{"id":14894,"date":"2026-09-17T08:50:39","date_gmt":"2026-09-17T05:50:39","guid":{"rendered":"https:\/\/www.kaspersky.com.tr\/blog\/?p=14894"},"modified":"2026-09-17T08:50:39","modified_gmt":"2026-09-17T05:50:39","slug":"gputhor-rowhammer-class-attack","status":"publish","type":"post","link":"https:\/\/www.kaspersky.com.tr\/blog\/gputhor-rowhammer-class-attack\/14894\/","title":{"rendered":"GPUThor: GPU \u00fczerinden ger\u00e7ekle\u015ftirilen bir ba\u015fka sunucu sald\u0131r\u0131 giri\u015fimi"},"content":{"rendered":"

Bir end\u00fcstriyel sunucu, iz b\u0131rakmadan grafik i\u015flemcisi (GPU) yoluyla nas\u0131l ele ge\u00e7irilebilir? Bu kadar karma\u015f\u0131k ve b\u00fcy\u00fck \u00f6l\u00e7\u00fcde teorik bir sald\u0131r\u0131, genellikle donan\u0131mdaki zay\u0131f noktalardan yararlan\u0131r. Bu, donan\u0131m\u0131n kendisindeki tasar\u0131m kusurlar\u0131yla ilgili de\u011fildir, daha \u00e7ok, bazen fiziksel d\u00fczeyde de g\u00f6r\u00fclen \u00e7al\u0131\u015fma \u015fekline dair tuhafl\u0131klarla ilgilidir. Toronto \u00dcniversitesi\u2019ndeki Kanadal\u0131 ara\u015ft\u0131rmac\u0131lar taraf\u0131ndan yay\u0131nlanan yeni bir makalede<\/a>, tam da video belle\u011findeki bu t\u00fcr donan\u0131m davran\u0131\u015f\u0131ndan yararlanan, yeni ve daha etkili bir Rowhammer sald\u0131r\u0131s\u0131 olan GPUThor anlat\u0131l\u0131yor.<\/p>\n

Rowhammer ve grafik kartlar\u0131<\/h2>\n

GPUThor, ilk olarak 2014 y\u0131l\u0131nda Rowhammer ara\u015ft\u0131rmas\u0131nda<\/a> ortaya at\u0131lan, RAM\u2019e y\u00f6nelik orijinal sald\u0131r\u0131n\u0131n temelindeki fikri temel almaktad\u0131r. Rowhammer ve bu s\u0131n\u0131ftaki t\u00fcm sald\u0131r\u0131lar, basit bir ger\u00e7e\u011fe dayan\u0131r: Bellek h\u00fccreleri birbirlerinden tamamen izole de\u011fildir. Ayn\u0131 h\u00fccre sat\u0131r\u0131na tekrar tekrar eri\u015fmek (hammering), belirli ko\u015fullar alt\u0131nda kom\u015fu sat\u0131rlardaki verileri bozabilir (yani bitleri ters \u00e7evirebilir). Bu etkinin m\u00fcmk\u00fcn oldu\u011fu do\u011fruland\u0131\u011f\u0131nda, geriye kalan tek \u015fey onu bir silah haline getirmenin bir yolunu bulmakt\u0131r; \u00f6rne\u011fin, bir hizmet reddi sald\u0131r\u0131s\u0131 tetikleyerek ya da hatta rastgele kod \u00e7al\u0131\u015ft\u0131rarak.<\/p>\n

