Sounds nifty in concept, but no. This is a PCIe device. PCIe latencies are always slower than RAM. @intel's 3DXPoint (Optane drives) got us very close. There this idea of yours ultimately fails is that we're already limited in feasibility by the physical properties of PCB traces.

IM Flash….. 3DXPoint…. Intel Optane

从HBM到HBF：2027~2028商业化在即，AI推理先尝红利 From HBM to HBF: Commercialization Nears in 2027~2028, with AI Inference Benifits First 笔者注：韩国科学院（KAIST）教授，同时被称为“HBM之父”的Joungho (Jeong-ho) Kim前几日在“HBF 研究成果介绍及技术发展战略”上说，“在人工智能时代，每人将需要 100TB 的存储容量。这是一个名副其实的 “存算饥渴” 时代。未来，高带宽内存（HBM）将负责算力支撑的高速读写，而高带宽闪存（High Bandwidth Flash，HBF）将承担海量数据的存储需求”。当然，有的报道更是明确到了，“到2030年代中期，AI Agents普及时，每人将需要100TB的存储容量。”对的，WD也说为了应对数据暴增的时代，其在2029年的HDD产品容量将达到100TB。 “In the AI era, 100TB of memory per person will be required. It’s the so-called ‘memory-hungry’ era. In the future AI era, high-bandwidth memory (HBM) will handle speed, while high-bandwidth flash (HBF) will manage capacity.” Link to the report: chosun.com/english/industry-… 在HBM推动AI算力狂飙的同时，存储侧的瓶颈正在被重新定义。Kim教授明确提出，仅靠HBM已难以跟上AI演进节奏，行业需要一种兼顾高带宽与超大容量，且功耗显著更低的新型存储形态，高带宽闪存（HBF）。他预计，HBF商业化落地时间将落在2027年底至2028年初，并且初期将优先应用于AI推理场景。 说到这里，笔者只想感慨：Intel当年和Micron力推的3DXPoint（Persistent Memory），恐怕都要按不住棺材板了，不是它不对，而是生得太早。嗯，Intel起个大早，却注定难以凭一己之力撬动整个产业架构的迁移，只能在时代的节拍之外，提前把答案写在了试卷上。 为什么HBF会在“推理”先爆发？ Kim教授将HBM比作GPU旁的“专属书架”，而HBF更像“巨型图书馆”：前者解决热数据的极致带宽，后者面向多模态时代海量数据与特征库的数据访问与就近供给。 随着图像，视频，音频，文档深度融合，GPU即使算得快，也会被“从远端存储搬运数据”拖慢。这类瓶颈在推理侧更常见：推理并不总是需要像训练那样持续喂满算力，但却高度依赖大规模知识库，检索数据，历史特征的快速调取。 技术路线：把NAND做成“高带宽，可堆叠，靠近GPU”的存储层 从产业口径看，HBF被频繁描述为“NAND版HBM”：其核心思路是利用NAND的高密度与低成本优势，通过垂直堆叠与接口，封装，系统架构协同，把闪存从远端块存储推向“更靠近计算，具备高吞吐”的温数据存储。 HBF的定位不是替代HBM，而是补足HBM在容量与成本上的先天约束，形成“HBM（热数据） HBF（温数据）”的混合存储架构。 产业化进度：标准先行，样品在2026下半年，系统在2027 推动HBF从概念走向生态的关键一步，是规格与标准化。SanDisk与SK hynix在2025年8月宣布签署合作备忘录，联合推进HBF技术规范制定与生态构建，直指“下一代AI推理的容量与性能”缺口。双方公开口径中，目标是在2026年下半年提供HBF样品，并期待首批搭载HBF的AI推理设备在2027年初出现。 在此背景下，Kim教授给出的“2027年~2028年”商业化时间窗，与产业链“先标准，再样品，后系统”的节奏形成呼应：2026（样品/规格成熟）→ 2027（系统验证/推理先落地）→ 2027末~2028（规模化商业化）。 谁最可能吃到第一波红利？ Kim教授的判断是：HBF将为同时掌握HBM，NAND工艺与先进封装的厂商提供新机会，尤其是三星电子与SK Hynix这类具备“存储 封装”一体化能力的玩家，更容易把HBF做成可量产，可被系统接受的产品形态。 与此同时，行业也在探索多种实现路径。例如，Kioxia曾展示过面向AI的HBF原型模块思路：以更靠近I/O与总线的方式提供高吞吐与TB级容量，强调“更接近内存侧”的数据供给模式。 HBF真正改变的是“存储分层”和“系统架构成本曲线” 如果把AI系统看成“算力~内存~存储”三段式，过去两年大家主要在HBM上堆带宽；但接下来，决定推理性价比的往往是：你能否把更大规模的数据，更低功耗地放到更靠近GPU的位置。HBF的意义不止是“更快的NAND”，而是把NAND变成一个介于HBM与存储之间的一层： 🔹 容量扩张更容易（NAND密度与成本优势） 🔹 功耗压力更小（相比用HBM去硬顶容量） 🔹 推理场景更匹配（大知识库，多模态温数据就近访问） 未来两年值得重点观察三件事： 🔹 接口，协议与控制器架构是否统一。 🔹 封装与散热在机柜级能否规模化， 🔹 “HBM HBF”的软件栈（缓存，预取，数据放置策略）是否跟上。 这三者决定HBF能否从“工程样机”走向“可复制的系统产品”。 图片1：SanDisk HBF Concept 图片2：GPU with HBF，一个GPU4TB内存不是梦。

I made this point with 3dxpoint in my article. It needs to be able to outperform the DRAM-NAND memory pair/architecture. MRAM is fast and non-volatile memory but it's expensive. A memory trying to be the best of both worlds ends up being mediocre at both and can't find a place in the market. DRAM is fast and NAND is higher density and nonvolatile. When paired, it's the super star team.

