Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Tensorコアを使った PyTorch の高速化について
Search
fam_taro
April 08, 2019
Technology
4.1k
4
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
Tensorコアを使った PyTorch の高速化について
fam_taro
April 08, 2019
More Decks by fam_taro
See All by fam_taro
NeRFの概要と 派生系についてのふんわり紹介
fam_taro
3
4.4k
実践 PyTorch Lightning (2019/11/30 分析コンペLT会 #1)
fam_taro
3
4.6k
Paper:ShapeMask
fam_taro
0
94
Summary: Objects as Points
fam_taro
0
3.3k
Paper-Survey: Objects as Points
fam_taro
0
2.4k
Sequence to Sequence Learning with Neural Networks
fam_taro
1
1.1k
Other Decks in Technology
See All in Technology
紙 to デジタル
ichien178
0
360
Point Cloud as a Foreign Language for Multi-modal Large Language Model
takmin
0
300
DMMブックスのNext.js化を加速させるAI活用 / Migrating DMM Books to Next.js with AI
kentarom
0
180
All About Sansan – for New Global Engineers
sansan33
PRO
1
1.5k
「ミスを許さない手順書」を作ってみた 〜 個人的にはこれ以上できることはあまりなさそう/20260827-ssmjp-operation-procedure-update
opelab
13
9k
AWSとGitHub Actionsの責任境界と 組織で安全に使用する取り組み
nealle
1
220
When Token Pruning is Worse than Random: Understanding Visual Token Information in VLLMs
sansantech
PRO
0
130
Claude Codeの体系的な理解と知識のフック
oikon48
8
6k
GopherCon @シアトル に行ってきました
logica0419
0
390
Master Dataグループ紹介資料
sansan33
PRO
1
4.8k
EMの役割で 変わったこと・変わらなかったこと
sansantech
PRO
1
190
8bit CPU 2026
koba789
7
3k
Featured
See All Featured
Dealing with People You Can't Stand - Big Design 2015
cassininazir
367
27k
What the history of the web can teach us about the future of AI
inesmontani
PRO
1
670
ピンチをチャンスに:未来をつくるプロダクトロードマップ #pmconf2020
aki_iinuma
128
56k
Collaborative Software Design: How to facilitate domain modelling decisions
baasie
1
290
Navigating the moral maze — ethical principles for Al-driven product design
skipperchong
2
500
Practical Orchestrator
shlominoach
191
12k
How to make the Groovebox
asonas
2
2.4k
BBQ