Peki, sunucular ve grafik h\u0131zland\u0131r\u0131c\u0131lar\u0131n t\u00fcm bunlarla ne ilgisi var? Yapay zeka teknolojilerinin h\u0131zla yayg\u0131nla\u015fmas\u0131yla birlikte, \u00e7ok say\u0131da paralel ve benzer hesaplamay\u0131 ger\u00e7ekle\u015ftirebilen donan\u0131mlara olan talep de artm\u0131\u015ft\u0131r. Oyun grafik kartlar\u0131na entegre edilmi\u015f h\u0131zland\u0131r\u0131c\u0131lar ise bu t\u00fcr i\u015f y\u00fckleri i\u00e7in bi\u00e7ilmi\u015f kaftand\u0131r. Bu durum, grafik h\u0131zland\u0131r\u0131c\u0131lar\u0131n\u0131 herkese kiralayan bulut sa\u011flay\u0131c\u0131lar\u0131n\u0131 Rowhammer sald\u0131r\u0131lar\u0131 i\u00e7in cazip bir hedef haline getirir. Varsay\u0131msal senaryo \u015fu \u015fekildedir: Bir sald\u0131rgan, bir grafik yongas\u0131na eri\u015fim hakk\u0131 sat\u0131n al\u0131r ve bunu, sa\u011flay\u0131c\u0131n\u0131n t\u00fcm altyap\u0131s\u0131n\u0131 ele ge\u00e7irmeye \u00e7al\u0131\u015fmak i\u00e7in kullan\u0131r. \u0130\u015fte tam da bu nedenle, video belle\u011fine y\u00f6nelik sald\u0131r\u0131lar ara\u015ft\u0131rmac\u0131lar i\u00e7in h\u00e2l\u00e2 \u00f6zel bir \u00f6neme sahiptir.<\/p>\n

Unsaflok sald\u0131r\u0131s\u0131 nas\u0131l \u00e7al\u0131\u015f\u0131r?<\/h2>\n

Bu y\u0131l\u0131n ilkbahar\u0131nda, her biri GDDR6 bellekle donat\u0131lm\u0131\u015f Nvidia h\u0131zland\u0131r\u0131c\u0131lar\u0131na y\u00f6nelik farkl\u0131 bir sald\u0131r\u0131y\u0131 ortaya koyan \u00fc\u00e7 yeni makale<\/a> yay\u0131nland\u0131. \u00dc\u00e7\u00fc de olduk\u00e7a m\u00fctevaz\u0131 sonu\u00e7lar verdi: En b\u00fcy\u00fck hasar, hedef olarak t\u00fcketici s\u0131n\u0131f\u0131 bir grafik kart\u0131 se\u00e7ildi\u011finde meydana gelirken, Nvidia A6000 gibi end\u00fcstriyel bir h\u0131zland\u0131r\u0131c\u0131ya y\u00f6nelik sald\u0131r\u0131lar\u0131n \u00e7ok daha d\u00fc\u015f\u00fck etkili oldu\u011fu ortaya \u00e7\u0131kt\u0131. Bunun yan\u0131 s\u0131ra, ECC (hata d\u00fczeltme kodu) bellek korumas\u0131 etkinle\u015ftirildi\u011finde sald\u0131r\u0131lar\u0131n hi\u00e7biri ba\u015far\u0131l\u0131 olamad\u0131.<\/p>\n

GPUThor, Nvidia\u2019nin GDDR6 bellek kullanan eski Ampere h\u0131zland\u0131r\u0131c\u0131lar\u0131na sald\u0131rma olas\u0131l\u0131\u011f\u0131n\u0131 da ele al\u0131yor: Ara\u015ft\u0131rmac\u0131lar A4000, A4500, A5000 ve A6000 modellerini inceledi. Ancak yeni y\u00f6ntemin etkilili\u011fi \u2014 verileri zorla de\u011fi\u015ftirilen h\u00fccre say\u0131s\u0131 ile \u00f6l\u00e7\u00fcld\u00fc\u011f\u00fcnde \u2014 belirgin \u015fekilde daha y\u00fcksek. Ara\u015ft\u0131rmac\u0131lar ayr\u0131ca bu tekni\u011fin, teorik olarak, daha yeni h\u0131zland\u0131r\u0131c\u0131lara da uygulanabilece\u011fini savunuyor.<\/p>\n

\"GPUThor<\/a>

GPUThor\u2019un \u00f6nceki sald\u0131r\u0131lara k\u0131yasla etkinli\u011fi. Kaynak <\/a><\/p><\/div>\n

