データサイエンスを加速するAIチップ / Introduction to AI chips which accelerate data science

2024年6月14日開催 Women in Data Science Tokyo @ IBM
DATA SCIENTIST TALK 資料
Speaker: 伊藤 愛
日本アイ・ビー・エム株式会社
東京基礎研究所 セミコンダクター
スタッフ・リサーチ・サイエンティスト

https://widstokyoibm2024.splashthat.com/

Transcript

1. σʔλαΠΤϯεΛՃ଎͢Δ AIνοϓ ҏ౻ Ѫ ౦ژجૅݚڀॴ ηϛίϯμΫλʔ ೔ຊIBMגࣜձࣾ
2. ⾃⼰紹介 ࢯ໊ɿҏ౻ Ѫ ձࣾɿ೔ຊΞΠɾϏʔɾΤϜגࣜձࣾ ॴଐɿ౦ژجૅݚڀॴ ηϛίϯμΫλʔ ৬छɿελοϑɾϦαʔνɾαΠΤϯςΟετ ུྺɿ ೥ͷ೔ຊ*#.ೖࣾҎདྷɺ౦ژجૅݚڀॴͰ ੑೳ෼ੳɺੑೳ࠷దԽɺ"*νοϓͷݚڀʹैࣄɻ ೥݄ΑΓ/PSUI1PMFϓϩδΣΫτʹࢀՃ͠ɺ ࿦ཧઃܭɺݕূɺిྗධՁͳͲΛ୲౰ɻ ೥݄ΑΓηϛίϯμΫλʔॴଐͱͳΓɺ ΑΓ൒ಋମͷݚڀʹ஫ྗ͍ͯ͠Δɻ
3. ©2024 IBM Corporation AIʹΑΔσʔλ෼ੳ 3 ػցֶश $IBU(15 ੜ੒"* ը૾ೝࣝ σΟʔϓϥʔχϯά ࣗવݴޠॲཧ νϟοτϘοτ धཁ༧ଌ '"2 ނোݕ஌ ࣗಈӡస
4. ©2024 IBM Corporation AIͷֶशͱਪ࿦ 4 𝑓 𝐱 𝐱 𝐲 ਪ࿦: ߏங͞Εͨؔ਺ 𝑓 Λ࢖ͬͯɺ৽͍͠ 𝐱 ͔Β 𝐲 Λਪଌ ֶश: ೖྗσʔλ΍ֶशσʔλΛ࢖͍ɺ 𝐱 ͔Β 𝐲 Λਪଌ͢Δؔ਺ 𝑓 Λߏங͢Δ cat dog ߏங Cat: 80% Dog: 20% ֶशσʔλ ਪଌ݁Ռ ਪ࿦σʔλ ྫɿڭࢣ͋Γֶश
5. ©2024 IBM Corporation ؔ਺ 𝑓 ͷߏஙํ๏ͷྫ 5 ઢܗճؼ αϙʔτϕΫλʔϚγϯ ΫϥελϦϯά χϡʔϥϧωοτϫʔΫ ʢ"*Ϟσϧʣ ܾఆ໦ ڠௐϑΟϧλϦϯά 𝑥! 𝑥" 𝑥 𝑦 𝑦 = 𝛼𝑥 𝑥! 𝑥" 𝑥! 𝑥# 𝑦! 𝑥" 𝑦# 𝑦" ॻ੶ A B C D E ސ٬ 1 Like Like 2 Like Like Like Like 3 Like ? Like 4 Like Like φΠʔϒϕΠζ ཧ༝ ݁Ռ ؍ଌ ਪଌ
6. ©2024 IBM Corporation χϡʔϥϧωοτϫʔΫͷԋࢉ 𝑦! 𝑦! , 𝑦" , … , 𝑦# = 𝜑 𝑥! , 𝑥" , … , 𝑥# 𝑤!! 𝑤!" … 𝑤!# 𝑤"! 𝑤"" … 𝑤"# ⋮ ⋮ ⋱ ⋮ 𝑤#! 𝑤#" ⋯ 𝑤## + 𝑩 𝑥! 𝑥" 𝑥# 𝑤!! 𝑤"! 𝑤#! 𝜑 ∑𝑥𝑤 + 𝑏 ׆ੑԽؔ਺ ϕΫτϧߦྻੵ ∑𝑥𝑤 ∑𝑥𝑤 − όΠΞε 𝑦" 𝑦# … • χϡʔϥϧωοτϫʔΫ͸ߦྻͰද͢͜ͱ͕Ͱ͖Δ • χϡʔϥϧωοτϫʔΫͷԋࢉʹ͸ɺଟ͘ͷߦྻԋࢉؚ͕·Ε͍ͯΔ … … 6
7. ©2024 IBM Corporation AIͷԋࢉΛ࣮ߦ͢ΔνοϓΞʔΩςΫνϟ CPU AIνοϓ ݩʑ3Dը૾ॲཧͷͨΊʹ ։ൃ͞Εͨνοϓɻը૾ॲ ཧͷͨΊͷߦྻϕΫλ৐ࢉ ثΛ๛෋ʹඋ͓͑ͯΓɺAI ͷԋࢉ΋ಘҙɻ AIͷԋࢉઐ༻ʹઃܭ͞Εͨ νοϓɻAIͷॲཧʹಛԽ͠ ͨػೳΛ๛෋ʹඋ͑ɺߴ͍ ੑೳΛތΔɻ ҰൠతͳԋࢉͷͨΊʹઃܭ ͞Εͨνοϓɻ࠷ۙͷCPU ͸AIͷԋࢉʹඞཁͳߦྻϕ Ϋλ৐ࢉث΋උ͍͑ͯΔɻ GPU IntelʮIntel Launches Xeon D Processor Built for the Network and