Upgrade to Pro — share decks privately, control downloads, hide ads and more …

170526_LeapMindイベント_三室

 170526_LeapMindイベント_三室

2017/05/26(金) 19:00〜
[LeapMind主催]FPGAを用いたディープラーニングのハードウェアアクセラレーションの前半partの資料です。

Yuki Mimuro

May 26, 2017
Tweet

More Decks by Yuki Mimuro

Other Decks in Business

Transcript

  1. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE ձ໊ࣾ LeapMindגࣜձࣾ ઃཱ 2012೥12݄25೔ ࢿຊۚ 206,984,639ԁ ୅ද দా૯Ұ

    ओཁגओ ҏ౻஧ςΫϊϩδʔϕϯνϟʔζɺVisionnaire Ventures(ถ)ɺArchetype* Ventures(೔) ैۀһ਺ 35ਓ ॴࡏ஍ ౦ژ౎ौ୩۠ौ୩3-15-3ɹ౔԰Ϗϧ3F ࣄۀ಺༰ σΟʔϓϥʔχϯάʹؔΘΔγεςϜߏங͓Αͼࢧԉ ձࣾ֓ཁ
  2. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE -FBQ.JOEͷԊֵ ౦ژ౎໨ࠇ۠ʹͯ -FBQ.JOEגࣜձࣾ૑ۀ ೥݄ ೥݄ ,%%*ແݶϥϘʹͯ/FX-JGFTUZMF৆ Λड৆ ೥݄

    "SDIFUZQF 7FOUVSFT*ODΑΓ ສԁͷ$#ͰͷࢿۚௐୡΛ ͓͜ͳ͍·ͨ͠ ೥݄ ςϨϏ౦ژͷʮϫʔϧυϏδωεα ςϥΠτʯʹͯ঺հ͞Ε·ͨ͠ ೥݄ γϦʔζ"ϥ΢ϯυͰ૯ֹԯ ສԁͷୈࡾऀׂ౰૿ࢿΛൃද ೥݄ ʰ'PSCFT+BQBOʱ೥݄߸ͷ ʮੈք͕஫໨ʂظ଴ͷελʔτΞο ϓ೔ຊͷࣾʯʹબग़͞Ε·ͨ͠ ೥݄ גࣜձࣾ/55σʔλͱ5XJUUFSࣾͷ શྔσʔλղੳʹ͓͍ͯۀ຿ఏܞ ͍ͨ͠·ͨ͠
  3. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE -FBQ.JOEͷಛ௃ ϋʔυ΢ΣΞݚڀ লిྗߴޮ཰Λ໨ࢦͨ͠ ճ࿏ઃܭ χϡʔϥϧωοτϫʔΫΛճ࿏ଆͰ෺ ཧతʹ༻ҙ͢Δ͜ͱͰখ͍͞νοϓͰ ΋σΟʔϓϥʔχϯάͷܭࢉ͕ग़དྷΔ ಠࣗΞʔΩςΫνϟΛݚڀ

    ιϑτ΢ΣΞݚڀ σΟʔϓϥʔχϯάͷϞσϧ ਫ਼౓ͷ޲্ܭࢉྔѹॖ Ϟσϧࣗମͷਫ਼౓޲্ͱಉ࣌ʹɺܭࢉ ྔΛѹॖ͠খ͞ͳܭࢉ؀ڥͰ΋σΟʔ ϓϥʔχϯάͷҖྗΛ࠷େݶൃشͤ͞ ΔಠࣗΞϧΰϦζϜͷݚڀ
  4. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE ಠٕࣗज़Λ࣠ʹ ػց͕࣮ੈքೝࣝ Ұൠৄࡉ෺ମೝࣝ ࣮ੈքͷ෺ମʹରͯ͠ɺػց͕ ͦΕ͕Կ͔ೝࣝ͢Δ͜ͱ͕Ͱ͖Δ Α͏ʹͳΓ·͢ ෺ମݕग़ ೋ஋Խχϡʔϥϧωοτ

