{"id":35567,"date":"2019-10-31T22:05:06","date_gmt":"2019-10-31T19:05:06","guid":{"rendered":"https:\/\/prohoster.info\/blog\/sokratit-energopotreblenie-v-tsod-pomozhet-novyj-fotonnyj-chip\/"},"modified":"2019-10-31T22:05:06","modified_gmt":"2019-10-31T19:05:06","slug":"sokratit-energopotreblenie-v-tsod-pomozhet-novyj-fotonnyj-chip","status":"publish","type":"post","link":"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/sokratit-energopotreblenie-v-tsod-pomozhet-novyj-fotonnyj-chip","title":{"rendered":"Sn\u00ed\u017eit spot\u0159ebu energie v datov\u00e9m centru pom\u016f\u017ee nov\u00fd fotonick\u00fd \u010dip","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

Na MIT vyvinuli architekturu nov\u00e9ho fotonick\u00e9ho procesoru. Ten zv\u00fd\u0161\u00ed efektivitu optick\u00fdch neuronov\u00fdch s\u00edt\u00ed tis\u00edckr\u00e1t ve srovn\u00e1n\u00ed s podobn\u00fdmi za\u0159\u00edzen\u00edmi. <\/p>\n

\u010cip sn\u00ed\u017e\u00ed objem elektrick\u00e9 energie spot\u0159ebov\u00e1van\u00e9 datov\u00fdmi centry. P\u0159edstavujeme, jak je navr\u017een.<\/p>\n

\"Sn\u00ed\u017eit<\/a><\/noindex><\/a><\/noindex>
\nFoto \u2014 Ildefonso Polo<\/a><\/noindex> \u2013 Unsplash<\/i><\/p>\n

Pro\u010d je nov\u00e1 architektura pot\u0159ebn\u00e1<\/h2>\n

\nOptick\u00e9 neuronov\u00e9 s\u00edt\u011b pracuj\u00ed rychleji ne\u017e tradi\u010dn\u00ed \u0159e\u0161en\u00ed pou\u017e\u00edvaj\u00edc\u00ed elektronick\u00e9 komponenty. Sv\u011btlo nepot\u0159ebuje<\/a><\/noindex> izolaci sign\u00e1ln\u00edch cest, a laserov\u00e9 paprsky mohou proch\u00e1zet jeden skrz druh\u00fd bez vz\u00e1jemn\u00e9ho ovlivn\u011bn\u00ed. D\u00edky tomu mohou v\u0161echny sign\u00e1ln\u00ed cesty fungovat sou\u010dasn\u011b, co\u017e zaji\u0161\u0165uje vysokou rychlost p\u0159enosu dat.<\/p>\n

Ale je tu probl\u00e9m \u2014 \u010d\u00edm v\u011bt\u0161\u00ed je rozm\u011br neuronov\u00e9 s\u00edt\u011b, t\u00edm v\u00edce energie spot\u0159ebov\u00e1v\u00e1. K vy\u0159e\u0161en\u00ed tohoto probl\u00e9mu se vyv\u00edjej\u00ed speci\u00e1ln\u00ed \u010dipy-urychlova\u010de (AI accelerators), kter\u00e9 optimalizuj\u00ed p\u0159enos dat. Nicm\u00e9n\u011b se \u0161k\u00e1luj\u00ed ne tak dob\u0159e, jak bychom si p\u0159\u00e1li.<\/p>\n

Na probl\u00e9mu energetick\u00e9 efektivity a \u0161k\u00e1lovatelnosti optick\u00fdch \u010dip\u016f pracuj\u00ed na MIT a p\u0159edstavili<\/a><\/noindex> novou architekturou fotonick\u00e9ho urychlova\u010de, kter\u00e1 sni\u017euje spot\u0159ebu energie za\u0159\u00edzen\u00ed tis\u00edckr\u00e1t a pracuje s des\u00edtkami milion\u016f neuron\u016f. V\u00fdvoj\u00e1\u0159i tvrd\u00ed, \u017ee technologia nalezne v budoucnosti uplatn\u011bn\u00ed v datov\u00fdch centrech, kter\u00e1 interaguj\u00ed se slo\u017eit\u00fdmi inteligentn\u00edmi syst\u00e9my a algoritmy strojov\u00e9ho u\u010den\u00ed a tak\u00e9 prov\u00e1d\u011bj\u00ed anal\u00fdzu velk\u00fdch dat.<\/p>\n