matthewcrist
89
10k
Ten Tips & Tricks for a 🌱 transition
stuffmc
0
180
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
141
35k
Data-driven link building: lessons from a $708K investment (BrightonSEO talk)
szymonslowik
1
1.3k
Keith and Marios Guide to Fast Websites
keithpitt
413
23k
Transcript
'BTUFS1ZUIPO.FFUVQ-5ձ 5FOTPSίΞΛͬͨ 1Z5PSDIͷߴԽ !GBN@UBSP
"HFOEB 5FOTPSίΞ JTԿ 1Z5PSDIͰ 5FOTPSίΞ͏ʹ ਪ ༧ଌ
ͰͬͯΈͨ .%FU ֶशͰ 5FOTPSίΞΛ͍͍ͨ߹
5FOTPSίΞ JTԿ CJUුಈখ '1 ͷߦྻԋࢉʹಛԽͨ͠ܭࢉ༻ίΞ 1Z5PSDI ͷਂֶशϑϨʔϜϫʔΫΛ͏ͱɺσϑΥϧτͰ CJUුಈখ
ʢ'1 Ͱॾʑͷܭࢉ͕ߦΘΕΔ /7*%*"(16ͷ͏ͪ 7PMUBੈҎ߱ͷҰ෦ʹ͍ͬͯΔ 5VSJOHͰ 359͡Όͳ͍ͱͬͯͳ͍ (59 5Jʹ͍ͬͯͳ͍ 1BTDBM .BYXFMM ,FQMFSʹແ͍ ແ࣊൵ (16ྫ 5*5"/7 (F'PSDF359γϦʔζ 5J 5*5"/359 5FTMB7 ($1ͰࢼͤΔ ࠓޙΑΓීٴ͢Δͱݸਓతʹࢥͬͯ·͢ ࢀߟϦϯΫ IUUQTXJLJXJLJKQOWJEJBWHB(16&##&"&#&""&"" IUUQTKBXJLJQFEJBPSHXJLJ/7*%*"@(F'PSDF https://www.nvidia.com/content/apac/gtc/ja/pdf/2017/1040.pdf
5FOTPSίΞ JTԿ 5FOTPSίΞΘͳ͍ͱ '1ʹͯ͠ ഒఔ 5FOTPSίΞ͏ͱ '1ʹͨ͠Β ഒ
Ҏ্ʂ ͨͩ͠͏ͨΊʹ੍͕ଟ͍ʜ $6%"Ҏ߱ DV%// Ҏ߱ '1Ͱѻ͏͜ͱΛ໌ه ϥΠϒϥϦͷίϚϯυͰ $POWͷೖྗɾग़ྗνϟϯωϧ͕ ͷഒ શͯͷ͕ରԠ͍ͯ͠ΔΘ͚Ͱͳ͍ Ή͠ΖҰ෦͔͠ ͋͘·Ͱݱ࣌ͳͷͰࠓޙ؇͞ΕΔ͔ʜ https://www.slideshare.net/NVIDIAJapan/chainer-tensor-fp16
1Z5PSDIͰ 5FOTPSίΞ͏ʹ '1Λ͏͜ͱΛ໌ه͢ΕϑϨʔϜϫʔΫ͕উखʹ ͬͯ͘ΕΔ ͜ͱ͕ଟ͍ ࠷ۙͷόʔδϣϯʹ͠ͳ͍ͱ͍͚ͳ͍͕ʜ 1Z5PSDIͰʜ
.PEFMͱ *OQVUʹର͠ lIBMG zΛ͚Δ ਫ਼ʹ͢Δͱ͍͏ҙຯ '1ʹ͢Δ 0VUQVU '1ͱݶΒͳ͍ͷͰҙ $POWͷೖྗͱग़ྗνϟϯωϧΛ ͷഒʹ͢Δ ຯʹ੍͖͍ͭ 1SFUSBJONPEFMͩͱ͏·͑͘ͳ͍έʔε͕ग़ͯ͘Δ ໌ࣔతʹ 5FOTPSίΞΛ͏Α͏ʹͰ͖ͳ͍ Ά͍ʁ
ਪ ༧ଌ ͰͬͯΈͨ .%FU .%FUͱ จIUUQTRJKJF[IBPHJUIVCJPJNHTNEFUQEG ࣮IUUQTHJUIVCDPNRJKJF[IBP.%FU
࠷ۙग़͖ͯͨݕग़Ϟσϧ :0-0WΑΓͯ͘ਫ਼͕ߴ͍ Β͍͠ ࣗͷσʔλͰֶशશ͘Ͱ͖ͯͳ͍Ͱ͕͢! ࣮ݧ݅ $0$0EBUBTFU %FUFDUJPOͰϝδϟʔͳͭ ֶशࡁΈϞσϧ Y Λ༻ ެࣜϦϙδτϦʹ͋Δ UFTUQZ Λ༻ NPEFM JOQVU ͷ࣌ؒͱޙॲཧ /.4 ͷ࣌ؒΛܭଌ ͨ࣌ؒ͠Λͬͯ '14Λग़͢ ߴ͍΄Ͳྑ͍ ຕͷσʔλʹରͯ͠ਪ ༧ଌ Λͯ͠ฏۉ࣌ؒΛͱΔ ݩจ ຕͰऔ͍ͬͯͨ Ϟσϧͷਫ਼ N"1
ਪ ༧ଌ ͰͬͯΈͨ .%FU ࣮ߦ࣌ؒ 5FOTPSίΞΛ͍ͬͯΔ͔֬ೝ͢ΔͨΊʹͬͨίϚϯυ UJNFUJNF ؔ
(16བྷΉͨΊԼهͷํ͕ྑͦ͞͏Ͱ͕͢ࠓճ UJNF ؔͰ࣌ؒͱΓ·ͨ͠! UPSDIDVEB&WFOU FOBCMF@UJNJOH5SVF QSPGJMFS QZUIPONUPSDIVUJMTCPUUMFOFDL UFTUQZ ddd 1Z5PSDI ͷఏڙ͢Δ QSPGJMFS ݕग़ͷޙॲཧ ࠓճ /.4 ͕ϘτϧωοΫʹͳ͍ͬͯͳ͍͔֬ೝ͢ΔͨΊʹ༻ OWQSPG QZUIPOddd /7*%*"ͷఏڙ͢Δ QSPGJMFSίϚϯυɻ5FOTPSίΞ͍ͬͯΔ͔ݟΕΔ OWDD ddd ະ༻ OWQSPG ͷίϚϯυΛ (6*ͰϦονʹݟΕΔΒ͍͠ɻͨͩ͠ΞϓϦαΠζσΧ͍ (#