    ڧԽֶश ࣌ܥྻσʔλղੳ Ϟσϧѹॖ ಠࣗͷίάχςΟϒٕज़ ͋Δγʔϯͷத͔ΒɺԿ͕Ͳ͜ʹ ͋Δͷ͔Λػց͕೺Ѳ͢Δ͜ͱ͕ Ͱ͖ΔΑ͏ʹͳΓ·͢ ܭࢉΛѹॖ͠ܭࢉࣜΛܰྔԽ͠ɺ ඇৗʹখ͞ͳϝϞϦফඅͰσΟʔ ϓϥʔχϯάΛಈ͔͠·͢ χϡʔϥϧωοτϫʔΫΛͱ ͷΈͰදݱ͠௚͠ɺߴ଎Խͱলϝ ϞϦԽΛ࣮ݱ͠·͢ ࣌ؒ࣠Λߟྀͨ͠χϡʔϥϧωο τͰߴਫ਼౓ͷ෼ੳΛߦ͍·͢ ػց͕ঢ়گΛཧղ͠ͳ͕Βɺ࠷ద ͳղ౴ΛࣗΒݟ͚͍ͭͯ͘͜ͱ͕ Ͱ͖ΔΑ͏ʹͳΓ·͢
  5. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE ιϦϡʔγϣϯ׆༻ࣄྫ *OEVTUSZ $POTVN FS 8FC ɾࣗಈӡసϓϩδΣΫτ ɾנ᛽ఫڈϩϘοτ ɾൃిॴҟৗݕ஌

    ɾ৯඼ҟ෺ݕ஌ ɾυϩʔϯΛ࢖ͬͨݐ෺఺ݕ ɾ"*Ոి ɾ"*ϗʔϜ ɾ৯ࣄ؅ཧఏҊ ɾ೶ۀ ɾ4/4ͷτϨϯυղੳ ɾߪങσʔλղੳ ɾҩྍσʔλղੳ
  6. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE ا ը ࢧ ԉ اը ݕ౼ ϓ Ϩ

    Ϧ α ồ ν ཁ ݅ ఆ ٛ Ϧαʔν ઃܭɾ࣮૷ ֶशɾνϡʔχϯά ಋೖ ӡ༻ ཁ݅ఆٛ Ϧ α ồ ν ֶ श σ ồ λ ࡞ ੒ Ϟ σ ϧ ઃ ܭ Ϟ σ ϧ ࣮ ૷ Ϟ σ ϧ ֶ श Ϟ σ ϧ ݕ ূ ύ ϥ ϝ ồ λ ồ ν ϡ ồ χ ϯ ά ಋ ೖ ࢧ ԉ ӡ ༻ ࢧ ԉ Ϟ σ ϧ ͷ ࠶ ֶ श ݕূ༻ΠϯλϑΣʔε࡞੒ ӡ༻ઃܭɾӡ༻ϚχϡΞϧ࡞੒ %FFQ-FBSOJOHγεςϜ։ൃͷྲྀΕ
  7. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE 45&1͝ͱʹߦ͏͜ͱ ಺༰ Ϧαʔν Ϟσϧߏங ݕূγεςϜӡ༻ ໨త ࣮ݱ͢ΔͨΊͷΞϓϩʔνΛ໛ࡧ ֩ͱͳΔ//ϞσϧΛߏங

    ࣮γεςϜӡ༻Ͱ͖Δମ੍Λߏங ϝϯόʔ 3FTFBSDIFS %FFQ-FBSOJOH&OHJOFFS 3FTFBSDIFS %FFQ-FBSOJOH&OHJOFFS 8FC&OHJOFFS ํ๏ ࿦จͳͲ͔Βద੾ͳΞϓϩʔνΛ ୳͠ग़͠ɺTBNQMFDPEFΛॻ͍ͯ ؆୯ͳ࣮ݧΛߦ͍ɺQSJPSJUZΛܾΊΔ ˙؀ڥ 1ZUIPO9PS9  EFQFOETPOQSPKFDU  $ TUBOEBSE FH 5FOTPSqPXWFS ֶश͸෺ཧαʔόʔͱ"84 "[VSF ίʔυ؅ཧ͸(JU)VC ˙؀ڥ ɾόοΫΤϯυ ݴޠɿ1ZUIPOɺQIQͳͲ '8ɿqBTLɺMBSBWFMɺ$BLF1)1ͳͲ ɾϑϩϯτΤϯυ ݴޠɿ/PEF+4 3FBDU+4 )5.- $44 04ɿ։ൃ࣌͸ࣗ༝ɺσϓϩΠઌ͸-JOVY %#ɿ.Z42-
  8. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE Labeled data ML algorithm Input data Trained network

