{"id":1765,"date":"2021-09-03T10:00:00","date_gmt":"2021-09-03T02:00:00","guid":{"rendered":"https:\/\/blog.adlinktech.com\/?p=1765"},"modified":"2021-10-11T15:53:34","modified_gmt":"2021-10-11T07:53:34","slug":"improve-medical-imaging-ai-gtc-reflections","status":"publish","type":"post","link":"https:\/\/blog.adlinktech.com\/tw\/2021\/09\/03\/%e6%94%b9%e5%96%84%e9%86%ab%e5%ad%b8%e6%88%90%e5%83%8f-ai-gtc-%e5%8f%8d%e6%80%9d\/","title":{"rendered":"\u96a8\u8457\u4eba\u5de5\u667a\u80fd (AI) \u7684\u4f7f\u7528\uff0c\u91ab\u5b78\u6210\u50cf\u6b63\u5728\u5f97\u5230\u6539\u5584"},"content":{"rendered":"

\u6211\u4e00\u76f4 \u53cd\u601d\u9019\u500b GTC 2021<\/a> \u8a0e\u8ad6\u4eba\u5de5\u667a\u80fd\u5982\u4f55\u5f71\u97ff\u91ab\u5b78\u6210\u50cf\u7684\u6703\u8b70\uff0c\u7576\u7136\u9084\u6709\u5e6b\u52a9\u61c9\u5c0d\u884c\u696d\u8b8a\u5316\u7684\u908a\u7de3\u4eba\u5de5\u667a\u80fd\u89e3\u6c7a\u65b9\u6848\u3002\u5f88\u660e\u986f\uff0c\u908a\u7de3\u8a08\u7b97\u548c GPU \u4e00\u8d77\u767c\u63ee\u8457\u95dc\u9375\u4f5c\u7528\uff0c\u6240\u4ee5\u6211\u60f3\u7c21\u8981\u5206\u4eab\u6211\u5011\u6240\u770b\u5230\u7684\uff0c\u4ee5\u53ca\u6211\u6700\u559c\u6b61\u7684\u5169\u500b\u5728\u91ab\u5b78\u6210\u50cf\u6280\u8853\u4e2d\u5be6\u73fe\u4eba\u5de5\u667a\u80fd\u7684\u529f\u80fd\u3002  <\/p>\n\n\n\n

\u6211\u77e5\u9053\u6211\u4e0d\u6703\u8b93\u4efb\u4f55\u4eba\u611f\u5230\u9a5a\u8a1d\uff0c\u56e0\u70ba\u4eba\u5de5\u667a\u80fd\u786c\u4ef6\u548c\u8edf\u4ef6\u7684\u6210\u672c\u5f88\u5feb\u5c31\u6703\u98c6\u5347\u3002\u4f46\u662f\uff0c\u6211\u89ba\u5f97\u6709\u5fc5\u8981\u91cd\u7533\u9019\u4e00\u9ede\uff0c\u4ee5\u4fbf\u70ba\u672c\u535a\u5ba2\u7684\u5176\u9918\u90e8\u5206\u5960\u5b9a\u57fa\u77f3\u3002\u5728\u6700\u8fd1 \u5e02\u5834\u5206\u6790<\/a>\uff0c\u9810\u8a08\u5f9e 2021 \u5e74\u5230 2028 \u5e74\uff0c\u5168\u7403\u91ab\u5b78\u5f71\u50cf\u5e02\u5834\u5c07\u4ee5 5.2% \u7684\u8907\u5408\u5e74\u589e\u9577\u7387\u589e\u9577\uff0c\u5230 2028 \u5e74\u5c07\u9054\u5230 286 \u5104\u7f8e\u5143\u3002\u96d6\u7136\u8a31\u591a\u4f7f\u7528 AI \u7684\u91ab\u7642\u61c9\u7528\u8207\u5176\u4ed6\u61c9\u7528\u76f8\u6bd4\u5c07\u51fa\u73fe\u589e\u9577\uff0c\u4f46\u91ab\u5b78\u5f71\u50cf\u7684\u589e\u9577\u901f\u5ea6\u76f8\u7576\u5feb\u3002\u6fc0\u589e\u3002 <\/p>\n\n\n\n

\u91ab\u5b78\u6210\u50cf\u3001GPU \u548c\u908a\u7de3\u8a08\u7b97 <\/strong> <\/h2>\n\n\n\n