ਪ ༧ଌ ͰͬͯΈͨ .%FU ࣮ߦ݁Ռ (16 N"1 ਫ਼
%FUFDUUJNF QFS JNBHF<NT> /NT UJNF 1FS JNBHF<NT> 5PUBM<NT> '14 ݩจ ϦϙδτϦ্ͷ 5JUBO9 1"4$"- 1Z5PSDI ެࣜࢦఆόʔδϣϯ 5JUBO7 1Z5PSDI '1 5JUBO7 1Z5PSDIQPTU 5JUBO7 1Z5PSDIQPTU '1 5JUBO7
ਪ ༧ଌ ͰͬͯΈͨ .%FU ࣮ߦ݁Ռ (16 N"1 ਫ਼
%FUFDUUJNF QFS JNBHF<NT> /NT UJNF 1FS JNBHF<NT> 5PUBM<NT> '14 ݩจ ϦϙδτϦ্ͷ 5JUBO9 1"4$"- 1Z5PSDI ެࣜࢦఆόʔδϣϯ 5JUBO7 1Z5PSDI '1 5JUBO7 1Z5PSDIQPTU 5JUBO7 1Z5PSDIQPTU '1 5JUBO7 l1Z5PSDIʹͯ͠z ͔ͭ l'1͏zͱ࠷͘ͳΔ ˠ ࠷ॳͷ ഒ ͔Ζ͏ͯ͡ݩจͷ '14Λ͑ͨ ༧ଌ࣌ͷΈͳΒ '1ʹͯ͠ ਫ਼มΘΒͳ͍ 1Z5PSDIͷόʔδϣϯ্͛ͨΒ͘ ͳΔ͠ɺ'1ʹͯ͘͠ͳΔ
ֶशͰ 5FOTPSίΞΛ͍͍ͨ߹ ͱΓ͋͑ͣશ෦ '1ʹ͢Ε͍͍ΜͰ͠ΐʁˠ /Pʂ ޯܭࢉ࣌ʹ͔ͳΓӨڹ͢Δ ޯ͕ফ͑Δ߹͋Δ
ˠ ਫ਼͕େ͖͘Լ͕Δ߹͕͋Δ .JYFE1SFDJTJPO5SBJOJOH ࢀߟϦϯΫIUUQTXXXTMJEFTIBSFOFU/7*%*"+BQBODIBJOFSUFOTPSGQ '1ͱ '1ʹΑΔܭࢉΛ NJYֶͨ͠शํ๏ ۩ମతʹҎԼͷΑ͏ͳςΫχοΫ͕ඞཁ ϩεεέʔϦϯά ϩεΛཁॴཁॴͰεέʔϧ͋ͬͯ͠ޯফࣦΛ؇ '1ΣΠτߋ৽ 'PSXBSEͱ #BDLXBSE '1 6QEBUFͰ '1Λ༻ ্هΛશͯࣗͰΔͱେมʂ ֶͼ͋Δͱࢥ͍·͕͢ʜ ˠ BQFYΛ͏ͱൺֱతखܰʹͰ͖Δ IUUQTHJUIVCDPN/7*%*"BQFY /7*%*"͕ఏڙ͢Δ 1Z5PSDI༻ "VUPNBUJD.JYFE1SFDJTJPO ".1 πʔϧ ݩͷίʔυʹର͠ߦ͚ͩ͢Ͱ .JYFE1SFDJTJPO5SBJOJOHͰ͖Δͱͷ͜ͱ ͨͩ͠ JOTUBMM࣌ $6%" 1Z5PSDIͷόʔδϣϯʹؾΛ͚ͭͳ͍ͱ͍͚ͳ͍
3FGFSFODFT <൛ ػցֶशϋʔυΣΞͷ4UBUFPGUIF"SUΛߟ͑Δ d$16 (16 516Λఴ͑ͯd 2JJUB> IUUQTRJJUBDPNBSVUFNBJUFNTGCB
<5SBJOJOH/FVSBM/FUXPSLTXJUI.JYFE1SFDJTJPO /7*%*"> IUUQPOEFNBOEHQVUFDIDPOGDPNHUDUBJXBOQEG @*OUFSOBM
[email protected]
$BSJMMJ@1%''PS4IBSJOHQEG <$IBJOFS Ͱ 5FOTPSίΞ GQ Λ͍͜ͳ͢> IUUQTXXXTMJEFTIBSFOFU/7*%*"+BQBODIBJOFSUFOTPSGQ <$IBJOFSʹ͓͚ΔਂֶशͷߴԽ> IUUQTXXXOWJEJBDPNDPOUFOUBQBDHUDKBQEGQEG <70-5""/%563*/("3$)*5&$563&"/%1&3'03."/$&015*.*;"5*0/ /7*%*"> IUUQTXXXOWJEJBDPNDPOUFOUBQBDHUDKBQEGQEG <5SBJOJOHXJUI.JYFE1SFDJTJPO%FFQ-FBSOJOH4%,%PDVNFOUBUJPO> IUUQTEPDTOWJEJBDPNEFFQMFBSOJOHTELNJYFE QSFDJTJPOUSBJOJOHJOEFYIUNMQZUPSDI <(16༷Ұཡද /7*%*"(F'PSDF 8JLJ > IUUQTXJLJXJLJKQOWJEJBWHB(16&##&"&#&""&""
͓ΘΓ ͋Γ͕ͱ͏͍͟͝·ͨ͠
ิOWQSPG ͷ݁Ռ 1Z5PSDI
ิOWQSPG ͷ݁Ռ 1Z5PSDI