Not sure if read where I brought 3DXPoint: My points are: Any memory trying to replace an existing memory space will likely fail because currently DRAM/NAND/HBM has the FULL backing of the AI market/demands. Any memory that tries to be best of both worlds will likely fail. (High Density & High Bandwidth). A memory will likely succeed if it tries to be good at only one thing. Why? Because memory is an architectural teamwork, HBM is paired with SOCAMM with CXL. It's going to be very robust. HBM doesn't work "alone" per se. 3DXPoint tried to be both low latency (fast) & high density. It ended in billions of failure because DRAM paired with NAND just does thing fine. I am not saying HBF has no future. It may succeed in some applications but it will have to compete with HBM SOCAMM SSD setups. HBF is based on NAND has cycle limits and wears out. It is not write-friendly.

HBF may or may not be useful in the market. I think Samsung and Micron are having doubts about it but working on it in case it finds a home in the market. The problem with HBF is that it promotes "read-only" inference applications. Writing on SLC NAND will degrade the floating gates. NAND has limited cycles so HBF is not ideal for high traffic environments. Trying to be good at two things can be the doom of it. It happened with 3Dxpoint memory and Micron/Intel completely abandoned it after billions of investment. I am thinking HBM paired with SOCAMM might be an alternative solution to HBF.

Here's some history in this article (3DXPoint Memory)

I will never comprehend why they sold out on that, it was their second largest tower and provided them with a ridiculous advantage, hell 3DXpoint memory would be going crazy right now.

they'd be more likely to revive 3dxpoint before their dram sector (both are never)

Bob Swan was a mistake, not only he stalled the company, he regressed it, remembers they had l Optane 3DXpoint and all that cool stuff? They missed Mobile, fabbing as it's the hottest commodity now and AI, it's another one for the history books like Yahoo.

Intel (and Micron) are actively silent about Optane. *Something* big happened on the legal side, and clearly Intel was pressured to stop developing and selling 3DXPoint. The 'official' story involves some disputes over IP more/less 'taken' from Stanford R. Ovshinsky's Estate. Unofficially, Intel may have pissed off Military Industrial contractors and may have accidentally started selling Rad Hardened memory w/o ITAR controls (chalcogenide PCMs like Optane are inherently rad-resistant)

Intel Optane / Micron 3dxpoint was just too far ahead of the game. Shame Micron murdered it for Intel. They'd surely be making serious bank off of Optane by now.

Why not opening 3DXpoint as a free, open standard, or cheap licensing business model a la ARM? High random IOPS define everyday usable performance, and NAND is no match for it. OS-sized, on-board chips for mid and high end applications would significantly improve UX overall.

30 years from now, on a flash drive? 🤣 Most NAND Flash manufacturers don't spec more than 5-10years Data Retention, even for Industrial pSLC. 3DXpoint(Optane) or True SLC NAND are the only Solid State Memory technologies on the open market that could retain data that long.

Oh, good. Something for me to pine after that isn't Optane/3DXpoint/Ovonic PCM 🙄

SSD全部Samsungにするかなぁ 3DXPoint系は除くが

我们组就是做这些破玩意的）距离量产十万八千年）还是纯数字的SRAM DRAM在5年内比较现实，没办法突破dram的scaling law，就只能协同优化）需要一个topdown design的llm machine）X 不过确实，新一代存储技术中，我觉得最有希望的就是STT-MRAM或者有人复活Intel的3DXpoint，其他的都只能水paper

まあここ数年というか2017年のKabyLake以降が引き金よね。 少なくとも10nmが遅れに遅れまくった結果、Appleからは刺されるしARMにもボコられるしAMDにはシェア取られるしNVIDIAにはAI周りで食われるしとマジでいいとこない。 オマケに最新世代は焼けるし、3DXpointも不振に終わったしマジでヤバい

Rynek kryptowalut i nowy system walutowy Bitcoin i OpenCoin (Ripple) narodziły się pod tym samym adresem IP - 64.202.167.192. Lokalizacja GPS? Dosłownie obok Phoenix, AZ, gdzie znajduje się placówka NSA. Główne centrum danych NSA znajduje się w Utah, zaledwie 5 minut jazdy od zakładów Micron, gdzie produkowana jest pamięć 3D XPoint, wykorzystywana w komputerach o pamięci hiperdimensionalnej. Badania nad oprogramowaniem prowadzone są na Uniwersytecie Kalifornijskim w Berkeley. Czy podać więcej Faktów? #Bitcoin #Ripple #OpenCoin #Kryptowaluty #NSA #PhoenixAZ #Berkeley #3DXPoint #CyberSecurity #Blockchain #TechMystery

3DXPoint、ランダムリードと耐久性はお化けなのでロマンなのです（目玉グルグル）