\u6700\u65b0\u7684\u786c\u4ef6\u5728\u89e3\u6c7a\u8af8\u5982\u97ff\u61c9\u6027\u3001\u66f4\u9ad8\u6e96\u78ba\u6027\u548c\u589e\u5f37\u6210\u50cf\u7b49\u554f\u984c\u6240\u9700\u7684\u6027\u80fd\u65b9\u9762\u63d0\u4f9b\u4e86\u5f88\u591a\u3002\u5e78\u904b\u7684\u662f\uff0cGPU \u8a08\u7b97\u6539\u5584\u4e86\u91ab\u7642\u4fdd\u5065\u5c08\u696d\u4eba\u54e1\u7684\u91ab\u5b78\u6210\u50cf\uff0c\u56e0\u6b64\u4ed6\u5011\u53ef\u4ee5\u505a\u51fa\u66f4\u5feb\u3001\u66f4\u6e96\u78ba\u7684\u6c7a\u7b56\u3002 <\/p>\n\n\n\n

\u6211\u5011\u6e05\u695a\u5730\u770b\u5230\u4e86\u91ab\u7642\u6210\u50cf\u6280\u8853\u65b9\u9762\u7684\u884c\u696d\u9769\u547d\u3002\u9019\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u8981\u6b78\u529f\u65bc NVIDIA \u5728 GPU \u6280\u8853\u65b9\u9762\u7684\u9032\u6b65\uff0c\u5b83\u88dc\u5145\u4e86\u51cc\u83ef\u79d1\u6280\u5728\u908a\u7de3 AI \u786c\u4ef6\u548c\u8edf\u4ef6\u65b9\u9762\u7684\u9032\u6b65\u3002\u908a\u7de3\u8a08\u7b97\u901a\u904e\u52a0\u5165GPU\uff0c\u63d0\u5347\u5716\u50cf\u8cea\u91cf\uff0c\u589e\u5f37\u5716\u50cf\u69cb\u5efa\u904e\u7a0b\uff0c\u5b8c\u6210\u5716\u50cf\u5206\u6790\uff0c\u8f14\u52a9\u91ab\u7642\u5c08\u696d\u4eba\u58eb\u3002\u5c0d\u65bc\u81e8\u5e8a\u5be6\u8e10\u548c\u5916\u79d1\u624b\u8853\uff0c\u667a\u80fd\u8996\u983b\u5206\u6790 (IVA) \u662f GPU \u89e3\u6c7a\u65b9\u6848\u7684\u5de8\u5927\u512a\u52e2\u3002 <\/p>\n\n\n\n

\u5d4c\u5165\u5f0f\u79fb\u52d5 PCI Express \u6a21\u584a (MXM) \u662f\u53e6\u4e00\u7a2e GPU \u6a19\u6e96\uff0c\u5b83\u4e5f\u589e\u5f37\u4e86\u91ab\u5b78\u6210\u50cf\u61c9\u7528\u3002\u908a\u7de3\u8a08\u7b97\u8207\u91ab\u5b78\u6210\u50cf\u4e2d\u7684 GPU \u76f8\u7d50\u5408\uff0c\u63d0\u4f9b\u4e86\u5716\u50cf\u91cd\u5efa\u6240\u9700\u7684\u5f37\u5927\u8a08\u7b97\u5de5\u4f5c\u8ca0\u8f09\u4ee5\u53ca\u7522\u751f\u7684\u5927\u91cf\u71b1\u91cf\u3002\u5728\u81e8\u5e8a\u6aa2\u67e5\u3001\u6d3b\u6aa2\u3001\u8853\u4e2d\u76e3\u6e2c\u751a\u81f3\u624b\u8853\u6cbb\u7642\u7b49\u61c9\u7528\u4e2d\uff0c\u8a08\u7b97\u74b0\u5883\u4e2d\u7684\u6bcf\u74e6\u6027\u80fd\u751a\u81f3\u64f4\u5c55\u6eab\u5ea6\u9078\u9805\u53ef\u4ee5\u6c7a\u5b9a\u5716\u50cf\u91cd\u5efa\u7684\u901f\u5ea6\u548c\u6e96\u78ba\u6027\u3002  <\/p>\n\n\n\n

\u8207\u773e\u4e0d\u540c\u7684\u529f\u80fd\uff1aMXM\u3001RDMA <\/strong> <\/h2>\n\n\n\n

