kufika ku-4.96 Izikhathi ze-BERT-Large Inference
Umhlahlandlela Womsebenzisi
Finyelela kufika ku-4.96 Times the BERT-Large Inference
Izimo ze-M6i Zenze Umsebenzi Omuningi Wokuqondisisa kune-M6g Instances equkethe ama-AWS Graviton2 processors
Umsebenzi wokufunda ngomshini wolimi lwemvelo ugcizelela ama-chatbots nezinye izinhlelo zokusebenza zebhizinisi. Njengoba le mithwalo yomsebenzi ihlaziya umbhalo othayiphiwe amakhasimende nabanye abasebenzisi, ingafaka izimfuno ezinzima ezinsizakalweni zokubala. Lokhu kwenza kubeluleka ukukhetha izimo zamafu eziletha ukusebenza okuphezulu.
I-BERT-Large iyimodeli yenhloso evamile yokucubungula ulimi lwemvelo (NLP) esiyikhethe ukukala ukusebenza kwama-Amazon amabili. Web Amasevisi (AWS) EC2 izinhlobo zezibonelo zamafu. Sihlole osayizi ababili bezehlakalo ze-M6i ngamaphrosesa we-3rd Gen Intel Xeon Scalable kanye nezimo ze-M6g ngama-AWS Graviton2 processors.
Sithole ukuthi izimo zombili ze-32-vCPU kanye ne-64-vCPU M6i ezinamaphrosesa we-Intel Xeon Scalable wesi-3 asebenze kahle kakhulu kunozakwabo be-M6g. Okutholakele kwethu kubonisa ukuthi amabhizinisi angaletha umuzwa osheshayo kubasebenzisi bawo ngokukhetha izimo ze-M6i. Ukwengeza, ngesikhathi sokushicilela, kuyilapho ama-VM ochungechunge lwe-M6i abiza u-24.6% ngaphezu kwe-M6g yochungechunge lwe-VM, izimo ze-M6i-ngokusebenza okufika ku-4.96 izikhathi eziphumayo-zinikeza ukusebenza okungcono kakhulu ngedola ngalinye.
I-M6i Instances With 32 vCPUs
Ukuze siqhathanise ukusebenza kwe-BERT-Large kochungechunge lwesibonelo lwe-AWS, sisebenzise uhlaka lwe-TensorFlow. Sihlole amazinga amabili anembayo: i-FP32, esekela yomibili uchungechunge lwama-VM, kanye ne-INT8, esekelwa uchungechunge lwe-M6i kuphela ngamamodeli esiwasebenzisile. Njengoba Umfanekiso 1 ubonisa, izimo ezingu-32-v CPU m6i.8xlarge zisebenzisa ukunemba kwe-INT8 zilethe izikhathi ezingu-4.96 ukusebenza kwezimo ze-m6g.8xlarge kusetshenziswa ukunemba kwe-FP32.
Okuhlobene 32-vCPU BERT-Large Inference Performance
Ukusheshisa | Okuphakeme kungcono
Umfanekiso 1. Ukusebenza kwe-BERT-Enkulu okutholwe yiqoqo le-m6i.8xlarge elinesi-3
Ama-Gen Intel Xeon Scalable processors kanye ne-m6g.8xlarge example cluster enama-AWS Graviton2 processors. Okuphakeme kungcono.
I-BERT-Enkulu
Sukuma izikhathi ezifika kwezingu-4.96 umsebenzi we-BERT-Large (ukunemba kwe-INT8) ngezimo ezingu-32-vCPU m6i.8xlarge ezihlanganisa i-3rd Gen Intel Xeon Scalable processors.
iqhathaniswa nokunemba kwe-FP32 enezimo ze-m6g.8xlarge
Vuka izikhathi ezifika kwezingu-3.07 umsebenzi we-BERT-Large (ukunemba kwe-INT8) ngezimo ezingu-64‑vCPU m6i.16xlarge ezihlanganisa i-3rd Gen Intel Xeon Scalable processors
iqhathaniswa nokunemba kwe-FP32 enezimo ze-m6g.16xlarge
I-M6i Instances With 64 vCPUs
Njengoba umdwebo 2 ubonisa, izimo ezingu-64-vCPU m6i.16xlarge ezinamaphrosesa e-Intel® Xeon® Scalable e-3rd Gen asebenzisa ukunemba kwe-INT8 alethe izikhathi ezingu-3.07 ukusebenza kwezimo ze-m6g.16xlarge ngama-AWS Graviton2 processors asebenzisa ukunemba kwe-FP32.
