ho fihlela ho 4.96 Times ea BERT-Large Inference
Bukana ea Mosebelisi
Finyella ho fihla ho makhetlo a 4.96 ho latela maikutlo a BERT-Large
Mehato ea M6i e entse Mosebetsi o mongata oa Inference ho feta M6g Instances e nang le li-processor tsa AWS Graviton2
Merero e mengata ea ho ithuta ka mochini oa puo ea tlhaho e ipapisitse le li-chatbots le lits'ebetso tse ling tsa khoebo. Ha mesebetsi ena e ntse e sekaseka mongolo o tlatsitsoeng ke bareki le basebelisi ba bang, ba ka beha litlhoko tse boima ho lisebelisoa tsa komporo. Sena se etsa hore ho be bohlokoa ho khetha maemo a maru a fanang ka ts'ebetso e phahameng.
BERT-Large ke mofuta o akaretsang oa ts'ebetso ea puo ea tlhaho (NLP) eo re e khethileng ho lekanya ts'ebetso ea li-Amazon tse peli. Web Litšebeletso (AWS) EC2 mefuta ea mehlala ea maru. Re lekile mefuta e 'meli ea M6i ka li-processor tsa 3rd Gen Intel Xeon Scalable le maemo a M6g ka li-processor tsa AWS Graviton2.
Re fumane hore maemo a mabeli a 32-vCPU le 64-vCPU M6i a nang le li-processor tsa 3rd Gen Intel Xeon Scalable li atlehile ho feta balekane ba bona ba M6g. Liphuputso tsa rona li bontša hore likhoebo li ka fana ka boiphihlelo bo potlakileng ho basebelisi ba tsona ka ho khetha maemo a M6i. Ho phaella moo, nakong ea ho phatlalatsoa, ha li-VM tsa lihlopha tsa M6i li bitsa 24.6% ho feta li-VM tsa M6g, liketsahalo tsa M6i-tse nang le ts'ebetso ho fihlela makhetlo a 4.96 - li fana ka ts'ebetso e ntle haholo ka dollar.
Boemo ba M6i Ka 32 vCPUs
Ho bapisa ts'ebetso ea maikutlo a BERT-Large ea letoto la mehlala e 'meli ea AWS, re sebelisitse moralo oa TensorFlow. Re lekile maemo a mabeli a nepahetseng: FP32, eo ka bobeli letoto la li-VM li tšehetsang, le INT8, e leng feela letoto la M6i le tšehetsang ka mefuta eo re e sebelisitseng. Joalokaha Setšoantšo sa 1 se bontša, liketsahalo tse 32-v CPU m6i.8xlarge tse sebelisang ho nepahala ha INT8 li fane ka makhetlo a 4.96 ka ts'ebetso ea liketsahalo tsa m6g.8xlarge ho sebelisa FP32 ho nepahala.
E amanang le 32-vCPU BERT-Large Inference Performance
Ho potlaka | Holimo ho molemo
Setšoantšo sa 1. BERT-Large inference performance e fihletsoeng ke sehlopha sa mohlala oa m6i.8xlarge le 3rd
Gen Intel Xeon Scalable processors le ka m6g.8xlarge mohlala cluster e nang le AWS Graviton2 processors. Holimo ho molemo.
BERT-Kholo
Fumana makhetlo a ka bang 4.96 ho feta mosebetsi oa tlhahiso-pele ea BERT-Large (INT8 ho nepahala) ka li-32-vCPU m6i.8xlarge tse nang le li-processor tsa 3rd Gen Intel Xeon Scalable.
khahlano le ho nepahala ha FP32 ka maemo a m6g.8xlarge
Fumana makhetlo a ka bang 3.07 ho feta mosebetsi oa tlhahiso-pele oa BERT-Large (INT8 ka nepo) ka maemo a 64‑vCPU m6i.16xlarge o nang le li-processor tsa 3rd Gen Intel Xeon Scalable
khahlano le ho nepahala ha FP32 ka maemo a m6g.16xlarge
Boemo ba M6i Ka 64 vCPUs
Joalokaha setšoantšo sa 2 se bontša, liketsahalo tse 64-vCPU m6i.16xlarge tse nang le li-processor tsa 3rd Gen Intel® Xeon® Scalable tse sebelisang INT8 ka mokhoa o nepahetseng li fana ka makhetlo a 3.07 ho feta liketsahalo tsa m6g.16xlarge tse nang le li-processor tsa AWS Graviton2 tse sebelisang FP32 ho nepahala.