MXM<\/strong> <\/p>\n\n\n\n

\u53ef\u4ee5\u90e8\u7f72\u591a\u7a2e\u8655\u7406\u5668\u4ee5\u9069\u61c9\u7279\u5b9a\u61c9\u7528\u3002\u51cc\u83ef\u79d1\u6280\u7684\u5d4c\u5165\u5f0f MXM GPU \u6a21\u584a\u5177\u6709\u8d85\u8d8a\u50b3\u7d71 CPU \u7684\u6027\u80fd\u63d0\u5347\u80fd\u529b\u3002 <\/p>\n\n\n\n

\u6211\u5011\u7684 MXM \u6a21\u584a\u662f\u4f7f\u7528 NVIDIA Turing \u67b6\u69cb\u958b\u767c\u7684\uff0c\u8a72\u67b6\u69cb\u5728\u4e00\u500b GPU \u4e2d\u96c6\u6210\u4e86 CUDA \u5167\u6838\u3001RT \u5167\u6838\u548c Tensor \u5167\u6838\u3002\u4e00\u500b\u4f8b\u5b50\u662f EGX-MXM-RTX3000 \u6a21\u584a<\/a> \u5b83\u5728 MXM 3.1 Type B \u5916\u5f62\u4e2d\u63a1\u7528\u5148\u9032\u7684 NVIDIA Turing GPU \u6280\u8853\u3002\u9019\u53ea\u662f\u5b8c\u6574 PCI Express \u5716\u5f62 (PEG) \u5361\u5c3a\u5bf8\u7684\u56db\u5206\u4e4b\u4e00\u3002\u9019\u4e9b\u6a21\u584a\u7684\u8d77\u59cb\u529f\u7387\u70ba 80W\uff0c\u9069\u7528\u65bc\u5c3a\u5bf8\u3001\u91cd\u91cf\u548c\u529f\u7387\u53d7\u9650 (SWaP) \u7684\u95dc\u9375\u91ab\u5b78\u6210\u50cf\u61c9\u7528\u3002 <\/p>\n\n\n\n

\u91ab\u7642\u9818\u57df\u7684\u61c9\u7528\u7a0b\u5e8f\u4e5f\u6e34\u671b\u5de5\u696d\u9818\u57df\u61c9\u7528\u7a0b\u5e8f\u7684\u58fd\u547d\uff0c\u9019\u53ef\u4ee5\u901a\u904e\u806f\u5408\u51cc\u83ef\u79d1\u6280\u548c NVIDIA \u89e3\u6c7a\u65b9\u6848\u4f86\u5be6\u73fe\u3002   <\/p>\n\n\n\n

RDMA <\/strong> <\/p>\n\n\n\n

\u82f1\u5049\u9054\u7684 GPUDirect \u9060\u7a0b DMA<\/a> (RDMA) \u662f\u201c\u4e00\u7cfb\u5217\u6280\u8853\uff0c\u53ef\u589e\u5f37 NVIDIA \u6578\u64da\u4e2d\u5fc3 GPU \u7684\u6578\u64da\u79fb\u52d5\u548c\u8a2a\u554f\u201d\u3002\u5be6\u73fe RDMA \u70ba\u5916\u90e8\u6578\u64da\u6e90\u63d0\u4f9b\u4e86\u5c0d GPU \u5167\u5b58\u7684\u76f4\u63a5\u8a2a\u554f\u3002\u5982\u679c\u5728\u91ab\u5b78\u6210\u50cf\u4e2d\u6c92\u6709 RDMA\uff0c\u6578\u64da\u5c07\u9032\u5165 CPU \u7684\u5167\u5b58\uff0c\u9019\u53ef\u80fd\u6703\u5c0e\u81f4\u6578\u64da\u50b3\u8f38\u548c\u5ef6\u9072\u7684\u9032\u4e00\u6b65\u5ef6\u9072\u3002  <\/p>\n\n\n\n

