“AI的IPHONE时刻”系列8：如何初步理解GPU算力？

广发证券股份有...2023-04-03

核心观点：结论。GPU 峰值算力反映GPU 的主频和相应核心的硬件配置，如FP32 的核心的个数，FP64 的核心的个数以及Tensor core 的计算能力等峰值参数，并非其实际计算性能。GPU 峰值算力参数可用来初步区分其适用的应用场景，如在某GPU 内，其FP64 算力显著高于FP32 等算力，其可能更适用于科学计算等场景，如FP32、TF32、FP16、INT8 等算力较为均衡，其可能更...

网页链接

免责声明：本文观点仅代表作者个人观点，不构成本平台的投资建议，本平台不对文章信息准确性、完整性和及时性做出任何保证，亦不对因使用或信赖文章信息引发的任何损失承担责任。

精彩评论

我们需要你的真知灼见来填补这片空白

发表看法

{"i18n":{"language":"zh_CN"},"isChannel":false,"data":{"share":"https://www.laohu8.com/m/news/2324686508?lang=zh_CN&edition=full","thumbnail":"","is_english":false,"pubTime":"2023-04-03 18:25","share_image_url":"https://static.laohu8.com/9a95c1376e76363c1401fee7d3717173","id":"2324686508","market":"us","top_or_hot":-1,"title":"“AI的IPHONE时刻”系列8：如何初步理解GPU算力？","media":"广发证券股份有...","content":"<div>\n<p>核心观点：结论。GPU 峰值算力反映GPU 的主频和相应核心的硬件配置，如FP32 的核心的个数，FP64 的核心的个数以及Tensor core 的计算能力等峰值参数，并非其实际计算性能。GPU 峰值算力参数可用来初步区分其适用的应用场景，如在某GPU 内，其FP64 算力显著高于FP32 等算力，其可能更适用于科学计算等场景，如FP32、TF32、FP16、INT8 等算力较为均衡，其可能更...</p>\n\n<a href=\"http://gu.qq.com/resources/shy/news/detail-v2/index.html#/?id=nesSN202304031825158421d312&s=b\">网页链接</a>\n\n</div>\n","source":"tencent","html":"<!DOCTYPE html>\n<html>\n<head>\n<meta http-equiv=\"Content-Type\" content=\"text/html; charset=utf-8\" />\n<meta name=\"viewport\" content=\"width=device-width,initial-scale=1.0,minimum-scale=1.0,maximum-scale=1.0,user-scalable=no\"/>\n<meta name=\"format-detection\" content=\"telephone=no,email=no,address=no\" />\n<title>“AI的IPHONE时刻”系列8：如何初步理解GPU算力？</title>\n<style type=\"text/css\">\na,abbr,acronym,address,applet,article,aside,audio,b,big,blockquote,body,canvas,caption,center,cite,code,dd,del,details,dfn,div,dl,dt,\nem,embed,fieldset,figcaption,figure,footer,form,h1,h2,h3,h4,h5,h6,header,hgroup,html,i,iframe,img,ins,kbd,label,legend,li,mark,menu,nav,\nobject,ol,output,p,pre,q,ruby,s,samp,section,small,span,strike,strong,sub,summary,sup,table,tbody,td,tfoot,th,thead,time,tr,tt,u,ul,var,video{ font:inherit;margin:0;padding:0;vertical-align:baseline;border:0 }\nbody{ font-size:16px; line-height:1.5; color:#999; background:transparent; }\n.wrapper{ overflow:hidden;word-break:break-all;padding:10px; }\nh1,h2{ font-weight:normal; line-height:1.35; margin-bottom:.6em; }\nh3,h4,h5,h6{ line-height:1.35; margin-bottom:1em; }\nh1{ font-size:24px; }\nh2{ font-size:20px; }\nh3{ font-size:18px; }\nh4{ font-size:16px; }\nh5{ font-size:14px; }\nh6{ font-size:12px; }\np,ul,ol,blockquote,dl,table{ margin:1.2em 0; }\nul,ol{ margin-left:2em; }\nul{ list-style:disc; }\nol{ list-style:decimal; }\nli,li p{ margin:10px 0;}\nimg{ max-width:100%;display:block;margin:0 auto 1em; }\nblockquote{ color:#B5B2B1; border-left:3px solid #aaa; padding:1em; }\nstrong,b{font-weight:bold;}\nem,i{font-style:italic;}\ntable{ width:100%;border-collapse:collapse;border-spacing:1px;margin:1em 0;font-size:.9em; }\nth,td{ padding:5px;text-align:left;border:1px solid #aaa; }\nth{ font-weight:bold;background:#5d5d5d; }\n.symbol-link{font-weight:bold;}\n/* header{ border-bottom:1px solid #494756; } */\n.title{ margin:0 0 8px;line-height:1.3;color:#ddd; }\n.meta {color:#5e5c6d;font-size:13px;margin:0 0 .5em; }\na{text-decoration:none; color:#2a4b87;}\n.meta .head { display: inline-block; overflow: hidden}\n.head .h-thumb { width: 30px; height: 30px; margin: 0; padding: 0; border-radius: 50%; float: left;}\n.head .h-content { margin: 0; padding: 0 0 0 9px; float: left;}\n.head .h-name {font-size: 13px; color: #eee; margin: 0;}\n.head .h-time {font-size: 11px; color: #7E829C; margin: 0;line-height: 11px;}\n.small {font-size: 12.5px; display: inline-block; transform: scale(0.9); -webkit-transform: scale(0.9); transform-origin: left; -webkit-transform-origin: left;}\n.smaller {font-size: 12.5px; display: inline-block; transform: scale(0.8); -webkit-transform: scale(0.8); transform-origin: left; -webkit-transform-origin: left;}\n.bt-text {font-size: 12px;margin: 1.5em 0 0 0}\n.bt-text p {margin: 0}\n</style>\n</head>\n<body>\n<div class=\"wrapper\">\n<header>\n<h2 class=\"title\">\n“AI的IPHONE时刻”系列8：如何初步理解GPU算力？\n</h2>\n\n<h4 class=\"meta\">\n\n\n2023-04-03 18:25 北京时间&nbsp;&nbsp;&nbsp;<a href=http://gu.qq.com/resources/shy/news/detail-v2/index.html#/?id=nesSN202304031825158421d312&s=b><strong>广发证券股份有...</strong></a>\n\n\n</h4>\n\n</header>\n<article>\n<div>\n<p>核心观点：结论。GPU 峰值算力反映GPU 的主频和相应核心的硬件配置，如FP32 的核心的个数，FP64 的核心的个数以及Tensor core 的计算能力等峰值参数，并非其实际计算性能。GPU 峰值算力参数可用来初步区分其适用的应用场景，如在某GPU 内，其FP64 算力显著高于FP32 等算力，其可能更适用于科学计算等场景，如FP32、TF32、FP16、INT8 等算力较为均衡，其可能更...</p>\n\n<a href=\"http://gu.qq.com/resources/shy/news/detail-v2/index.html#/?id=nesSN202304031825158421d312&s=b\">网页链接</a>\n\n</div>\n\n\n</article>\n</div>\n</body>\n</html>\n","isBrief":false,"type":0,"news_type":1,"symbol":"LU0642271901.SGD","symbol_name":"Janus Henderson Horizon Global Technology Leaders A2 SGD-H","start_time":0,"source_url":"http://gu.qq.com/resources/shy/news/detail-v2/index.html#/?id=nesSN202304031825158421d312&s=b","article_id":"2324686508","we_media_id":null,"thumbnails":[],"rights":{"source":"tencent","url":"http://gu.qq.com/resources/shy/news/detail-v2/index.html#/?id=nesSN202304031825158421d312&s=b","rn_cache_url":null,"customStyle":"body{padding-top:10px;}#news_title{font-weight:bold;#titleStyle#;}#news_description span{font-size:12px;#descriptionStyle#;}.footer-note{#statement#}","selectors":".mod-LoadTzbdNews, body","filters":".relate-stock, .hot-list, .recom-box, .wx-sou","directOrigin":true},"url":"https://stock-news.laohu8.com/highlight/detail?id=2324686508","pubTimestamp":1680517513,"columns":[],"sourceInfo":{"source_id":"tencent","name":"腾讯"},"weMediaInfo":null,"summary":"GPU 峰值算力参数可用来初步区分其适用的应用场景，如在某GPU 内，其FP64 算力显著高于FP32 等算力，其可能更适用于科学计算等场景，如FP32、TF32、FP16、INT8 等算力较为均衡，其可能更适用于AI 训练和推理环节。