2025 (@2025Updated) profile picture
Joined January 2015
Photo by 2025 (@2025Updated)Photo by 2025 (@2025Updated)Photo by 2025 (@2025Updated)
Replying to @Boba_Ball_
When it's colder outside, more heat escapes through the walls and windows, leaving them colder than the air inside the room. The heating system is constantly fighting to replace this lost energy and maintain an air temperature of 19°C. However, the colder walls and windows emit less infrared radiation back towards our bodies, increasing our net heat loss. Consequently, we feel colder despite the thermostat reading exactly the same temperature. Basically, the same as @Andrew_int_wood said.
9
45
2
1,097
216,913
2025 retweeted
十年没写过重度 C++、一辈子没写过一行 Verilog 的软件工程师,能不能靠 Coding Agent 徒手搓出一颗跑大模型推理的硬件芯片? 开发者 Ankit Gupta 在 AWS 上租了一台带 16GB HBM 显存的 f2 FPGA 实例,把 Claude Opus 和 GPT-6.2 Astra 丢进 Conductor 云端循环里连续跑了几天。全程只烧掉 10,000 美元 Token(累计生成 300 万行实验代码、解析 3000 万行仿真波形输出),硬生生把一套 DeepGrove 三值权重(Ternary Weight)大模型的 FPGA 硬件推理引擎,从最初跑通逐比特对齐的 6.43 tok/s,经历 14 轮无人干预的架构迭代,一路干到了 213.13 tok/s(整整提速 33.13 倍)。 扒开它自动演进出来的 14 步硬件架构,有几个连资深 FPGA 工程师看了都会拍大腿的底层设计: 1. 彻底砍掉乘法器(No Multipliers):三值权重只有 -1、0、+1,Agent 直接用 2-bit 权重驱动一个 3:1 多路选择器(Mux) 来挑出 x、~x 或 0,闲置的第 4 个编码直接触发硬件错误标志用来抓权重内存损坏;取反操作(~x + 1)里的 +1 不在每个通道里做,而是延迟到行归约器(Row Reducer)每行统一只加一次,加上延迟执行 RMSNorm 缩放,直接比同速率的 INT8/INT16 架构省下了一半的 DSP 资源。 2. 32 通道菊花链直灌,跳过片上权重缓存:在 Batch=1 解码时,搬运权重比算数本身贵得多。引擎直接把 2-bit 权重从 HBM 流式直灌进计算阵列,完全跳过片上 Buffer;同时放弃沉重的全局交叉开关(Crossbar),改成 32 个 HBM 通道各自独立走一条菊花链,让 32 个内存通道全程满载。 3. 双 Bank 并行预取与烤机自动降频:第 14 轮迭代里,Agent 给注意力头加载器加上了双内存 Bank 并行读取(单步直接提速 17.3%)。更绝的是,它在 231.25 MHz 下跑长时间压力烤机(Soak Test)时自己抓到了偶发的随机硬件翻转错误,随后主动把时钟往回退了一档锁在 225/425 MHz,换取绝对稳定的 213 tok/s 交付。 以前大家觉得 Vibe Coding 的边界在前端和 CRUD,因为硬件底层有物理时序和面积死锁;但事实证明,只要验证信号是确定性的(仿真波形与延迟周期),Agent 在寄存器传输级(RTL)的搜索效率比人类还要恐怖。
10
33
4
214
17,996
2025 retweeted
SIA applauds Sens. John Cornyn (R-TX) and Mark Kelly (D-AZ) for introducing bipartisan legislation to extend the advanced manufacturing investment tax credit (AMIC) for five years. This proven incentive has helped catalyze more than $825 billion in private-sector semiconductor investments across 30 states, supporting hundreds of thousands of American jobs and strengthening our nation’s supply chain resilience and technology leadership. Extending the credit is essential to sustaining this momentum and ensuring the United States remains competitive in the global race for semiconductor and super intelligence leadership. We thank Senators Cornyn and Kelly for their leadership and urge Congress to advance this important legislation this year. Read our full statement here: semiconductors.org/sia-appla…
6
1
9
1,884
Look how far we’ve come
135
305
79
4,269
482,143
Replying to @PhysInHistory
Same idea explained visually
7
1
24
1,081
Replying to @elonmusk @garrytan
Honestly, deployment friction might actually be the best thing that could happen to the AI infra sector. It forces the AI buildout to get stretched over a much longer timeline. That keeps pricing power strong for the companies above, and it gives the applications side of AI enough breathing room to demonstrate commercial viability and catch up to the datacenter capex. Avoids having a big capacity spike followed by a glut.