    prediction Training Prediction Training is usually off-line Needs a lot of time and power Usually requires big GPU. (some new ASIC chips are expected soon) Has to be on-line Time and power constraint Usually less computation than training. Currently GPU based. FPGA use expected in the future cloud edge %FFQ-FBSOJOHPOFEHF
  9. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE ։ൃ؀ڥܥ౷ਤ ౷߹؀ڥ ϓϥοτϑΥʔϜ %BUB %BUB %BUB 'SBNFXPSL %BUB

    ਺ֶతʹਖ਼͍͠ܗ ͰܭࢉํࣜΛখ͞ ͘͢ΔϥΠϒϥϦ NJOJ TFHOFU ౷߹؀ڥ ϓϥοτϑΥʔϜ νοϓʹ χϡʔϥϧωοτ Λߏங͢Δ NJOJ $// NJOJ 3$// NJOJ "MFYOFU NJOJ 7((OFU NJOJ 3FTOFU Ոి༻ νοϓ ޻৔༻ νοϓ ࣗಈं༻ νοϓ υϩʔϯ༻ νοϓ ϩϘοτ༻ νοϓ 4PGU -BZFS )BSE -BZFS
  10. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE %FFQ-FBSOJOH͕޿·Δϓϩηε *O *#. 1$ $IJQ 4PGU  *O

    -.#MBDLTUBS *OUFSGBDF $IJQ 4PGU  ΞϓϦέʔγϣϯɾϋʔυ΢ΣΞɾσΟʔϓϥʔχϯά શͯͷύʔπ͕όϥόϥͳͷͰɺ౷߹ͨ͠αʔϏε͕ඞཁʹͳΔ
  11. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE ࢓ࣄ಺༰ ɾΫϥΠΞϯτ͔Βड͚ͨ%FFQ-FBSOJOHʹؔ͢Δ։ൃҊ݅ͷऔ૊Έ ɾ%FFQ-FBSOJOH .BDIJOF-FBSOJOH΍$PNQVUFS7*TJPO ؔ܎ͷϥ ΠϒϥϦɾϑϨʔϜϫʔΫͷௐࠪɺ࣮૷ɺվྑ ඞਢεΩϧ ɾ1ZUIPOɺ$

    ͷߴ͍ϓϩάϥϛϯάεΩϧ ɾߦྻ΍ภඍ෼ͷ਺ֶత஌ࣝ ׻ܴεΩϧ ɾ৘ใ޻ֶܥͷֶҐͷอ༗ ɾ"OESPJEJ04΍ɺ3BTQCFSSZ1JͳͲ૊ΈࠐΈػث΁ͷϓϩάϥϜ ࣮૷ͷܦݧ ٻΊΔਓ෺૾ ɾϓϩάϥϛϯά͢Δ͜ͱʹ೤தͰ͖Δ ɾࣗ෼ࣗ਎ͷϓϩάϥϛϯάྗ޲্ͨΊʹɺϑΟʔυόοΫΛड͚ೖ ΕΔ͜ͱ͕Ͱ͖Δ ɾ࠷ઌ୺ͷٕज़ʹର͢ΔڧΈڵຯ͕͋Δʢ୳ڀ৺ʣ ɾ੒Ռ૑ग़ʹ޲͚ͯɺνʔϜͱͯ͠ઓ͏͜ͱʹތΓΛ͍࣋ͬͯΔ ɾৗʹ৽͍͠खஈʹ໨Λ͚ͭɺ࣮ݧݕূ͢Δ͜ͱ͕Ͱ͖Δ ։ൃ؀ڥ $B⒎F΍5FOTPS'MPXΛओʹ࢖༻ ίʔυن໿ʹ৐ͬऔ͍ͬͯΕ͹ΤσΟλͳͲ͸ࣗ༝ %FFQ-FBSOJOHΤϯδχΞ
  12. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE 8FCόοΫΤϯυΤϯδχΞ ࢓ࣄ಺༰ ɾࣗࣾαʔϏεϓϩμΫτͷ։ൃ ɾ৽نٕज़ͷௐ͓ࠪΑͼݚڀ։ൃͱͦͷڞ༗ ɾαʔϏεʹ߹Θͤͯɺ্ཱ͔͛Βӡ༻·Ͱશͯͷ؀ڥ΍ίʔυΛߏங ඞਢεΩϧ ɾ8FCΞϓϦέʔγϣϯͷ։ൃɾӡ༻ܦݧ ɾ-JOVY6OJYʹؔ͢Δ஌ࣝɾܦݧ