Qaphela: Imodeli ye-BERT-Large esiyisebenzisele ama-AWS Graviton2 processors ayisekeli i-INT8 ku-TensorFlow. 
Isiphetho
Sihlole ukusebenza kwe-BERT-Large yokucubungula ulimi lwemvelo kochungechunge lwezibonelo ezimbili ze-AWS: izehlakalo ze-M6i ezifaka okwesithathu
Ama-Gen Intel Xeon Scalable processors kanye nezimo ze-M6g ezihlanganisa ama-AWS Graviton2 processors. Ngosayizi ababili abahlukene, izehlakalo ze-M6i zisebenze kangcono kunezimo ze-M6g, zafinyelela izikhathi ezifika ku-4.96 umsebenzi wokuqagela. Ukuze ulethe okuhlangenwe nakho okushesha kakhulu kumakhasimende akho nabanye abasebenzisi, sebenzisa i-NLP yakho eyisithenjwa enzima ezimweni ze-AWS M6i usebenzisa i-3rd Gen Intel Xeon Scalable processors.
Funda kabanzi
Ukuze uqale ukusebenzisa i-NLP yakho yokubeka imithwalo yemisebenzi ezimweni ze-AWS M6i nge-3rd Gen Intel Xeon Scalable processors, vakashela https://aws.amazon.com/ec2/instance-types/m6i/.
Ukuhlolwa kwe-VM eyodwa okwenziwa yi-Intel ngomhla ka-11/10/2021 nangomhla ka-12/01/2021. Wonke ama-VM alungiselelwe ngo-Ubuntu 20.04 LTS, 5.11.0-1022-aws, isitoreji se-EBS, GCC=8.4.0, Python=3.6.9, tensorflow=2.5.0, Docker=20.10.7,
containerd=1.5.5, imodeli ye-BERT, usayizi weqoqo 1, ubude bokulandelana okungu-384, FP32 nokunemba kwe-INT8. Imininingwane yesibonelo: m6i.8xlarge, 32vcpus, Intel® Xeon® Platinum 8375C CPU @ 2.90GHz, 128 GB isamba sememori engu-DDR4; m6g.8xlarge, 32vcpus, ARM Neovers N1, Arm v8.2 @2.5GHz, 128 GB inkumbulo ephelele ye-DDR4; m6i.16xlarge, 64vcpus, Intel® Xeon® Platinum 8375C CPU @ 2.90GHz, 256 GB inkumbulo ephelele ye-DDR4; m6g.16xlarge, 64vcpus, ARM Neovers N1, Arm v8.2 @2.5GHz, 256 GB isamba sememori engu-DDR4.
Ukusebenza kuyehluka ngokusetshenziswa, ukumisa nezinye izici. Funda kabanzi ku www.Intel.com/PerformanceIndex.
Imiphumela yokusebenza isekelwe ekuhlolweni kusukela kumadethi aboniswe ekucushweni futhi ingase ingabonisi zonke izibuyekezo ezitholakala esidlangalaleni. Bona ikhophi yasenqolobaneni ukuze uthole imininingwane yokucushwa. Awukho umkhiqizo noma ingxenye engavikeleka ngokuphelele. Izindleko zakho nemiphumela ingahluka.
Ubuchwepheshe be-Intel bungadinga ihadiwe enikwe amandla, isofthiwe noma isevisi isebenze.
© Intel Corporation. I-Intel, ilogo ye-Intel, nezinye izimpawu ze-Intel yizimpawu zokuthengisa ze-Intel Corporation noma izinkampani ezingaphansi kwayo. Amanye amagama namabhrendi angafunwa njengempahla yabanye.
Iphrintwe e-USA 0722/JO/PT/PDF US002
Sicela usebenzise kabusha
Amadokhumenti / Izinsiza
 |
intel Finyelela kufikela kokungu-4.96 Izikhathi ze-BERT-Large Inference [pdf] Umhlahlandlela Womsebenzisi Finyelela kufika ku-4.96 Times the BERT-Large Inference, Finyelela kufika ku-4.96, Times the BERT-Large Inference |