Hlokomela: Mofuta oa BERT-Large oo re o sebelisitseng bakeng sa li-processor tsa AWS Graviton2 ha o tšehetse INT8 ho TensorFlow. 
Qetello
Re lekile ts'ebetso ea tlhahiso ea puo ea tlhaho ea BERT-E kholo ea letoto la mehlala e 'meli ea AWS: liketsahalo tsa M6i tse nang le 3rd.
Li-processor tsa Gen Intel Xeon Scalable le liketsahalo tsa M6g tse nang le li-processor tsa AWS Graviton2. Ka mefuta e 'meli e fapaneng, maemo a M6i a ile a sebetsa ho feta maemo a M6g, a fihlela makhetlo a 4.96 ho feta mosebetsi oa boithuto. Ho fana ka boiphihlelo bo potlakileng ho bareki ba hau le basebelisi ba bang, tsamaisa meroalo ea hau ea ts'ebetso ea NLP maemong a AWS M6i ka li-processor tsa 3rd Gen Intel Xeon Scalable.
Ithute haholoanyane
Ho qala ho tsamaisa meroalo ea hau ea ts'ebetso ea NLP maemong a AWS M6i ka li-processor tsa 3rd Gen Intel Xeon Scalable, etela https://aws.amazon.com/ec2/instance-types/m6i/.
Liteko tsa VM e le 'ngoe ke Intel ka 11/10/2021 le 12/01/2021. Li-VM tsohle li hlophisitsoe ka Ubuntu 20.04 LTS, 5.11.0-1022-aws, polokelo ea EBS, GCC = 8.4.0, Python=3.6.9, tensorflow=2.5.0, Docker=20.10.7,
containerd=1.5.5, BERT model, batch size 1, bolelele ba tatelano 384, FP32 le INT8 ho nepahala. Lintlha tsa mohlala: m6i.8xlarge, 32vcpus, Intel® Xeon® Platinum 8375C CPU @ 2.90GHz, 128 GB kakaretso ea memori ea DDR4; m6g.8xlarge, 32vcpus, ARM Neovers N1, Arm v8.2 @2.5GHz, 128 GB kakaretso ea memori ea DDR4; m6i.16xlarge, 64vcpus, Intel® Xeon® Platinum 8375C CPU @ 2.90GHz, 256 GB kakaretso ea memori ea DDR4; m6g.16xlarge, 64vcpus, ARM Neovers N1, Arm v8.2 @2.5GHz, 256 GB kakaretso ea memori ea DDR4.
Ts'ebetso e fapana ho latela ts'ebeliso, tlhophiso le lintlha tse ling. Ithute haholoanyane ho www.Intel.com/PerformanceIndex.
Liphetho tsa ts'ebetso li ipapisitse le liteko ho tloha matsatsing a bontšitsoeng ho litlhophiso 'me li kanna tsa se hlahise lintlafatso tsohle tse fumanehang phatlalatsa. Sheba bekapo bakeng sa lintlha tsa tlhophiso. Ha ho sehlahisoa kapa karolo e ka bolokehang ka botlalo. Litšenyehelo tsa hau le liphetho li ka fapana.
Mahlale a Intel a ka hloka lisebelisoa tse thata tsa software, software kapa ts'ebetso.
© Intel Corporation. Intel, logo ea Intel, le matšoao a mang a Intel ke matšoao a khoebo a Intel Corporation kapa lithuso tsa eona. Mabitso a mang le mabitso a mang a ka nkoa e le thepa ea ba bang.
E hatisitsoe USA 0722/JO/PT/PDF US002
Ka kopo Recycle
Litokomane / Lisebelisoa
 |
intel Finyella ho fihla ho 4.96 Times the BERT-Large Inference [pdf] Bukana ea Mosebelisi Fumana ho fihla ho 4.96 Times the BERT-Large Inference, Finyella ho fihla ho 4.96, Times the BERT-Large Inference |