60
25
7
1,043
331,459
2025 retweeted
GB: "So you think for internal Intel products you will be at parity with Taiwan Semi late 27, 28." LBT: "Yes" All I needed to hear. $INTC
Some highlights from the talk. $INTC "number one priority to get CPU to be the leader again and then XPU and then the GPU and eventually QPU (Quantum Processing Unit)" "we have the EMIB-T, the world best (packaging?) technology" "first outside customer beside the internal customer for 18A (18A-P?)" "need to have the defect density (...) the holy grail is 0.1 (...) 14 nanometer you really need to get to that level so that customers will be happy to use you and so that's why October we announced the 0.9PDK and then the Q1 I will have 1.0PDK is 0.1 and that can serve the customer" "18A yield is at 70%" "in 2027 we will have our first product on the 14A internally and then 28/29 will be outside customer to come in" GB: "So you think for internal Intel products you will be at parity with Taiwan Semi late 27, 28." LBT: "Yes" "(after earning) * 3.5 hour postmortem to evaluate with my team, 12 of 15 are new, I want to make sure they really work like a basketball team (...) for example, how do we order the memory, how to move quickly from PC shortage to data center. One of the problems we have is the architecture (...) * architecture of CPU for laptop and PC and data center are different so you cannot shift so quickly (...) we're going to fix that (...) * foundry and product supposed to be under PG, he wanted to separate (them) but we are still IDM company (...) (instead of having two separate weekly meeting) have one meeting to really have a discussion, debate, decide and communicate." "internal team working on Diamond Rapid and Coral Rapid with SMT (Simultaneous multithreading) coming in with the performance so we can catch up with AMD then we like to leap frog it and then we also have a team that I put together so that we can really spin in and they can do much faster and this team is only two months old they already have 141 people and really good CPU architects. (...) Intel used to have tick-tock (...) So I want to go back to the tick-tock, internal team and external team we can continue that way" GB: "has there been any change in demand that you see from AI customers?" LBT: "So far I don't see a lot of changes only the PC client side you know there's some softness. Data center and cloud and frontier model the demand is very strong and I just cannot supply enough. So overall, AI in my personal view is bigger than Internet and is much bigger. (...) CPU demand is going to continue. In fact I mentioned you know 40 billion to 200 billion in the next five years. That's a huge opportunity. CPU is really good at orchestration and also very good in reinforced learning and then agentic AI and that part I think I validate that and I see that I believe in that. So overall, I see that nothing going to slow down." GB: "A shared portfolio company (which is part of both GB's and LBT's VC portfolios?) at the absolute cutting edge is using Intel's EMIB-T (...) instead of TSM's alternative and is happy." Intel CEO Lip-Bu Tan on Saving Intel and Catching TSMC | TomorrowX Summit youtu.be/TMGyj5BFHv0?si=gx4h… via @YouTube