    ɾ1ZUIPO 1)1ͳͲͷϓϩάϥϛϯάܦݧ ׻ܴεΩϧ ɾΞϧΰϦζϜʹؔ͢Δ஌ࣝ ɾϑϨʔϜϫʔΫʹؔ͢Δ஌ࣝɾܦݧ ɾσʔλϕʔεʹؔ͢Δ஌ࣝɾܦݧ ٻΊΔਓ෺૾ ɾ৽ͨͳαʔϏεʹର͢Δٕज़બఆ͕Ͱ͖Δ ɾଞऀΛר͖ࠐΈͳ͕Β࢓ࣄΛߦ͏͜ͱ͕Ͱ͖Δ ɾ࠷ઌ୺ͷٕज़ʹର͢ΔڧΈڵຯ͕͋Δʢ୳ڀ৺ʣ ɾ৽ͨͳٕज़ΛࣗΒ࣮ݧɾݕূ͠ɺࣗࣾαʔϏεʹ׆͔͢ྗ͕͋Δ ɾૉ௚Ͱ৽ֶ͘͠Ϳ͜ͱΛԀΘͳ͍
  13. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE 8FCϑϩϯτΤϯυΤϯδχΞ ࢓ࣄ಺༰ ɾσβΠϯɺاը͔ΒͷཁٻʹԠͨ͡)5.-$44+BWB4DSJQUQIQʹΑΔ΢Σ ϒϑϩϯτΤϯυͷ࣮૷ ɾϞοΫΞοϓ΍ϓϩτλΠϓͷ࡞੒ͱ࣮ূ ɾ69Λߟྀͨ͠6*ͷఏҊ͓Αͼઃܭ ɾ৽نٕज़ͷௐ͓ࠪΑͼݚڀ։ൃͱͦͷڞ༗ ɾύϑΥʔϚϯεվળ΍඼࣭ҡ࣋ͷͨΊͷࢪࡦ࣮ࢪ

    ɾ6*ίϯϙʔωϯτ΍ϥΠϒϥϦͷ੍࡞ ඞਢεΩϧ ɾ)5.-$44+BWB4DSJQU΁ͷઐ໳తͳਂ͍஌ࣝ ɾ1$΍εϚʔτϑΥϯͷ΢ΣϒΞϓϦέʔγϣϯͷ։ൃܦݧ ɾΞΫηγϏϦςΟɾϢʔβϏϦςΟʹؔ͢Δ஌ࣝ ׻ܴεΩϧ ɾ$44ϓϦϓϩηοαʹؔ͢Δ஌ࣝ ɾJ04ɺ"OESPJEͳͲͷωΠςΟϒΞϓϦέʔγϣϯͷ։ൃܦݧ ɾ1ZUIPO ɺ+BWBͳͲͷαʔόʔαΠυݴޠͰͷ։ൃܦݧ ٻΊΔਓ෺૾ ɾ৽ͨͳαʔϏεʹର͢Δٕज़બఆ͕Ͱ͖Δ ɾ69Λߟ্͑ͨͰXFCσβΠϯΛ࡞੒Ͱ͖Δ ɾ࠷ઌ୺ͷٕज़ʹର͢ΔڧΈڵຯ͕͋Δʢ୳ڀ৺ʣ ɾ৽ͨͳٕज़ΛࣗΒ࣮ݧɾݕূ͠ɺࣗࣾαʔϏεʹ׆͔͢ྗ͕͋Δ ɾૉ௚Ͱ৽ֶ͘͠Ϳ͜ͱΛԀΘͳ͍
  14. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE ݚڀऀϦαʔνϟʔ ࢓ࣄ಺༰ ɾ௕ظతͳࢹ఺Ͱɺٕज़ͷϩʔυϚοϓΛͱΒ͑ɺݚڀΛߦ͍ࣗࣾಠ ࣗͷٕज़ΛੜΈग़͢ ɾੜΈग़ͨ͠ಠࣗͷٕज़ΛɺࠃࡍతͳֶձͳͲͰൃද͠ɺٕज़Λੈ ͷதʹؐݩ͢Δ ඞਢεΩϧ ɾػցֶशҨ఻తΞϧΰϦζϜίϯϐϡʔλʔϏδϣϯ౳ͷ෼໺Ͱ