Edge”ʯ https://www.intel.com/content/www/us/en/newsr oom/news/intel-launches-new-processor-built- edge.html#gs.0mokh7 (2024/6/11) ZDNETʮ NVIDIAͷ৽νοϓʮH100ʯɺAIʹΑΔਓؒཧղͷՃ଎Λ໨ࢦ͢ʯ https://japan.zdnet.com/article/35187234/ (2024/6/11) νοϓ FPGA AMD ʮADM-PCIE-9H3ʯ https://japan.xilinx.com/products/boards-and-kits/1-zihv8r.html (2024/6/11) ಛఆͷԋࢉઐ༻ͷճ࿏Λϓ ϩάϥϜ͢Δ͜ͱ͕Ͱ͖Δ νοϓɻAIԋࢉઐ༻ͷճ࿏ Λ࡞Δ͜ͱ͕Ͱ͖Δɻ 7
8. ©2024 IBM Corporation νοϓΞʔΩςΫνϟؒͷੑೳൺֱ Nature 「Hardware implementation of memristor-based artificial neural networks」 https://www.nature.com/articles/s41467-024-45670-9 (2024/6/11) 8
9. ©2024 IBM Corporation AIͷීٴʹΑΔফඅిྗ໰୊ IDC ʮࠃ಺σʔληϯλʔ಺ͷAI޲͚ిྗਪఆΛൃදʯ https://www.idc.com/getdoc.jsp?containerId=prJPJ51802224 (2024/6/11) IEA 「Electricity 2024 Executive summary」 https://www.iea.org/reports/electricity-2024/executive-summary (2024/6/11) IDC JapanʹΑΔࠃ಺σʔληϯλʔʹ͓͚ ΔAIαʔόʔ޲͚ిྗΩϟύγςΟͷਪܭ஋ IEAʢࠃࡍΤωϧΪʔػؔʣʹΑΔϨϙʔτ ʢElectricity 2024ʣ “After globally consuming an estimated 460 terawatt-hours (TWh) in 2022, data centres’ total electricity consumption could reach more than 1000 TWh in 2026. This demand is roughly equivalent to the electricity consumption of Japan. ” “2022೥ʹશੈքͰਪఆ460TWhΛফඅͨ͠ σʔληϯλʔͷ૯ిྗফඅྔ͸ɺ2026೥ʹ 1000TWhΛ௒͑ΔՄೳੑ͕͋Δɻ͜Ε͸೔ຊ શମͷిྗফඅྔʹ΄΅ඖఢ͢Δɻ” “2027೥຤࣌఺ʹ͓͚ΔAIαʔόʔ޲͚ిྗ ΩϟύγςΟ͸ɺ2024೥຤࣌఺ͷ໿1.5ഒͱ ͳΔ” 9
10. ©2024 IBM Corporation ओཁͳAIνοϓ Google TPU ֶश༻ʢਪ࿦΋Մೳʣ ਪ࿦ઐ༻ Intel Gaudi PFN MN-Core LeapMind Efficiera IBM NorthPole Google͕։ൃͨ͠AIઐ༻ͷϋʔυ΢ΣΞɻ Google CloudͰ׆༻͞Ε͍ͯΔɻ Intel͕։ൃͨ͠AIઐ༻ͷϋʔυ΢ΣΞɻ σʔληϯλʔͰͷੜ੒AI΍LLMͷ࣮ߦΛ ໨తʹ͍ͯ͠Δɻ Preferred Networks͕։ൃͨ͠AIઐ༻ͷ ϋʔυ΢ΣΞɻ AI/ج൫ϞσϧͷֶशϑΣʔζʹ࠷దԽ͍ͯ͠Δɻ LeapMind͕։ൃͨ͠AIઐ༻ͷϋʔυ΢ΣΞIPɻ CNN (Convolutional Neural Network) ͷ ਪ࿦ԋࢉॲཧʹಛԽ͍ͯ͠Δɻ IBM͕։ൃͨ͠AIਪ࿦ઐ༻ͷϋʔυ΢ΣΞɻ AIϞσϧͷ৘ใΛશͯνοϓ಺ʹ֨ೲ͢Δ ͜ͱʹΑΓσʔλͷҠಈΛେ෯ʹݮΒ͠ɺ ফඅిྗΛ཈͍͑ͯΔɻ ͜ͷAIνοϓͷݚڀʹ5೥൒ܞΘ͍ͬͯ·͢ 10
