Tr8自主 LC VCO 设计系统 · 私有 · 展示页Autonomous LC VCO design · private · showcase
规格进去,
冻结的 VCO 出来。Hand it a spec.
Get back a frozen oscillator.
VCO Design System 是围绕开源的 VCO Design Studio 搭起来的完整设计系统:agent 读规格、查实测证据选结构,驱动商用电路仿真器做真实仿真,读诊断、自己修,最后把每一个控制码都回归一遍,冻结、交付。每一步都留回执,每个结论都只写测到的东西。VCO Design System is the complete design system built around the open-source VCO Design Studio. Agents read the spec, pick a structure from measured evidence, drive a commercial circuit simulator, read the diagnosis and repair the design themselves, then regress every control code before freezing and delivering. Every step leaves a receipt, and every claim is limited to what was measured.
五层 × 十二步:亮点是这一步用到的层,黄色虚线是两条返工回路。点某一列可停在那一步,再点一次继续播放。示意,不是仿真数据。Five layers × twelve steps: lit nodes are the layers a step uses; the dashed amber paths are the two rework loops. Click a column to hold that step, click again to resume. Illustration, not simulation data.
规模At a glance
不是一个脚本,
是一套系统。Not a script.
A system.
设计技能、流程程序、知识与证据、多 agent 协作,各自独立维护,靠共同的 ID 和来源连在一起。数字为当前状态。Design skills, the flow program, knowledge and evidence, and multi-agent collaboration are maintained separately and joined by shared IDs and sources. Figures reflect the current state.
架构Architecture
五层,
各守各的边界。Five layers,
each with a hard edge.
每一层只做自己的事,也写明不能替别的层做什么:知识不替你选架构,技能不能用叙述绕过仿真,程序不会把“工具跑完了”升级成“设计合格了”。点一层看细节。Each layer does its own job and states what it must not do for the others: knowledge does not choose the architecture, skills cannot talk their way around a simulation, and the program never upgrades “the tool finished” into “the design qualifies”. Select a layer for details.
- 负责Owns
- 不替代Never replaces
设计流程Design flow
十二步,
两段,两个回路。Twelve steps,
two phases, two loops.
所有结构共用同一条流程。设计段在三个代表码上做,交付段把每个码重测一遍;不达标就沿回路退回去重做。每个节点下方标着它用到哪几层。Every structure shares one flow. The design phase works on three representative codes; the delivery phase re-measures every code; a miss goes back along a loop. Each node shows which layers it uses.
设计段Design phase三个代表码:最低、中间、最高three representative codes: lowest, middle, highest
A · 起步A · Start
选出有实测依据的候选结构Shortlist structures backed by measured evidence
先确认够得着频带、能起振Confirm the tank can reach the band and start
B · 落地B · Realize
中心频率与带中功耗落到目标Put center and band-center power on target
整个控制范围起振一样稳Even startup across the control range
三个代表码上测相噪与 FoMPhase noise and FoM on three codes
C · 加固C · Harden
按结构设计谐波返回网络Design each structure's harmonic return
每个评估码都要能起振Every evaluated code must start
自动找修复并完整验证Find a repair, verify it in full
预算有余再找更优解Look for better while budget lasts
交付段Delivery phase每一个控制码every control code
交付前每个码都重测Re-measure every code before delivery
换数值设置复核,再冻结Re-check with other numerics, then freeze
设计、码表、回执和缺口一起交付Design, code table, receipts and gaps
最新进度Latest progress
36 个格子,
34 个通过。Thirty-six cells,
thirty-four pass.
六种结构 × 三个频率 × 每个频率两档功耗上限,在一套商用 CMOS 工艺上从模板起步全自动设计。格子里是最差码在 1 MHz 频偏处的相噪 FoM(dB);判定同时看中心频率、调谐范围 TR ≥ 30 %、频带中心的功耗和每个评估码的起振。Six structures × three frequencies × two power caps per frequency, designed fully automatically from templates in a commercial CMOS process. Each cell shows the worst-code phase-noise FoM at a 1 MHz offset (dB); a cell passes on center frequency, tuning range TR ≥ 30 %, band-center power and startup at every evaluated code.
原则Principles
每个失败,
都让流程变好一点。Every failure
makes the flow better.