1
4
2
37
5,436
$INTC Intel CEO, Lip-Bu Tan's Interview with Gavin Baker, at TomorrowX Summit 2026: Watch here: youtube.com/watch?v=TMGyj5BF…
5
16
2
80
13,434
"TSMC is such a cyclical company" TSMC: $TSM
30
50
4
998
42,569
2025 retweeted
MIT 有门很特别的课程,叫《How to Make (Almost) Anything》,中文可以理解为“如何制造(几乎)任何东西”。 课程编号 MAS.863,官网就是一份现成的自学路线图。 这门课有意思的地方是,它不满足于让学生只会写代码、画电路图或者设计 3D 模型,而是要求他们真正把东西做出来。不按传统机械、电子、计算机专业拆分,而是逼着学生在三个月里每周攻克一个物理制造环节: - 结构与加工:CAD 参数化建模、激光切割、3D 扫描与打印、大型 CNC 数控铣削、模具设计与浇铸 - 电子与嵌入式:EDA 原理图与 PCB 布局、桌面雕刻机亲手铣板焊接、传感器输入与电机驱动输出 - 通信与智能:Mesh 组网、上位机界面开发,以及 2026 年课表里新加的 AI Agent 与边缘端 AI(Edge AI)实战 官网上不仅挂着完整的每周课表和工具清单,还能直接点进历届学生的个人主页,看他们每周踩了什么坑、怎么一步步把零件拼成能跑起来的机器。 课程主页:fab.cba.mit.edu/classes/MAS.…
74
1,047
34
7,796
372,132
2025 retweeted
Intel 刚发了一份「350W 新卡暴打英伟达 600W RTX PRO 6000 Blackwell、性能快 5.9 倍」的官方跑分。 刚看到标题第一反应是:Intel 疯了还是喝多了? 结果把测试报告和小字附录扒完,发现这波操作既鸡贼,又真捅到了老黄的肺管子: 1. 5.9 倍是怎么跑出来的?专挑老黄阉割的 1:64 软肋打 Intel 跑出 5.9 倍领先的项目,全是华尔街量化金融在用的双精度浮点(FP64)期权定价(欧式蒙特卡洛模拟 5.9x、Black-Scholes 5.6x、二叉树定价 4.6x)。 为什么差这么多?因为老黄为了保 B200 / H200 的天价利润,把工作站旗舰 RTX PRO 6000 Blackwell 的 FP64 双精度算力阉到了 1:64;而 Intel 这张 D270(Xe3P 架构) 给的是原生全速 FP64(每核心 64 个 FP64 FMA)。 更绝的是报告末尾的小字附录:Intel 的测试机里插了 1 张 D270 + 3 张 Arc Pro B70,英伟达那台插了 8 张 RTX PRO 6000,但小字里压根没写跑金融测试时到底开了几张卡。 2. 切回大模型推理,真实的能效比居然真能打 抛开田忌赛马的 FP64,真拿去单卡跑 Qwen3-30B-A3B(FP8),D270 的输出吞吐是 RTX PRO 6000 Blackwell 的 1.12 倍,跑 Drex 2.0 延迟低了 29%。 5.9 倍的噱头没了,但别忘了功耗:350W 打 600W,花一半出头的电,在 30B MoE 模型上正面反超了 Blackwell 工作站顶配卡。 3. 真正让本地大模型玩家流口水的是显存:单卡最高 480GB 老黄死卡显存容量,一张 600W 的 RTX PRO 6000 才给 96GB GDDR7。 Intel 这次干脆不抢又贵又缺的 HBM,直接在 350W 风冷 PCIe 卡上堆 LPDDR5X:公版起步就是 160GB(带宽 1.5 TB/s),板卡厂定制版单卡最高直接干到 480GB。 单张风冷 PCIe 卡塞进 480GB 显存是什么概念?以前得凑一整机柜的卡或者买天价集群才能塞下的满血超大模型权重,现在一张 350W 的卡就能全装进去。在 SC26 上 Intel 已经拉着 SambaNova 搞出了新玩法:用超大显存的 D270 专门吃下大模型海量上下文的 Prefill(预填充),再丢给高带宽的 SambaNova SN50 吐 Token。 老黄要是再接着搞小显存精准刀法,大显存推理这块肥肉是真要被这种「LPDDR5X 堆料怪」撬走了。
137
309
58
2,863
267,677
2025 retweeted
Replying to @mylifcc
计算物理真的刚需双精度,要是真的能原生1:1 fp64并且驱动还不出什么大问题的话应该会有很多实验室买的吧。 我以前做的张量网络研究就是终极吃显存大户
2
1
110
10,801
Replying to @LIWEI_TWCapital
2028 is mostly tpu 8i max for anthropic (10gw)...and it's a dual tpu 8i chip package...meaning 2x the TDP of tpu 8i...thus needs half the number of units vs tpu 8i..so $ value is still much higher vs 2027
1
1
7
252