11. ©2024 IBM Corporation IBM NorthPole • 2018೥͔ΒIBM ResearchͰݚڀ։ൃ ͞ΕͨAIਪ࿦ઐ༻νοϓ • 2023೥10݄ʹScienceࢽʹ࿦จܝࡌ • σʔλͷҠಈʹ͔͔Δίετʹண໨͠ɺ AIͷԋࢉʹ͓͚ΔσʔλҠಈΛۃྗ ࡟ݮͨ͠ • ଟ͘ͷAIϞσϧͰɺGPUʹൺ΂ͯ ѹ౗తͳ௿ফඅిྗΛ࣮ݱ • ҏ౻͸ϩδοΫઃܭɺγϛϡϨʔγϣϯɺ ফඅిྗͷݕূͳͲΛ୲౰ 11
12. ©2024 IBM Corporation ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟͱͦͷϘτϧωοΫ ϝϞϦ ΞΫηεͷน CPU ϝϞϦ ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟ 12
13. ©2024 IBM Corporation ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟͱͦͷϘτϧωοΫ ϝϞϦ ΞΫηεͷน CPU ϝϞϦ ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟ ·ͣɺը૾σʔλͷ࠷ॳͷ෦ ෼ͱχϡʔϥϧωοτϫʔΫ ͷ࠷ॳͷ෦෼Λ͖࣋ͬͯͯॲ ཧ͢Δ 13
14. ©2024 IBM Corporation ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟͱͦͷϘτϧωοΫ ϝϞϦ ΞΫηεͷน CPU ϝϞϦ ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟ ࣍ʹɺχϡʔϥϧωοτϫʔ Ϋͷ࣍ͷ෦෼Λ͖࣋ͬͯͯɺ ࠷ॳͷॲཧ݁Ռͱ߹Θͤͯॲ ཧ͢Δ 14
15. ©2024 IBM Corporation ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟͱͦͷϘτϧωοΫ ϝϞϦ ΞΫηεͷน CPU ϝϞϦ ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟ ը૾σʔλͷ࠷ॳͷ෦෼ͷॲ ཧ͕શͯऴΘͬͨΒɺը૾σ ʔλͷ࣍ͷ෦෼ͱɺχϡʔϥ ϧωοτϫʔΫͷ࠷ॳͷ෦෼ Λ͖࣋ͬͯͯॲཧ͢Δ 15
16. ©2024 IBM Corporation ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟͱͦͷϘτϧωοΫ ϝϞϦ ΞΫηεͷน ϑΥϯɾϊΠϚϯɾϘτϧωοΫ CPU ϝϞϦ CPU͸χϡʔϥϧωοτϫʔΫͱσʔλΛ ϝϞϦ͔Β͖࣋ͬͯͯॲཧ͢Δඞཁ͕͋Δ ϑΥϯɾϊΠϚϯɾΞʔΩςΫνϟ 16
17. ©2024 IBM Corporation IBM NorthPoleͰͷχϡʔϥϧωοτϫʔΫͷਪ࿦ԋࢉ cat 0.9 NorthPole͸νοϓ֎ͷϝϞϦΛඞཁͱͤͣɺϑΥϯɾϊΠϚϯɾϘτϧωοΫ͕ଘࡏ͠ͳ͍ ೴ͷΑ͏ʹɺχϡʔϥϧωοτϫʔΫશମ͕νοϓͷதͰ࣮ߦ͞ΕΔ 17
18. ©2024 IBM Corporation NorthPoleͷΞʔΩςΫνϟ֓ཁ ԋࢉث ϝϞϦ TPU NorthPole GPU (A100) CPU (Zen 3) 10ԯτϥϯδελ 18
