ＳＴ０１０－１４

every time i'm in the CPU mines. i'm reminded that assembly is a really high level language

the distance from assembly to the physical machine is further than compiler IR to ASM on any reasonably interesting machine. It's fun to poke around them :3

if there was an award for coolest looking CPU ever i think the PowerPC G4 chips would win

void xorSwap (int* x, int* y) { if (x != y) { *x ^= *y; *y ^= *x; *x ^= *y; } } One of the most clever and the worst performing methods of swapping 2 numbers. Came across this gem as I attempted to learn prog from scratch, and got some good flashbacks to my second year when I found this trick, worked out a couple examples, and was adamant on using this everywhere I could!

coolest looking and coolest....you know...👀

Every humanoid robot being built right now depends on a $88M Japanese company that almost nobody in Western markets has looked at. While capital is flooding into robot software, AI brains, and system integrators, everyone is ignoring the one company that decides whether robotic joints actually function. And that oversight is the opportunity. The name is $6464.T — Tsubaki Nakashima Co., Ltd. Tesla Optimus requires 14 to 16 planetary roller screws per unit to actuate its limbs. Each screw is only as precise as the rolling elements housed inside it. Sub-micron Silicon Nitride ceramic balls and specialty steel rollers are the foundational material the entire physical AI execution layer runs on. Tsubaki Nakashima is the dominant global supplier of those elements. They supply SKF, Schaeffler, NSK, and Timken without competing with any of them. That neutral upstream position makes them the one indispensable node in the entire advanced motion industry. Every major bearing manufacturer on earth is their customer. Their proprietary in-house grinding and sintering equipment ensures no competitor can replicate their yield at scale. The Si3N4 precision ceramic ball market grows from $267M in 2025 to $612M by 2032 at a 12.54% CAGR. Their engineered plastics division, TN Plastics, adds another layer: a collaboration with a major robotics manufacturer achieved a 20% reduction in robot weight and a 1.3x improvement in operational speed by replacing steel elements with precision polymer components. In February 2026, TN Plastics launched microfluidic plate manufacturing, entering a medical diagnostics market projected to grow from $40B to $75B by 2030. Now the financials, because this is where most investors got it completely wrong: The company reported a ¥27.21B GAAP net loss in FY2025. Screeners flagged it as distressed. Most investors stopped reading there. What they missed: ¥16.7B was non-cash goodwill impairment. ¥6.4B was a non-cash inventory write-down on legacy steel stock. ¥2.9B was a deferred tax adjustment. None of it touched the cash account. Actual FY2025 cash generation: Operating Cash Flow ¥10.52B. Free Cash Flow ¥11.64B — an all-time company record. Q1 2026 confirmed the full return to GAAP profitability: Operating Profit: ¥1.13B (up 214.6% YoY) Net Profit: ¥308M (vs. a net loss of ¥559M in Q1 2025) EBITDA: ¥2.00B (up 62.1% YoY) The CEO executing this is Itaru Matsuyama — BCG, DuPont Electronic Materials, KKR Capstone. Appointed July 2024 with one mandate: private-equity-style transformation. He has divested the entire ball screw division to MinebeaMitsumi, closed plants in the Netherlands, Korea, and Tennessee, and suspended all dividends to direct 100% of cash toward debt reduction. The valuation: Tsubaki Nakashima trades at 0.31x to 0.34x Price-to-Book and 0.17x Price-to-Sales. Harmonic Drive Systems, supplying reducers for the same robotic joints, trades at 4.23x P/B and 10.24x P/S. A re-rating to just 0.8x P/S — still a massive discount to every robotics peer — implies 470% upside from current levels. The risks, because a real thesis requires them: European auto stagnation drove a 2.7% revenue decline in Q1 2026. Net debt sits at ¥58.2B. Tariff exposure is real across global facilities. Mitigations: Indian manufacturing expansion and TN Plastics medical microfluidics buffer the auto cyclicality. The ¥20B in subordinated debt is rated BBB Stable by R&I and doesn't mature until 2051. A Local for Local manufacturing model with US plants in Michigan and Georgia neutralizes tariff risk by supplying American customers from domestic, tariff-exempt soil. The market is pricing this as a failing cyclical auto supplier. What it actually is: the material science chokepoint of the entire physical AI supply chain, run by a KKR-trained CEO generating record cash and trading at a fraction of liquidation value. A $88M company sitting underneath the entire humanoid robotics buildout. NFA. Do your own due diligence.

继续思考， 华为在挑战里面没有谈散热，这是我比较诧异的。 目前两层堆叠，我觉得还有些散热的解法，但如果到三层Active Logic Stack或者更多之后，散热会从工程问题变成架构主问题。。。 目前流行的双层堆叠的技术AMD V-Cache，Intel Foveros和TSMC SoIC，还属于用冷cache叠热logic，因为SRAM功耗较低，热密度低，可用做Top Die，所以散热还能接受。 结构如下所示： SRAM |||| CPU 但华为的论文里是Logic-on-Logic。 即： Active Logic |||| Active Logic |||| Active Logic 这就完全不同了，这种多层Active Logic，热无法横向扩散，所以中间的Logic Die直接变成了烤箱，传统散热是完全扛不住的。 三层或者三层Active Logic堆叠之后，必须进入主动式散热时代！冷却液必须进入封装内部。 变成， Active Logic + Microfluidic Channel |||| Active Logic + Microfluidic Channel |||| Active Logic + Microfluidic Channel 液冷液冷液冷是关键！关键得说三次。。。 以后芯片设计里面需要Thermal Topology Architect，因为：热路径本身会决定Layout。 对的，本人的判断是：华为将来3层和3层以上的LogicFolding路径里面，Thermal将是最大的未解难题，甚至比EDA还难！