{"id":18,"date":"2011-07-21T07:17:33","date_gmt":"2011-07-21T07:17:33","guid":{"rendered":"http:\/\/team.inria.fr\/cupseli1\/?page_id=18"},"modified":"2011-07-21T07:17:33","modified_gmt":"2011-07-21T07:17:33","slug":"research","status":"publish","type":"page","link":"https:\/\/project.inria.fr\/cupseli\/research\/","title":{"rendered":"Research"},"content":{"rendered":"<p><\/p>\n<h2>Context<\/h2>\n<p>The demand for AI computation is growing faster than the availability of high-end GPUs and dedicated infrastructures. Cupseli\u2019s vision is to harness underused distributed power for large-scale AI workloads, without building new data centers, thus lowering costs and carbon footprint.<\/p>\n<h2>Ambition<\/h2>\n<ul class=\"list\">\n<li>Fine-tuning &amp; inference across volatile resources<\/li>\n<li>Optimized memory, computation, and communications<\/li>\n<li>Security, privacy, and fault tolerance by design<\/li>\n<li>Greener, sovereign alternative to centralized data centers<\/li>\n<\/ul>\n<div class=\"card\">\n<h3>Axis 1 \u2014 Frugality<\/h3>\n<p>Adapt training and inference to limited, dynamic resources: fault tolerance, activation\/weight compression, memory offloading, partitioning, and re-materialization.<\/p>\n<\/div>\n<div class=\"card\">\n<h3>Axis 2 \u2014 Security &amp; Confidentiality<\/h3>\n<p>Protect data and models: encryption, homomorphic operations, confidential computing, and poisoning\/backdoor defenses.<\/p>\n<\/div>\n<div class=\"card\">\n<h3>Axis 3 \u2014 Volatility<\/h3>\n<p>Resilient scheduling, failure-aware execution, checkpointing, and straggler mitigation on heterogeneous, intermittent resources.<\/p>\n<\/div>\n<p><\/p>","protected":false},"excerpt":{"rendered":"<p>Context The demand for AI computation is growing faster than the availability of high-end GPUs and dedicated infrastructures. Cupseli\u2019s vision is to harness underused distributed power for large-scale AI workloads, without building new data centers, thus lowering costs and carbon footprint. Ambition Fine-tuning &amp; inference across volatile resources Optimized memory,\u2026<\/p>\n<p> <a class=\"continue-reading-link\" href=\"https:\/\/project.inria.fr\/cupseli\/research\/\"><span>Continue reading<\/span><i class=\"crycon-right-dir\"><\/i><\/a> <\/p>\n","protected":false},"author":791,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"open","template":"","meta":{"footnotes":""},"class_list":["post-18","page","type-page","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/project.inria.fr\/cupseli\/wp-json\/wp\/v2\/pages\/18","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/project.inria.fr\/cupseli\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/project.inria.fr\/cupseli\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/project.inria.fr\/cupseli\/wp-json\/wp\/v2\/users\/791"}],"replies":[{"embeddable":true,"href":"https:\/\/project.inria.fr\/cupseli\/wp-json\/wp\/v2\/comments?post=18"}],"version-history":[{"count":0,"href":"https:\/\/project.inria.fr\/cupseli\/wp-json\/wp\/v2\/pages\/18\/revisions"}],"wp:attachment":[{"href":"https:\/\/project.inria.fr\/cupseli\/wp-json\/wp\/v2\/media?parent=18"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}