Co to p\u0159edstavuje<\/h2>\n

\nNov\u00fd \u010dip je postaven na b\u00e1zi optoelektronick\u00e9ho obvodu. P\u0159en\u00e1\u0161en\u00e1 data jsou st\u00e1le k\u00f3dov\u00e1na optick\u00fdmi sign\u00e1ly, ale k maticov\u00e9mu n\u00e1soben\u00ed se pou\u017e\u00edv\u00e1 vyv\u00e1\u017een\u00e9 homodynn\u00ed detekce (str. 30<\/a><\/noindex>). Je to technika, kter\u00e1 umo\u017e\u0148uje formovat elektrick\u00fd sign\u00e1l na z\u00e1klad\u011b dvou optick\u00fdch.<\/p>\n

Pro p\u0159enos sv\u011bteln\u00fdch impuls\u016f s informacemi o vstupn\u00edch a v\u00fdstupn\u00edch neuronech se pou\u017e\u00edv\u00e1 jedna sign\u00e1ln\u00ed cesta. \u00dadaje o vah\u00e1ch neuron\u016f se naopak p\u0159en\u00e1\u0161ej\u00ed po odd\u011blen\u00fdch kan\u00e1lech. V\u0161echny tyto informace se rozch\u00e1zej\u00ed po uzlech s\u00edt\u011b homogenn\u00edch fotodetektor\u016f, kter\u00e9 vypo\u010d\u00edt\u00e1vaj\u00ed v\u00fdstupn\u00ed hodnotu pro ka\u017ed\u00fd neuron (ur\u010duj\u00ed \u00farove\u0148 sign\u00e1lu). Pot\u00e9 tyto informace putuj\u00ed do modul\u00e1toru, kter\u00fd p\u0159ev\u00e1d\u00ed elektrick\u00fd sign\u00e1l zp\u011bt na optick\u00fd. N\u00e1sledn\u011b je odes\u00edl\u00e1na na dal\u0161\u00ed vrstvu neuronov\u00e9 s\u00edt\u011b a proces se opakuje.<\/p>\n

Ve sv\u00e9 v\u011bdeck\u00e9 pr\u00e1ci in\u017een\u00fd\u0159i z MIT uv\u00e1d\u011bj\u00ed<\/a><\/noindex> n\u00e1sleduj\u00edc\u00ed sch\u00e9ma pro jednu vrstvu:<\/p>\n

\"Sn\u00ed\u017eitObr\u00e1zek: Velkoplo\u0161n\u00e9 optick\u00e9 neuronov\u00e9 s\u00edt\u011b zalo\u017een\u00e9 na fotoelektrick\u00e9m n\u00e1soben\u00ed<\/a><\/noindex> \/ CC BY<\/i><\/p>\n

Nov\u00e1 architektura AI-akceler\u00e1toru vy\u017eaduje pouze jeden vstupn\u00ed a jeden v\u00fdstupn\u00ed kan\u00e1l pro ka\u017ed\u00fd neuron. V d\u016fsledku toho po\u010det fotodetektor\u016f odpov\u00edd\u00e1 po\u010dtu neuron\u016f, nikoliv jejich vah\u00e1m.<\/p>\n

Takov\u00fd p\u0159\u00edstup \u0161et\u0159\u00ed m\u00edsto na \u010dipu, zvy\u0161uje po\u010det u\u017eite\u010dn\u00fdch sign\u00e1ln\u00edch cest a optimalizuje spot\u0159ebu energie. Nyn\u00ed in\u017een\u00fd\u0159i z MIT vytv\u00e1\u0159ej\u00ed prototyp, kter\u00fd umo\u017en\u00ed otestovat mo\u017enosti nov\u00e9 architektury v praxi.<\/p><\/blockquote>\n

