[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"content-doc-dcioacb15da0":3},{"user":4,"document":8,"mainDocument":27,"columnUrl":29,"subscription":30,"footer":42,"text":80},{"isAuthenticated":5,"isAdmin":5,"displayName":6,"avatarUrl":6,"nid":6,"groupLevel":7},false,"",-10,{"id":9,"fullTitle":10,"subTitle":6,"url":11,"columnId":12,"columnName":13,"columnUrl":14,"summary":6,"contentHtml":15,"mainContentHtml":6,"posterUrl":16,"createDate":17,"displayDate":18,"displayDateSlash":19,"pageviews":20,"tags":21,"hidden":5,"isSubContent":5,"replyDocOrTargetId":6,"contentType":23,"videoId":6,"liveVideoUrl":6,"useContentVideo":5,"duration":24,"price":24,"priceText":25,"priceBadgeText":25,"priceBadgeClass":26,"freeForMinGroupLevel":24,"redirectUrl":6,"readyToStream":5},"dcioacb15da0","闪迪HBF不是HBM替代品，而是AI推理时代新型容量方案","\u002Fdoc\u002Fdcioacb15da0","col18178739ee","美股资讯","\u002Fcol\u002Fcol18178739ee","\u003Cp>\n\u003C\u002Fp>\u003Cp>\u003Cspan>SemiAnalysis \u003Cspan>这次讨论\u003C\u002Fspan> HBF \u003Cspan>高带宽闪存，最重要的结论不是“它能不能替代\u003C\u002Fspan> HBM\u003Cspan>”。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>恰恰相反，\u003C\u002Fspan>HBF \u003Cspan>不是\u003C\u002Fspan> HBM \u003Cspan>的替代品，而是\u003C\u002Fspan> AI \u003Cspan>推理进入细分场景后，出现的一种新型容量方案。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>过去大家看\u003C\u002Fspan> AI \u003Cspan>内存，核心矛盾是带宽。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>HBM \u003Cspan>解决的是高带宽、高并发、高性能训练和推理需求。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>但随着长上下文、\u003C\u002Fspan>MoE \u003Cspan>混合专家模型、小\u003C\u002Fspan> batch \u003Cspan>推理越来越多，另一个问题开始变得重要：\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>模型太大，能不能完整放进一台机器里。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>这就是\u003C\u002Fspan> HBF \u003Cspan>的机会。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>它更适合顺序读取，比如加载模型权重；但它并不擅长小规模随机散读。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>所以它最理想的应用场景，不是超大云厂商的大规模\u003C\u002Fspan> HBM \u003Cspan>资源池，而是少量\u003C\u002Fspan> GPU \u003Cspan>的本地化、私有化企业部署。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>但问题也不少\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>第一，成本还不清楚。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>现在行业里能看到的成本说法，基本都还是厂商宣传口径。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>闪迪说单位比特成本大约是\u003C\u002Fspan> HBM \u003Cspan>的\u003C\u002Fspan> 1\u002F8\u003Cspan>；\u003C\u002Fspan>Hot Chips \u003Cspan>上也有说法认为每\u003C\u002Fspan> GB \u003Cspan>成本可能是\u003C\u002Fspan> SSD \u003Cspan>的\u003C\u002Fspan> 10 \u003Cspan>倍左右。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>规模化之后，良率、测试、控制器摊销这些成本确实可能下降。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>但\u003C\u002Fspan> TSV\u003Cspan>、堆叠、堆叠后全栈测试、\u003C\u002Fspan>pSLC \u003Cspan>高速\u003C\u002Fspan> NAND \u003Cspan>模式这些成本是结构性的，很难完全消除。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>就像\u003C\u002Fspan> HBM \u003Cspan>永远不可能做到普通\u003C\u002Fspan> DRAM \u003Cspan>的价格，\u003C\u002Fspan>HBF \u003Cspan>也不太可能变成普通\u003C\u002Fspan> SSD \u003Cspan>的价格。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>第二，耐久度是最大变量。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>目前擦写寿命规格还没公开。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>如果单元磨损快于预期，厂商就必须配置更多冗余容量，实际成本会继续上升。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>第三，散热和可靠性还没有真正被验证。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>闪存本身并不喜欢高温环境，但\u003C\u002Fspan> HBF \u003Cspan>的设计却可能要靠近高热\u003C\u002Fspan> GPU\u003Cspan>。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>现在看到的方案，比如用\u003C\u002Fspan> UCIe \u003Cspan>把\u003C\u002Fspan> HBF \u003Cspan>和\u003C\u002Fspan> GPU \u003Cspan>物理分开，或者在\u003C\u002Fspan> 85\u003Cspan>℃环境下每天刷新全盘数据，听起来都还没有经过大规模验证。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>尤其是后者，对超大规模云厂商来说，运维负担可能很难接受。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>美股投资网认为，\u003C\u002Fspan> HBF \u003Cspan>目前更像一个“期权级变量”。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>它有可能打开一类新的\u003C\u002Fspan> AI \u003Cspan>推理部署方式，但还不是成熟的大规模通用方案。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>更大的行业信号是：\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>内存市场正在从标准化大宗商品，走向分层、专用化。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>对投资来说，\u003C\u002Fspan>HBF \u003Cspan>最大的意义不是马上颠覆\u003C\u002Fspan> HBM\u003Cspan>，而是提醒市场：\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>AI \u003Cspan>硬件已经从“通用算力扩张”，进入“专用架构分层”的阶段。