GPU 算力参数从哪里来。由此可见，GPU 峰值算力与主频和硬件配置数量等参数相关。据Trendforce 数据，预估2022 年搭载GPGPU 的AI 服务器年出货量占整体服务器比重近1%，即约14 万台。","collect":0,"end_time":0,"defaultTopTitle":"qq.com","property":[],"viewcount":null,"language":"zh","relate_stocks":{"LU0642271901.SGD":"Janus Henderson Horizon Global Technology Leaders A2 SGD-H","LU2063271972.USD":"富兰克林创新领域基金","BK4551":"寇图资本持仓","LU0786609619.USD":"高盛全球千禧一代股票组合Acc","BK4554":"元宇宙及AR概念","LU0082616367.USD":"摩根大通美国科技A（dist）","BK4581":"高盛持仓","BK4527":"明星科技股","LU0719512351.SGD":"JPMorgan Funds - US Technology A (acc) SGD","LU1803068979.SGD":"FTIF - Franklin Technology A (acc) SGD-H1","BK4567":"ESG概念","LU1923623000.USD":"Natixis Thematics AI & Robotics Fund R/A USD","LU0348723411.USD":"ALLIANZ GLOBAL HI-TECH GROWTH \"A\" (USD) INC","LU0320765059.SGD":"FTIF - Franklin US Opportunities A Acc SGD","BK4141":"半导体产品","BK4534":"瑞士信贷持仓","LU1720051108.HKD":"ALLIANZ GLOBAL ARTIFICIAL INTELLIGENCE \"AT\" (HKD) ACC","LU1720051017.SGD":"Allianz Global Artificial Intelligence AT Acc H2-SGD","BK4503":"景林资产持仓","LU2125909247.SGD":"Natixis Thematics Meta H-R/A SGD","SG9999000418.SGD":"Aberdeen Standard Global Technology SGD","NVDA":"英伟达","BK4533":"AQR资本管理(全球第二大对冲基金)","LU0316494557.USD":"FRANKLIN GLOBAL FUNDAMENTAL STRATEGIES \"A\" ACC","BK4529":"IDC概念","LU0109391861.USD":"富兰克林美国机遇基金A Acc","LU1861215975.USD":"贝莱德新一代科技基金 A2","LU1923622614.USD":"Natixis Thematics Meta R/A USD","LU1548497426.USD":"安联环球人工智能AT Acc","LU1429558221.USD":"Natixis Loomis Sayles US Growth Equity RA USD","SG9999002224.SGD":"Allianz Global High Payout SGD","LU0417517546.SGD":"Allianz US Equity Cl AT Acc SGD","LU0708995401.HKD":"FRANKLIN U.S. OPPORTUNITIES \"A\" (HKD) ACC","LU1435385759.SGD":"Natixis Loomis Sayles US Growth Equity RA SGD-H","LU2125909593.SGD":"Natixis Thematics Meta R/A SGD","LU0889565833.HKD":"FRANKLIN TECHNOLOGY \"A\" (HKD) ACC","GB00BDT5M118.USD":"天利环球扩展Alpha基金A Acc","LU0056508442.USD":"贝莱德世界科技基金A2","LU1623119135.USD":"Natixis Mirova Global Sustainable Equity R-NPF/A USD","LU0640476718.USD":"THREADNEEDLE (LUX) US CONTRARIAN CORE EQ \"AU\" (USD) ACC","LU0097036916.USD":"贝莱德美国增长A2 USD","BK4579":"人工智能","BK4549":"软银资本持仓","LU1316542783.SGD":"Janus Henderson Horizon Global Technology Leaders A2 SGD","LU1712237335.SGD":"Natixis Mirova Global Sustainable Equity H-R-NPF/A SGD","BK4550":"红杉资本持仓","BK4548":"巴美列捷福持仓","LU0127658192.USD":"EASTSPRING INVESTMENTS GLOBAL TECHNOLOGY \"A\" (USD) ACC","LU0109392836.USD":"富兰克林科技股A","IE00B19Z9505.USD":"美盛-美国大盘成长股A Acc"},"translate_title":"\"IPHONE Moment of AI\" Series 8: How to understand GPU computing power preliminarily?","themeId":null,"isJumpTheme":false,"ttsUrl":null,"symbols_score_info":{"NVDA":0.9},"content_text":"核心观点：结论。