\u4f8b\u5982\uff0c\u6d89\u53ca RDMA \u7684\u4e00\u500b\u7528\u4f8b\u662f\u8d85\u8072\u6210\u50cf\u3002\u8d85\u8072\u6ce2\u5229\u7528\u524d\u7aef\u8a2d\u5099\uff0c\u4f8b\u5982 \u73fe\u5834\u53ef\u7de8\u7a0b\u9580\u9663\u5217<\/a> \u7528\u65bc\u5728\u6578\u64da\u5230\u9054\u5176\u6700\u7d42\u76ee\u7684\u5730 GPU \u4e4b\u524d\u9032\u884c\u6a21\u6578\u8f49\u63db\u3002\u9019\u4ee3\u8868\u4e86 FPGA \u548c GPU \u4e4b\u9593\u7684\u5927\u91cf\u901a\u4fe1\u3002 RDMA \u5141\u8a31\u589e\u52a0\u5e36\u5bec\uff0c\u5f9e\u800c\u63d0\u4f9b\u8d85\u8072\u6ce2\u6240\u9700\u7684\u8a08\u7b97\u6027\u80fd\u3002 <\/p>\n\n\n\n

\u932f\u904e\u4e86\u6703\u8b70\uff1f <\/strong> <\/h2>\n\n\n\n

\u9084\u6709\u66f4\u591a\uff01\u5e0c\u671b\u6211\u5011\u5f88\u5feb\u6703\u518d\u6b21\u89aa\u81ea\u56de\u4f86\uff0c\u5426\u5247\uff0c\u5982\u679c\u60a8\u932f\u904e\u4e86\u6700\u521d\u7684\u6703\u8b70\uff0c\u60a8\u53ef\u4ee5\u4e86\u89e3\u66f4\u591a\u95dc\u65bc\u51cc\u83ef\u79d1\u6280\u7684\u91ab\u5b78\u5f71\u50cf\u6280\u8853 \u5728\u9019\u6b21\u8ac7\u8a71\u4e2d<\/a> \u6216\u8a2a\u554f\u6211\u5011 \u9019\u88e1<\/a>. <\/p>\n\n\n\n

\"\u4f5c\u8005\uff1a\u8d0a\u6069\u00b7\u8521\"\/<\/figure>
\u4f5c\u8005\uff1a\u8d0a\u6069\u00b7\u8521<\/span>

\u51cc\u83ef\u79d1\u6280\u5d4c\u5165\u5f0f\u5e73\u53f0\u8207\u6a21\u584a\u5e73\u53f0\u7522\u54c1\u4e2d\u5fc3\u7e3d\u76e3<\/p>\n

\u8a2a\u554f\u51cc\u83ef\u79d1\u6280<\/a><\/div>\n<\/div><\/div>\n\n\n\n

<\/p>","protected":false},"excerpt":{"rendered":"

I\u2019ve been reflecting on this GTC 2021 session discussing how AI is affecting medical imaging, and of course the edge AI solutions to help address industry changes. It\u2019s clear that edge computing and the GPU, together, are playing a key role so I wanted to briefly share what we\u2019ve been seeing, and two of my favorite features […]<\/p>\n","protected":false},"author":188103281,"featured_media":1766,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_stc_notifier_status":"sent","_stc_notifier_sent_time":"2022-03-29 15:20:02","_stc_notifier_request":false,"_stc_notifier_prevent":false,"_stc_subscriber_keywords":"","_stc_subscriber_search_areas":"","_coblocks_attr":"","_coblocks_dimensions":"","_coblocks_responsive_height":"","_coblocks_accordion_ie_support":"","_crdt_document":"","jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2},"_wpas_customize_per_network":false},"categories":[1,702134936],"tags":[],"class_list":{"0":"post-1765","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","6":"hentry","7":"category-uncategorized","8":"category-japanese","10":"fallback-thumbnail"},"jetpack_publicize_connections":[],"jetpack_featured_media_url":"https:\/\/i0.wp.com\/blog.adlinktech.com\/wp-content\/uploads\/2021\/09\/Banner-Tech_ADLINK_Tech_Medical-Imaging-Reflections_GTC-2021-blog.jpg?fit=1024%2C536&ssl=1","jetpack_likes_enabled":true,"jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/pc6JDC-st","_links":{"self":[{"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/posts\/1765","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/users\/188103281"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/comments?post=1765"}],"version-history":[{"count":5,"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/posts\/1765\/revisions"}],"predecessor-version":[{"id":1874,"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/posts\/1765\/revisions\/1874"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/media\/1766"}],"wp:attachment":[{"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/media?parent=1765"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/categories?post=1765"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.adlinktech.com\/tw\/wp-json\/wp\/v2\/tags?post=1765"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}