Bu sonu\u00e7lar nas\u0131l elde edildi? Rowhammer sald\u0131r\u0131lar\u0131na kar\u015f\u0131 standart olan ve Target Row Refresh (TRR) olarak adland\u0131r\u0131lan bu savunma mekanizmas\u0131, Kanadal\u0131 ara\u015ft\u0131rmac\u0131lar taraf\u0131ndan daha yak\u0131ndan incelendi. G\u00f6r\u00fcn\u00fc\u015fe g\u00f6re, TRR tekrarlanan eri\u015fim giri\u015fimlerini tespit etti\u011finde kom\u015fu h\u00fccrelerin yenilenmesini zorlamaktad\u0131r ve bu da verilerin bozulmas\u0131n\u0131 zorla\u015ft\u0131r\u0131r ya da imkans\u0131z hale getirir. Sald\u0131rganlar genellikle h\u00fccrelere rastgele eri\u015ferek TRR\u2019yi a\u015fmaya \u00e7al\u0131\u015f\u0131rlar; bu sayede savunma sistemini kar\u0131\u015ft\u0131r\u0131r ve etkinli\u011fini azalt\u0131rlar. Ara\u015ft\u0131rmac\u0131lar, Nvidia Ampere kartlar\u0131nda TRR\u2019nin yaln\u0131zca her 72 bellek h\u00fccresi yenileme d\u00f6ng\u00fcs\u00fcnde bir kez tetiklendi\u011fini ke\u015ffettiler. Bu bulgudan yola \u00e7\u0131karak, dengesiz bir eri\u015fim modeli uygulad\u0131lar ve hedef h\u00fccrelere eskisinden \u00e7ok daha agresif bir \u015fekilde sald\u0131rd\u0131lar. Sonu\u00e7: Referans noktas\u0131 olarak \u201cGPUHammer\u201d ad\u0131yla bilinen orijinal GDDR sald\u0131r\u0131s\u0131 ile kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda, GPUThor\u2019un yakla\u015f\u0131k 7.000 ila 23.000 kat daha etkili oldu\u011fu ortaya \u00e7\u0131kt\u0131.<\/p>\n

Sonu\u00e7lar ve gelece\u011fe dair \u00f6ng\u00f6r\u00fcler<\/h2>\n

Bu daha agresif sald\u0131r\u0131 modelini di\u011fer iyile\u015ftirmelerle birle\u015ftirerek, gigabayt ba\u015f\u0131na 72.000 ila 377.000 bit ters \u00e7evirme elde edildi. Daha \u00f6nceki Rowhammer varyantlar\u0131 en iyi ihtimalle birka\u00e7 y\u00fcz tane ba\u015farabilmi\u015fti. Bu sayede ara\u015ft\u0131rmac\u0131lar, iki bitlik ve hatta \u00fc\u00e7 bitlik hatalar elde etmeyi ba\u015fard\u0131lar. ECC, tek bitlik bir hatay\u0131 kolayca d\u00fczeltir, ancak \u00e7ift bitlik bir hatay\u0131 d\u00fczeltemez.<\/p>\n

Bu yeni y\u00f6ntem, bir Rowhammer sald\u0131r\u0131s\u0131n\u0131n ger\u00e7ek d\u00fcnyada yol a\u00e7abilece\u011fi zarar\u0131 da ortaya koymaktad\u0131r: GPUThor kullan\u0131larak video belle\u011fine tekrar tekrar eri\u015fim, hizmet reddi durumuna yol a\u00e7maktad\u0131r. H\u0131zland\u0131r\u0131c\u0131 \u00f6nce yeniden ba\u015flat\u0131l\u0131r ve bu s\u00fcre\u00e7te veriler kaybolur, ard\u0131ndan y\u00f6neticiye de\u011fi\u015ftirilmesi gerekti\u011fi konusunda bildirimde bulunur.<\/p>\n