系统不靠一次做对,而靠一个闭环:问题被测出来、被分类,修成对所有结构都成立的通用修复,经验写回知识层,之后的每一次设计都直接受益。The system does not rely on getting it right once. It relies on a loop: a problem is measured and classified, fixed as a general repair that holds for every structure, and the lesson is written back, so every later design benefits.
- 01
发现问题Find the problem
实测失败,带原因分类:流程可修,还是物理受限A measured failure, classified: flow-fixable or physics-limited
- 02
解决问题Solve it
做成通用的流程修复;复跑验证,另一模型家族独立审查A general flow fix, verified by re-running and independent cross-model review
- 03
沉淀经验Capture it
写进知识卡、实测证据、技能和规则Into knowledge cards, measured evidence, skills and rules
- 04
流程变好The flow improves
之后每一次设计都从更好的起点出发Every later design starts from a better place
贯穿全程的硬规矩:Hard rules throughout: 只写测到的;未知不等于不可行;起振判定失败即关闭;每次仿真都有回执;研究通过不等于产品通过。report only what was measured; unknown is not infeasible; startup fails closed; every simulation leaves a receipt; a research pass is not a product pass.
Crescendo 接口The Crescendo interface
参数搜索,
交给 Crescendo。Parameter search,
by the Crescendo engine.
系统里的参数搜索不另写优化器,而是调用 Crescendo 引擎的无头优化器。引擎负责提出候选,系统负责把每个候选真正仿真一遍。The system does not carry its own optimizer: parameter search calls the Crescendo engine's headless optimizer. The engine proposes candidates; the system actually simulates every one.
- 算法Algorithms
- ask/tell 种群优化:贝叶斯优化、差分进化、NSGA-II、随机搜索ask/tell population search: Bayesian optimization, differential evolution, NSGA-II, random
- 评价Evaluation
- 系统提供显式评价器,每个候选都跑真实仿真An explicit evaluator from the system runs a real simulation for every candidate
- 采纳Adoption
- 失败或不可行的候选一律不采纳Failed or infeasible candidates are never adopted
- 回执Receipts
- 每次运行都记下引擎版本与源码哈希Every run records the engine version and source hashes
- 版本Release
- 固定使用一份只读的引擎发布版本A pinned, read-only engine release
和 VCO Design Studio 的关系Relation to VCO Design Studio
开源的是底座,
这里是整栋楼。The foundation is open.
This is the building.
VCO Design Studio 负责电路本身:五个文献家族的拓扑、唯一的电气图、分层原理图和原生导出,MIT 开源。VCO Design System 在它上面加了会做设计的部分:流程程序、技能、知识与证据、多 agent 协作。后者是私有系统,不开源,本页只做展示。VCO Design Studio handles the circuit itself: five published topology families, one electrical graph, packaged schematics and native export, open source under MIT. VCO Design System adds the part that designs: the flow program, skills, knowledge and evidence, and multi-agent collaboration. It is private and not open source; this page is a showcase.
数据口径About the data
- 36 格进度表来自维护者本机在一套商用 CMOS 工艺上的前仿:典型工艺角、室温、标称电源,磁性元件为带显式损耗的理想模型;历史实测证据和系统规模数字按各自口径统计。不含 PVT、蒙特卡洛、版图与后仿,也不代表可靠性。The 36-cell progress table comes from the maintainer's local pre-layout simulation in one commercial CMOS process: typical corner, room temperature, nominal supply, ideal magnetics with explicit loss; historical evidence and system counts use their own scopes. No PVT, Monte Carlo, layout or post-layout results, and no reliability claim.
- 进度表用的是查表研究条件,和具体产品规格的验收分开;“通过”是研究意义上的通过。The progress table uses lookup research conditions, separate from acceptance against a product spec; “pass” is a research pass.
- 数字做了取整;进度表只标频率和功耗档,不给具体功耗值。设计方法的阈值、算法和参数有意不在本页公开。Figures are rounded; the progress table labels frequency and power tier only. The method's thresholds, algorithms and parameters are deliberately not published here.
- 首屏动画是示意,不是仿真数据。结构家族的出处见 VCO Design Studio 页面的参考文献。The hero animation is an illustration, not simulation data. Sources of the structure families are listed on the VCO Design Studio page.
其他工具Other tools