I expect the market to go through a brutal correction over the next 6-10months. I'm seeing 75% drawdown in $NVDA, 70% in $MU. All the way down the AI trade chain. Application layer will be massive benifitiaries of this capital unlock imo. This means $LMND $UPST and others have a good chance to really outperform, they are isolated beneficiaries of AI adoption at inflection. Perhaps I'm wrong but that capital has to go somewhere, and bonds are not appealing. Such a hypothetical drawdown would result in rate collapse + consumer exiting recession. Huge tailwind once this massive flush is complete. NFA
One of many exceptionally bearish charts I’m seeing in the semi’s/hyper scaler list. Though, perhaps gains and free money continue in perpetuity. Who knows.
41
33
6
388
217,183
2025 retweeted
Oracle has secured all the funding we need to fulfill our contract with OpenAI. With our final capital raise earlier this month, we became the first hyperscaler fully funded for this investment cycle. The financing we outlined at Oracle Investor Day, including vendor financing, is in place. Any further private discussions with vendors are just additional options. Our ability to deliver on our OpenAI contract does not depend on those discussions. Any media reports that will come out and say the contrary are incorrect and wrong. Oracle and OpenAI are making rapid progress on our infrastructure partnership. Our Abilene campus, where Astra-6 was trained, is nearly complete. Shackelford, our next gigawatt-scale campus, is fully permitted. The handover of NVIDIA Vera Rubin racks in its data halls is on track for the coming weeks. We continue to expect positive free cash flow in the second half of FY2028. We are targeting more than $30 billion in free cash flow in FY2029 and expect to self-fund future growth beginning that year.
601
526
641
5,625
2,964,381
2025 retweeted
投资里最反人性的时刻又来了: 2024—2025 年美国科技股远期市盈率超过 30 倍的时候,所有人都在抢着上车;如今估值一路砸到 16 倍左右,直接追平 2020 年 3 月疫情熔断低点(甚至比 2022 年熊市底部还低),大家反而开始垂头丧气、割肉离场了。 过去十年,这种级别的估值窗口只出现过两次。市场从来不会在人声鼎沸时打折,便宜筹码永远只出现在情绪最冰冷的时候。
42
21
3
194
45,590
I was always bearish on $NVDA even at $200 when people kept pressing me due its monopoly and low forward P/E. I'm doubling down. I think AI chip design is not a moat anymore. $NVDA was the first mover. Everyone else catching on like $AVGO. Broadcom is priced like a duopoly gaining 10x in a matter of 3 years. I think it will disappear too. You also have INTC, AMD, ARM, CBRS, and MediaTek(few other private companies). I think there is a case for moat for foundry like TSMC and INTC. The market is incredibly good at chasing returns. Any significant edge will be chipped down.
35
4
2
241
74,430
我最新的產業調查顯示,Google 下修 2027 年、由 Broadcom 設計的 TPU 8i 系列共約 100 萬顆(以單價較高的 8i Max / Whalefish 為主)。此下修約佔 2027 年 TPU 8i 系列出貨量約 20%。 本次調查也指出,Google 預計增加 2027 年由聯發科設計的 TPU 8t 系列出貨,但細節尚未確定。Google 近幾個月反覆強調前沿模型研發優先取得算力,此次 8t 與 8i 系列出貨比重的改變,正是在供應鏈緊張與資料中心供電受限下所做的調整。
8
36
2
551
71,670