Bu etkileyici ara\u015ft\u0131rma sonu\u00e7lar\u0131na ra\u011fmen GPUThor sald\u0131r\u0131lar\u0131 tam anlam\u0131yla ba\u015far\u0131ya ula\u015fm\u0131\u015f de\u011fil. Birincisi ara\u015ft\u0131rmac\u0131lar veri bozulmas\u0131 sonucunda rastgele kod \u00e7al\u0131\u015ft\u0131r\u0131labilece\u011fini kan\u0131tlayamad\u0131lar \u2013 her ne kadar ECC etkinle\u015ftirildi\u011finde bile bunun m\u00fcmk\u00fcn oldu\u011funu iddia etseler de. \u00a0\u0130kincisi, binlerce kat daha etkili bir sald\u0131r\u0131, daha yeni h\u0131zland\u0131r\u0131c\u0131lar\u0131n da ele ge\u00e7irilebilece\u011fine dair teorik bir olas\u0131l\u0131\u011fa i\u015faret ediyor; ancak bu durum da \u015fimdilik kan\u0131tlanabilmi\u015f de\u011fil.<\/p>\n

Buna ra\u011fmen, Kanadal\u0131 ara\u015ft\u0131rmac\u0131lar, grafik h\u0131zland\u0131r\u0131c\u0131lara y\u00f6nelik Rowhammer sald\u0131r\u0131lar\u0131n\u0131n h\u00e2l\u00e2 kullan\u0131lmam\u0131\u015f bir potansiyele sahip oldu\u011funu ortaya koydu. Gelecekteki ara\u015ft\u0131rmalar\u0131n, daha \u00f6nce bu t\u00fcr sald\u0131r\u0131lara kar\u015f\u0131 son derece diren\u00e7li oldu\u011fu d\u00fc\u015f\u00fcn\u00fclen \u00e7ok daha geli\u015fmi\u015f cihazlara y\u00f6nelik benzer sald\u0131r\u0131lar\u0131 ortaya \u00e7\u0131karmas\u0131 hi\u00e7 de \u015fa\u015f\u0131rt\u0131c\u0131 olmayacakt\u0131r.<\/p>\n\n","protected":false},"excerpt":{"rendered":"

Ara\u015ft\u0131rmac\u0131lar, grafik h\u0131zland\u0131r\u0131c\u0131lar\u0131na yap\u0131lan k\u0131smen etkili bir Rowhammer sald\u0131r\u0131s\u0131n\u0131 tan\u0131mlad\u0131. <\/p>\n","protected":false},"author":665,"featured_media":14895,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[1726,1194,1727],"tags":[754,2880],"class_list":["post-14894","post","type-post","status-publish","format-standard","has-post-thumbnail","category-enterprise","category-business","category-smb","tag-donanim","tag-gpu"],"hreflang":[{"hreflang":"tr","url":"https:\/\/www.kaspersky.com.tr\/blog\/gputhor-rowhammer-class-attack\/14894\/"},{"hreflang":"es","url":"https:\/\/www.kaspersky.es\/blog\/gputhor-rowhammer-class-attack\/32472\/"},{"hreflang":"ru","url":"https:\/\/www.kaspersky.ru\/blog\/gputhor-rowhammer-class-attack\/42623\/"},{"hreflang":"x-default","url":"https:\/\/www.kaspersky.com\/blog\/gputhor-rowhammer-class-attack\/56362\/"},{"hreflang":"de","url":"https:\/\/www.kaspersky.de\/blog\/gputhor-rowhammer-class-attack\/33868\/"},{"hreflang":"ru-kz","url":"https:\/\/blog.kaspersky.kz\/gputhor-rowhammer-class-attack\/31011\/"}],"acf":[],"banners":"","maintag":{"url":"https:\/\/www.kaspersky.com.tr\/blog\/tag\/donanim\/","name":"donan\u0131m"},"_links":{"self":[{"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/posts\/14894","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/users\/665"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/comments?post=14894"}],"version-history":[{"count":3,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/posts\/14894\/revisions"}],"predecessor-version":[{"id":14898,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/posts\/14894\/revisions\/14898"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/media\/14895"}],"wp:attachment":[{"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/media?parent=14894"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/categories?post=14894"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/tags?post=14894"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}