1
消息称 HBM4 价格明年有望翻倍，产能瓶颈与长期合同加剧供应紧张
IT之家 7 月 13 日消息，受到人工智能需求激增与产能结构性瓶颈的双重推动，高带宽内存（HBM）的价格到 2027 年有望翻倍。 据 DigiTimes 上周五报道，业内人士透露，下一代 HBM4 的价格有望升至每 Gb 4 至 5 美元甚至更高，而 2026 年下半年的价格约为每 Gb 2 美…
2
HBM 之父金正浩：AI 的本质是内存，GPU 真正工作的时间只有 10%-30%
HBM之父一语道破：AI计算的真正瓶颈不在GPU，而在内存，多数时间GPU都在“等待”。
IT之家 7 月 6 日消息，被誉为“HBM 之父”的韩国科学技术院（KAIST）电气系教授金正浩近日接受《东亚日报》采访时表示，AI 的核心竞争力正在从 GPU 转向内存。 金正浩认为 AI 的本质是内存，GPU 在 AI 推理中的利用率远低于理论水平。AI 每次输出结果，都必须先从 HBM 读取…
3
Buy or wait on local LLM hardware during the memory crunch
本地运行大模型正面临内存紧缺，是现在入手硬件还是继续观望？这篇文章帮你理清关键权衡点。
Article URL: https://thefrontierlab.ai/buy-or-wait-local-llm-hardware/ Comments URL: https://news.ycombinator.com/item?id=48760089 Points: 2 # Comment…
4
北大与 DeepSeek 联合开源 DSpark：破解 AI 大模型高并发推理瓶颈，速度提升 60% 至 85%
北大与DeepSeek联手开源DSpark，将大模型高并发推理速度提升60%-85%，破解瓶颈有实证。
IT之家 6 月 27 日消息，今日，DeepSeek 联合北京大学正式发布 DSpark 推理加速框架，旨在解决大语言模型在高并发生产环境中的推理效率瓶颈。 该框架已部署于 DeepSeek-V4-Flash 与 DeepSeek-V4-Pro 的预览版服务引擎中，相比此前生产环境采用的单 tok…
5
The AI Conundrum: We are living in highly subsidized, interesting times
AI从梦想到日常工具不到十年，但高度补贴的繁荣背后正撞上行业之墙。
If you trace the timeline of how LLMs went from a technologist's dream to early text-generation toys, to the world-shifting launch of ChatGPT, and fin…
6
闪迪新专利：将 NAND 闪存堆叠在计算芯片下方，破解存储瓶颈
IT之家 6 月 22 日消息，闪迪正在研发更多创新方案以解决存储容量受限问题，例如在芯片内部堆叠 NAND 闪存。 IT之家注意到，人工智能行业飞速发展，算力需求同步激增，各类性能瓶颈随之暴露，这倒逼 DRAM 与 NAND 闪存厂商跳出固有思维、探索突破性技术路线。 过去，芯片厂商只需推出全新存…
7
The Download: AI bottleneck debates, and BCI trials take off
AI发展遭遇瓶颈辩论，脑机接口临床试验全面启动，科技前沿动态速览。
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. …
8
视频转码性能最高提升 215%：AMD 线程撕裂者 Threadripper 处理器突破 HandBrake 瓶颈
IT之家 6 月 19 日消息，AMD 昨日（6 月 17 日）发布公告，通过修复 HandBrake 软件的线程管理限制， 线程撕裂者（ Threadripper）系列处理器视频转码性能最高提升 215%。 IT之家注：HandBrake 是一款免费开源的视频转码软件，支持转换视频编码格式（如 H…
9
I trust my car more than my AI agent
从对AI代理的信任缺失切入，揭示未来AI硬件与产品发展的核心瓶颈——信任差距，值得每个关注智能体的人思考。
Article URL: https://thoughts.jock.pl/p/ai-agent-future-where-this-goes-2026 Comments URL: https://news.ycombinator.com/item?id=48569889 Points: 2 # C…
10
Ask HN: What computer are you using for AI coding tools?
AI编码工具重度用户的硬件瓶颈实战分享，多代理配合Chrome调试拖垮M3 Pro的真实反馈。
I'm a heavy user of Claude code, had an M3 pro with 18gb ram for a few years. Until a few months ago it was always a beast. Now I'm spinning up more C…
11
What AI benchmarks miss about real-world performance
AI基准测试忽略的致命短板：买够GPU却卡在数据路径的突发瓶颈上。
Presented by F5 Enterprise AI teams have spent years solving for compute, securing GPU allocations, negotiating cloud capacity, and benchmarking train…
12
Is Text All You Need? Text as a Universal Information Bottleneck for Speech LLMs
探讨文本作为信息瓶颈在语音大模型中的核心作用，为连续声学信号集成提供新思路
arXiv:2606.09366v1 Announce Type: new Abstract: Large language models (LLMs) provide a powerful reasoning backbone for speech understanding, but integ…
13
AI Budget Is Growing. Your Returns Aren't
AI投资预算飙升，但回报为何停滞不前？Bain深度剖析企业AI落地的效率陷阱与破局之道。
Article URL: https://www.bain.com/insights/your-ai-budget-is-growing-your-returns-arent-heres-why/ Comments URL: https://news.ycombinator.com/item?id=…
14
Causal Neural Probabilistic Circuits
融合概念瓶颈与因果概率电路，支持测试时人工干预，提升模型可解释性与预测准确度。
arXiv:2603.01372v2 Announce Type: replace-cross Abstract: Concept Bottleneck Models (CBMs) enhance the interpretability of end-to-end neural networks …
15
The design bottleneck for solo founders? AI has solved it.
AI正成为独立创始人突破设计瓶颈的关键工具，助力小企业高效创业
Presented by Design.com Something significant is happening in how people build businesses. There are currently 29.8 million solopreneurs contributing …
16
Memory-Bound but Not Bandwidth-Limited: The Physical AI Inference Gap in Batch-1 LLM Decode
揭示LLM推理瓶颈新视角：batch-1解码受内存限制而非带宽限制，挑战传统认知。
arXiv:2605.30571v1 Announce Type: cross Abstract: Physical AI systems, including robots, autonomous vehicles, embodied agents and edge copilots, often…
17
Why LLM decode is memory-bound, not compute-bound
LLM推理核心瓶颈不在算力而在内存带宽，本文透彻解释为什么decode阶段受限于内存而非计算。
Article URL: https://github.com/harshuljain13/llm-inference-at-scale/blob/master/content/00_foundations/00.1_why_llm_inference_is_different/why_llm_in…
18
Amdahl's law for AI agents
用Amdahl定律揭示AI代理并行与串行的效率瓶颈，团队价值最大化的关键就在这里。
Article URL: https://electric.ax/blog/2026/02/19/amdahls-law-for-ai-agents Comments URL: https://news.ycombinator.com/item?id=48304513 Points: 2 # Com…
19
HiSpec: Hierarchical Speculative Decoding for LLMs
针对大模型推理中投机解码的验证瓶颈，提出层次化投机解码方法，显著提升推理速度。
arXiv:2510.01336v2 Announce Type: replace-cross Abstract: Speculative decoding accelerates LLM inference by using a smaller draft model to speculate t…
20
SIA: Self Improving AI with Harness & Weight Updates
提出SIA方法，让AI通过Harness与权重更新实现自我改进，打破人类调参瓶颈。
arXiv:2605.27276v1 Announce Type: new Abstract: Humans are the bottleneck in building and improving AI. Both the models and the agents that wrap them …