GPU 峰值算力反映GPU 的主频和相应核心的硬件配置，如FP32 的核心的个数，FP64 的核心的个数以及Tensor core 的计算能力等峰值参数，并非其实际计算性能。GPU 峰值算力参数可用来初步区分其适用的应用场景，如在某GPU 内，其FP64 算力显著高于FP32 等算力，其可能更适用于科学计算等场景，如FP32、TF32、FP16、INT8 等算力较为均衡，其可能更适用于AI 训练和推理环节。首先我们引入数据类型。计算机运行必须考虑计算资源的分配，不同场景下使用不同的数据类型是必要的，数据类型即标定数据的正负、数据的范围大小、数据的精度大小。AI 芯片的主要应用场景包括训练、推理与高性能计算，根据英伟达官网的表述，AI 训练为缩短训练时间，主要使用FP8、TF32 和FP16；AI 推理为在低延迟下实现高吞吐量，主要使用TF32、BF16、FP16、FP8 和INT8；HPC（高性能计算）为实现在所需的高准确性下进行科学计算的功能，主要使用FP64。GPU 算力参数从哪里来。NVIDIA GA100 GPU 由多个GPU 处理群集(GPC)，纹理处理群集(TPC)，流式多处理器 (SM) 和HBM2 显存控制器等组成。A100 中总共有108 个SM，每个SM 有64 个FP32 CUDA 核，64个INT32 CUDA 核，32 个FP64 CUDA 核，以及4 个第三代Tensor Core，其中Tensor Core 专注于矩阵运算，在模型训练与推理的过程中，Tensor Core 将是主要的计算内核。GPU 峰值算力的测算公式为：峰值计算能力= GPU Core 的运行频率*GPU SM 数量*单个SM 对应的特定数据类型的指令吞吐量*2。根据该公式，可以得到A100 FP16（Tensor Core 加速）峰值算力为312T FLOPS，FP32（Cuda Core）峰值算力=19.5TFLOPS，与英伟达官方披露的性能参数一致。由此可见，GPU 峰值算力与主频和硬件配置数量等参数相关。模型训练&推理算力需求测算。根据我们的测算：在训练侧，训练一个GPT-3.5 175B 模型的NVIDIA A100需求为1080 个、AI 服务器需求为135 台；训练一个万亿参数量AI 大模型对A100、AI 服务器的需求分别为8521 个、1065 台。在推理侧，一个谷歌级应用使用GPT-3.5 175B 进行推理，对NVIDIA A100 需求为72 万个、对AI 服务器需求为9 万台；一个谷歌级应用使用万亿参数大模型进行推理，对NVIDIA A100 需求为378万个、对AI 服务器需求为47.3 万台。据Trendforce 数据，预估2022 年搭载GPGPU 的AI 服务器年出货量占整体服务器比重近1%，即约14 万台。我们以14 万台为基数测算训练、推理侧算力需求对AI 服务器出货量的拉动。根据测算，训练100 个GPT-3.5 175B 模型对AI 服务器出货量的拉动为9.6%，10 个使用GPT-3.5 175B 模型的谷歌级推理应用对全球AI 服务器出货量的拉动为643%。投资建议。建议关注算力芯片厂商：海光信息（与计算机组联合覆盖）、寒武纪（计算机组覆盖）、芯原股份、龙芯中科等；连接芯片厂商：源杰科技、澜起科技、裕太微、聚辰股份、帝奥微等；存储芯片厂商：深科技、兆易创新、东芯股份等；服务器硬件厂商：工业富联、沪电股份、胜宏科技等。AIGC 应用：海康威视、大华股份等。风险提示：AIGC 发展不及预 期， AI服务器出货量不及预期 ，国产厂商技术和产品进展不及预期。","kind":"news","is_publish_news":true,"is_publish_highlight":false,"is_publish_live":false,"is_publish_wemedia":null,"editions":null,"column":"","sentiment":"0","news_tag":"","news_rank":0,"symbols":[],"gpt_button":1,"need_auth":false,"code":"91000000","status":"200"},"commentList":[],"isCommentEnd":true,"newsSizeData":{"likeSize":0,"commentSize":0,"repostSize":0,"favoriteSize":1,"likeStatus":false,"favoriteStatus":false},"APP":{"userAgent":"Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; ClaudeBot/1.0; +claudebot@anthropic.com)","isDev":false,"isTTM":false,"tenantId":"TBCN","deviceId":"web-server-community-laohu8-v3","version":"4.36.2","shortVersion":"4.36.2","platform":"web","vendor":"web","appName":"laohu8","isIOS":false,"isAndroid":false,"isTiger":false,"isTHS":false,"isWeiXin":false,"isWeiXinMini":false,"isWeiBo":false,"isQQ":false,"isBaiduSwan":false,"isBaiduBox":false,"isDingTalk":false,"isToutiao":false,"isOnePlus":false,"isHuaWei":false,"isXiaomi":false,"isXiaomiWebView":false,"isOppo":false,"isVivo":false,"isSamsung":false,"isMobile":false},"href":"/m/news/2324686508","isCrawlerRequest":true}