19. ©2024 IBM Corporation NorthPoleͷΞʔΩςΫνϟ • ϝϞϦͱԋࢉث͕1ͭͷίΞͷத Ͱີ઀ʹ഑ஔ͞Ε͍ͯΔ • χϡʔϥϧωοτϫʔΫͷσʔλ ͸ɺ֤ίΞͷϝϞϦ಺ʹϩʔυ ͞ΕΔ • νοϓશମͰ͸192MBͷϝϞϦΛ ࣋ͭ • χϡʔϥϧωοτϫʔΫͷԋࢉʹ ಛԽͨ͠ԋࢉثΛ࣋ͭ • 8/4/2ϏοτͷྔࢠԽΛαϙʔτ 19 ϢχϑΝΠυϝϞϦ ʢॏΈɺϕΫτϧɺ ϓϩάϥϜʣ ॏΈ όοϑΝ ෦෼࿨ όοϑΝ ϓϩάϥϜ όοϑΝ ؅ཧ ϕΫτϧ ԋࢉث ੵ࿨ ԋࢉث ϢχϑΝΠυϝϞϦ ʢॏΈɺϕΫτϧɺ ϓϩάϥϜʣ ॏΈ όοϑΝ ෦෼࿨ όοϑΝ ϓϩάϥϜ όοϑΝ ؅ཧ ϕΫτϧ ԋࢉث ੵ࿨ ԋࢉث 1ίΞ 2 x 2 ίΞߦྻ ໛ࣜਤ ԋࢉ ϝϞϦ ؅ཧ ϨΠΞ΢τ ໛ࣜਤ ωοτϫʔΫɾΦϯɾνοϓ ϨΠΞ΢τ ϑϧνοϓ
20. ©2024 IBM Corporation NorthPoleͷιϑτ΢ΣΞελοΫ 20 NorthPole SDK NorthPole Validator NorthPole Runtime OS, Libraries RedHad Enterprise Linux NorthPole Compiler PyTorch python OpenCV podman OpenShift PyTorch Λ࢖͍ɺ GPUΛ࢖ͬͯNorthPoleʹ޲͚ͯ χϡʔϥϧωοτϫʔΫͷྔࢠԽΛߦ͏ ֶशͨ͠χϡʔϥϧωοτϫʔΫΛ Compiler Λ࢖ͬͯ NorthPoleͰ࣮ߦՄೳͳܗʹม׵͢Δ ίϯύΠϧͨ͠χϡʔϥϧωοτϫʔΫΛ Runtime API Λ࢖ͬͯNorthPole্ʹల։͠ɺೖྗσʔλͱ݁ Ռͷ΍ΓऔΓΛߦ͏ Validator Λ࢖ͬͯίϯύΠϧͨ͠χϡʔϥϧωοτ ϫʔΫͷݕূΛιϑτ΢ΣΞͰߦ͏͜ͱ͕Ͱ͖Δ ֶश ਪ࿦ NorthPole Runtime NorthPole Validator NorthPole Compiler PyTorch
21. ©2024 IBM Corporation NorthPoleͷը૾ೝࣝϞσϧʢResNet-50ʣʹ͓͚Δੑೳ ిྗޮ཰ (frames/joule) ۭؒޮ཰ (frames/ඵ)/10ԯτϥϯδελ NorthPoleΑΓߴ౓ͳ൒ಋମٕज़Ͱ੡଄͞ΕͨνοϓΛؚΊͨ ଞͷΞʔΩςΫνϟશͯͷੑೳΛNorthPole͸྇կ͍ͯ͠Δ • NorthPoleͱಉ͡12nmͰ੡଄ ͞ΕͨGPU (V100) ͱൺֱͯ͠ɺ 25ഒͷిྗޮ཰ͱ5ഒͷۭؒޮ ཰Λୡ੒ • 4nmͰ੡଄͞ΕͨGPU (H100) ͱ ൺֱͯ͠ɺ5ഒͷిྗޮ཰Λୡ੒ 21
22. ©2024 IBM Corporation NorthPoleͷଞͷχϡʔϥϧωοτϫʔΫʹ͓͚Δੑೳ ෺ମݕग़ϞσϧʢYolo-v4ʣ ݴޠϞσϧʢBERT-Baseʣ ଞͷχϡʔϥϧωοτϫʔΫʹ͓͍ͯ΋ɺಉ༷ͷిྗޮ཰ͱۭؒޮ཰Λ࣮ݱ 22
23. ©2024 IBM Corporation NorthPoleͷϢʔεέʔεྫ • ௿ফඅిྗ͕ٻΊΒΕΔΤοδͰͷAIར༻ • αʔόʔͰͷAIਪ࿦ΛΑΓ௿ফඅిྗͰ࣮ߦ • ࣗಈӡసʹ͓͚ΔΧϝϥը૾ͷ෺ମݕग़ • υϩʔϯͷΧϝϥը૾ͷ෺ମݕग़ • ߤۭը૾ʹ͓͚Δ෺ମݕग़ • … ϦαʔνϓϩτλΠϓ 23
24. ©2024 IBM Corporation ंࡌΧϝϥΛ૝ఆͨ͠෺ମݕग़ϞσϧʢYolo-v4ʣͷ࣮ߦྫ 24
25. ©2024 IBM Corporation Thank you! 25