NowMatrixMultiplicationorMatmulcombinesthesesetsbymultiplyingandaddingcorrespondingelements, creating a totallynewmatrixthatrepresentstherelationshipbetweentheoriginalmatricesorfeatures.
Imagineyouhave a matrixofimagesand a matrixoffiltersthatdetectedges.
画像の行列と、エッジを検出するフィルターの行列があるとする。
Matmulcombinesthesematricestoproduce a newmatrixthatrepresentstheedgesineachimage.
Matmulはこれらの行列を結合し、各画像のエッジを表す新しい行列を生成する。
Thispaper, whichyoucanseeonyourscreen, hastakenthesocialmediaofAIbystormbygivingit a newtwist.
画面上でご覧いただけるこの論文は、AIのソーシャルメディアに新風を吹き込み、一世を風靡した。
Before I tellyouindetailwhatexactlythispaperhasdonehere, letmegive a bitmoretechnicaldetailonthisMatmulbecauseit's a pivotalconceptinthiswholescenariotounderstandthisbetter.
Similarly, ifyoulookatthisdiagram, thisprimarilyshowsyouwhatexactlyisthecomparisonwithothermodelsand a lotofotherstuffaroundperformancecomparisonandanalysisofdifferentmodels.
I willalsodropthelinktothispaperinVute's descriptionandyoucanreaditasyourlayerbecauseit's quiteaninterestingreadinmyhumbleopinion.
ヴュートの説明の中にこの論文へのリンクも落としておくので、自分のレイヤーとして読んでみてほしい。
Now, lookatthisdiagram.
さて、この図を見てほしい。
Thisactuallyshowsyou a bitmorein-depthandeasywayaswhatishappeninghere.
これは実際に、ここで起こっていることとして、もう少し踏み込んだ簡単な方法を示している。
So, totestthepowerusageandeffectivenessoftheMatmul-freeLMoncustomhardwarethatcanbetterexploitternaryoperations, theseresearchershavecreatedanFPGAacceleratorinSystemVerilogandthisisthewholeoverviewofit.
Theyalsowrote a customassemblerfortheircustominstructionsetwhichwasusedtoconvertassemblyfilesintoaninstructionrowandthereis a lotofdetailaroundtherethattheyalsohavethisregisterrouterasyoucanseeinthemiddlethatdelegatesincominginstructionsintoavailableregisters.
Iftwoinstructionsdonotblockeachother, theyexecutesimultaneouslyandthenalsoweseethereis a root-means-squarefunctionalunitthatuses a specializedhardware.
I'llgotopreserveprecisionandtherearefewstagesofitandthenwehavethisternarymatrixmultiplicationfunctionalunitthattakesin a DRAMaddressfor a ternarymatrixthenperforms a T-MATMULonspecifiedvector.
Thisarchitectureentirelyplaces a ternarymatricesandDRAMwhilerunning a T-MATMULinstructionandSRAMFIFOissimultaneouslyfilledwithsequentialDRAMfetchresult.
Asthedemandfordeployinglanguagemodelsonvariousplatformsgrows, MATMUL-freeLMspresent a promisingdirectionforcreatingmodelsthatarebotheffectiveandresourceefficient.