayo samandla aphansi kakhulu. "Wengeze ngokuthi ucabanga ukuthi ukushintshela kuma-optics wokupakisha "kuyeza."

Ukufeza lokho, iziqalo ezimbalwa ze-silicon photonics sezivele zethula imikhiqizo engenza lokho kanye.Ngokwesibonelo, i-Ayar Labs, ithuthukise i-UCIe ye-photonic chip ehlanganiswe ne-prototype graphics analytics accelerator Intel eyakhelwe ngonyaka odlule.

Ukuthi ama-chips amancane enkampani yangaphandle (ama-photonics noma obunye ubuchwepheshe) azothola indlela yawo emikhiqizweni ye-AMD kusazobonakala.Njengoba sike sabika phambilini, ukumiswa kungenye yezinselelo eziningi ezidinga ukunqotshwa ukuze kuvunyelwe ama-chip chips amaningi.Sicele i-AMD ukuthola ulwazi olwengeziwe mayelana neqhinga labo le-chip encane futhi sizokwazisa uma sithola noma iyiphi impendulo.

I-AMD ike yahlinzeka ngama-chips ayo amancane kubakhi bama-chip abayizimbangi.Ingxenye ye-Intel's Kaby Lake-G, eyethulwe ngo-2017, isebenzisa umgogodla wesizukulwane sesishiyagalombili seChipzilla kanye ne-AMD's RX Vega Gpus.Ingxenye isanda kuvela kabusha ebhodini le-Topton's NAS.