<\/p>\n

Kdo je\u0161t\u011b vyv\u00edj\u00ed fotonov\u00e9 \u010dipy<\/h2>\n

\nNa v\u00fdvoji podobn\u00e9 technologie se zab\u00fdv\u00e1<\/a><\/noindex> Lightelligence \u2014 mal\u00fd startup z Bostonu. Zam\u011bstnanci spole\u010dnosti tvrd\u00ed, \u017ee jejich AI-akceler\u00e1tor umo\u017en\u00ed vy\u0159e\u0161it \u00fakoly strojov\u00e9ho u\u010den\u00ed stokr\u00e1t rychleji ne\u017e klasick\u00e1 za\u0159\u00edzen\u00ed. V lo\u0148sk\u00e9m roce t\u00fdm dokon\u010dil v\u00fdrobu prototypu sv\u00e9ho za\u0159\u00edzen\u00ed a p\u0159ipravoval se na proveden\u00ed test\u016f.<\/p>\n

Pracuje v oblasti fotonov\u00fdch \u010dip\u016f i Cisco. Na za\u010d\u00e1tku roku spole\u010dnost ozn\u00e1mila n\u00e1kup<\/a><\/noindex> startup spole\u010dnosti Luxtera, kter\u00e1 navrhuje fotonov\u00e9 \u010dipy pro datov\u00e1 centra. Zejm\u00e9na spole\u010dnost vyr\u00e1b\u00ed hardwarov\u00e9 rozhran\u00ed, kter\u00e9 umo\u017e\u0148uje p\u0159ipojit optick\u00e9 vl\u00e1kno p\u0159\u00edmo k server\u016fm. Tento p\u0159\u00edstup zvy\u0161uje propustnost s\u00edt\u00ed a urychluje p\u0159enos dat. Za\u0159\u00edzen\u00ed Luxtera vyu\u017e\u00edvaj\u00ed speci\u00e1ln\u00ed lasery k k\u00f3dov\u00e1n\u00ed informac\u00ed a germaniov\u00e9 fotodetektory k jejich de\u0161ifrov\u00e1n\u00ed.<\/p>\n

\"Sn\u00ed\u017eit
\nFoto \u2014 Thomas Jensen<\/a><\/noindex> \u2013 Unsplash<\/i><\/p>\n

Optick\u00fdmi technologiemi se zab\u00fdvaj\u00ed i dal\u0161\u00ed velk\u00e9 IT spole\u010dnosti, jako nap\u0159\u00edklad Intel. Ji\u017e v roce 2016 za\u010dali vyr\u00e1b\u011bt vlastn\u00ed optick\u00e9 \u010dipy, kter\u00e9 optimalizuj\u00ed p\u0159enos dat mezi datov\u00fdmi centry. Ned\u00e1vno z\u00e1stupci organizace uv\u00e1d\u011bli<\/a><\/noindex>, co pl\u00e1nuj\u00ed zav\u00e1d\u011bt tyto technologie mimo datov\u00e1 centra \u2014 v LIDAR pro bezpilotn\u00ed automobily.<\/p>\n

Co nakonec<\/h2>\n

\nZat\u00edm nelze fotonov\u00e9 technologie nazvat univerz\u00e1ln\u00edm \u0159e\u0161en\u00edm. Jejich implementace vy\u017eaduje vysok\u00e9 n\u00e1klady na technick\u00e9 p\u0159epracov\u00e1n\u00ed datov\u00fdch center. Ale v\u00fdzkumy, jako ty, kter\u00e9 vyv\u00edj\u00ed MIT a dal\u0161\u00ed organizace, u\u010din\u00ed optick\u00e9 \u010dipy levn\u011bj\u0161\u00ed a pravd\u011bpodobn\u011b umo\u017en\u00ed jejich 'progres' na masov\u00e9m trhu s vybaven\u00edm pro datov\u00e1 centra.<\/p>\n

\nITGLOBAL.COM poskytujeme slu\u017eby priv\u00e1tn\u00edho a hybridn\u00edho cloudu a pom\u00e1h\u00e1me firm\u00e1m \u0159\u00eddit IT slu\u017eby. O \u010dem p\u00ed\u0161eme na firemn\u00edm blogu:<\/a><\/noindex>\u00a0pom\u00e1h\u00e1me firm\u00e1m rozv\u00edjet IT infrastrukturu a poskytujeme slu\u017eby priv\u00e1tn\u00ed a hybridn\u00ed cloud. O tom, co p\u00ed\u0161eme v firemn\u00edm blogu:<\/p>\n