    म࢜ Ϛελʔ Ҏ্ͷֶҐΛ༗͍ͯ͠Δ͜ͱ ׻ܴεΩϧ ɾࠃࡍֶձ΁ͷ࿦จܝࡌ࣮੷ ɾຽؒݚڀػؔ౳Ͱͷ࣮຿ܦݧ ٻΊΔਓ෺૾ ɾ৽͍ٕ͠ज़Λ௥ٻ͢Δ͜ͱ͕ͱ͜ͱΜ޷͖ ɾ࿦จΛग़͢ͳͲɺੈͷதʹର͢ΔΠϯύΫτΛॏࢹ͢Δ ɾࣗ෼ࣗ਎ͱͯ͠ͷݚڀςʔϚΛ͍࣋ͬͯΔ ɾ৽͍ٕ͠ज़ʹର͢ΔΞϯςφ͕Ӷཱ͍ͬͯ͘Δ ɾ࠷େ੒ՌΛग़ͨ͢Ίʹɺ஌ࣝͷڞ༗ΛνʔϜͰߦ͏͜ͱ͕Ͱ͖Δ ։ൃ؀ڥ $B⒎F΍5FOTPS'MPXΛओʹ࢖༻ ίʔυن໿ʹ৐ͬऔ͍ͬͯΕ͹ΤσΟλͳͲ͸ࣗ༝
  15. ˜-FBQ.JOE *OD"MMSJHIUTSFTFSWFE '1("ΤϯδχΞ ࢓ࣄ಺༰ ɾ%FFQ-FBSOJOH ਂ૚ֶश Λߴ଎͔ͭলిྗͰఏڙ͢Δج൫ͷઃܭ ฐࣾϓϩμΫτ#MBDL4UBS ඞਢεΩϧ ɾߴҐ߹੒

    )-4 ʹؔ͢Δઐ໳తͳ஌ࣝ ɾ%FFQ-FBSOJOHʹؔ͢Δجૅతͳ஌ࣝ ɾ7FSJMPH)%-·ͨ͸7)%-ΛؚΉ35-։ൃͷ஌ࣝ ׻ܴεΩϧ ɾ-JOVYͷ஌ࣝ ɾ$ݴޠͷϓϩάϥϛϯάεΩϧ ɾ૊ΈࠐΈϚΠίϯपลճ࿏ઃܭͷ஌ࣝ ɾϚΠίϯɺ'1(" $1-% Λ࢖ͬͨࣗ࡞ϋʔυ΢ΣΞͷ੡࡞ܦݧ ٻΊΔਓ෺૾ ɾੈքతͳτϨϯυΛৗʹଊ͑ͭͭɺࣗࣾʹ࠷దͳٕज़ΛಋೖͰ͖Δ ɾ'1("ʹݶΒͣɺϋʔυ΢ΣΞશൠʹର͢Δ஌ࣝΛ͍࣋ͬͯΔ ɾϋʔυ͚ͩͰ͸ͳ͘ɺιϑτʹؔ͢Δ஌ࣝ΋͍࣋ͬͯΔ ɾଞ෦໳ͱͷ࿈ܞͷͨΊʹɺԁ׈ͳίϛϡχέʔγϣϯΛऔΔ ։ൃ؀ڥ ɾ7JWBEP  ɺ7JWBEP)-4   ɾ%FHJMFOUࣾͷ;ZCPɺ5SFO[ࣾͷϘʔυ