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>谁能解决具体场景下的容量、带宽、功耗、散热和成本问题，谁就可能拿到下一轮\u003C\u002Fspan> AI \u003Cspan>基础设施的增量价值。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>$SNDK $MU $SKHY $WDC $STX #\u003Cspan>美股\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>━━━━━━━━━━━━━━━━━━━━━━\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>【原创声明】\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>本文由美股投资网（\u003C\u002Fspan>TradesMax.com\u003Cspan>）研究团队原创完成。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>原创机构：美股投资网\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>英文品牌：\u003C\u002Fspan>TradesMax\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>本文基于公开财报、\u003C\u002Fspan>SEC\u003Cspan>文件、市场数据、机构资金、期权交易及产业链信息进行独立分析\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>如需引用本文，请注明：\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>\u003Cspan>“美股投资网（\u003C\u002Fspan>TradesMax.com\u003Cspan>）原创研究”。\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>━━━━━━━━━━━━━━━━━━━━━━\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003C\u002Fp>","https:\u002F\u002Fwww.tradesmax.com\u002Fimages\u002Fa_Stock\u002FS\u002FSNDK\u002FSNDK.jpg","2026-08-28T15:11:49","2026.08.28","2026\u002F08\u002F28",37564,[22],"SNDK","Article",0,"免费","success",{"id":9,"fullTitle":10,"subTitle":6,"url":11,"columnId":12,"columnName":13,"columnUrl":14,"summary":6,"contentHtml":15,"mainContentHtml":6,"posterUrl":16,"createDate":17,"displayDate":18,"displayDateSlash":19,"pageviews":20,"tags":28,"hidden":5,"isSubContent":5,"replyDocOrTargetId":6,"contentType":23,"videoId":6,"liveVideoUrl":6,"useContentVideo":5,"duration":24,"price":24,"priceText":25,"priceBadgeText":25,"priceBadgeClass":26,"freeForMinGroupLevel":24,"redirectUrl":6,"readyToStream":5},[22],"\u002Fcol\u002Fstocknews",{"visible":5,"marketingHtml":31,"services":32,"recentDocuments":41},"\u003Cfigure class=\"image\">\u003Ca href=\"https:\u002F\u002Fstockwe.com\u002Fdoc\u002Fdcio537efad5\" rel=\"noopener noreferrer\">\u003Cimg src=\"\u002Fimg\u002Fstockwewebfiles\u002Fweb-202408-stk\u002F1586109431mceclip0.jpg\">\u003C\u002Fa>\u003C\u002Ffigure>\u003Cdiv class=\"text-center\">\u003Ch2 class=\"card-title mx-auto\">\u003Cbr>\u003Ca rel=\"noopener noreferrer\" href=\"https:\u002F\u002Fstockwe.com\u002Fdoc\u002Fdcio537efad5\">案例介绍：英伟达深度研究报告\u003C\u002Fa>\u003C\u002Fh2>\u003C\u002Fdiv>",[33,37],{"productId":34,"serviceName":35,"priceText":36},"prod_PPxdDdK87QaiLv","月付","$12.95美元",{"productId":38,"serviceName":39,"priceText":40},"prod_PPxeMs3bix1da5","年付","$149.00美元",[],{"links":43,"images":71,"summaryHtml":76,"aboutTitle":77,"aboutHtml":78,"copyrightHtml":79},[44,47,50,53,56,59,62,65,68],{"label":45,"url":46},"深度报告","\u002Fcol\u002FdepthReport",{"label":48,"url":49},"VIP会员","\u002Fvip",{"label":51,"url":52},"期权推荐","\u002FOption",{"label":54,"url":55},"低价暴涨股","\u002FPenny",{"label":57,"url":58},"AI智能体","\u002FAiAgent",{"label":60,"url":61},"常见问题","https:\u002F\u002Fstockwe.com\u002FFAQ",{"label":63,"url":64},"美股课程","\u002Fcol\u002Fvideos",{"label":66,"url":67},"免责声明","\u002Fdisclaimer",{"label":69,"url":70},"联系我们","\u002FContactUs",[72,73,74,75],"\u002Fimg\u002Fstockwebsiteblob\u002Fweb-202509-stk\u002FUploaderzic2tuwsol2_2025_09_11_18_21_07.gif","\u002Fimg\u002Fstockwebsiteblob\u002Fweb-202509-stk\u002FUploadercakzdvydksw_2025_09_03_09_00_56.png","\u002Fimg\u002Fstockwebsiteblob\u002Fweb-202509-stk\u002FUploadergtjyagwvoyk_2025_09_14_08_32_05.png","\u002Fimg\u002Fstockwebsiteblob\u002Fweb-202509-stk\u002FUploader3u0tt4jhlqh_2025_09_23_22_30_48.png","邮箱: buy@TradesMax.com 美国电话 626-378-3637","公司介绍","\u003Cp class=\"MsoNormal\">美股大数据 \u003Ca href=\"https:\u002F\u002Fstockwe.com\" rel=\"noopener\">StockWe.com\u003C\u002Fa> 是一个美国领先的金融和美股信息大数据提供商，紧盯华尔街金融市场和行情，2008年成立于美国硅谷，创始人是前纽约证券交易所资深分析师Ken，联合多位摩根斯坦利分析师，谷歌 Meta工程师利用AI和大数据，配合十多年美股实战经验和业内量化交易模型，每天处理海量股票数据：挖掘潜力大牛股，捕捉期权异动大单，实时主力资金流向、机构持仓变化、川普突发新闻，美股买卖信号第一时间发到您手机APP。\u003C\u002Fp>","专业美股投资者都在这里",{"loading":81,"search":82,"searchPlaceholder":82,"hotContent":83,"draft":84,"noData":85,"searchNoData":86,"courseContent":87,"more":88,"buyNow":89,"subscribeNow":90,"encoding":91,"paidContent":92},"Loading...","搜索","热门内容","草稿","目前没有任何内容公布","当前检索内容没有数据","课程内容","更多","立即购买后观看","- 立即订阅 -","视频编码中...","付费内容"]