AI News AI资讯 17h ago Updated 1h ago 更新于 1小时前 52

Achieving Automation in Chip Design Verification, Boosting Development Efficiency by Over 10 Times, 'Zhiwei Chuangxin' Completes Angel Round Financing of Tens of Millions of Yuan | 36Kr Exclusive 实现芯片设计验证自动化,提升开发效率10倍以上,「智维创芯」完成数千万元天使轮融资|36氪首发

As the semiconductor industry charges forward amid AI frenzy, one of the most fundamental yet headache-inducing stages has become the tightest and most fragile link in the entire industry chain: chip verification. The first-time tape-out success rate can be as low as 14%, and verifying a single module consumes nearly two months of engineering work—these aren't sensational claims, but the harsh reality of the industry. Now, Zhiwei Chuangxin steps in with ChatDV, "the world's first large-model AI 在半导体行业被AI狂热裹挟着一路狂飙时,一个最基础却又最令人头疼的环节,正成为整个产业链条上那根绷得最紧、也最脆弱的弦。那就是芯片验证。首次流片成功率低至14%,一个模块的验证工作要耗费工程师近两个月的人力——这些不是危言耸听,而是行业现状。于是,智维创芯带着“全球首个数字芯片验证大模型智能体”ChatDV来了,宣称要将效率提升10倍以上。这听起来像是给在黑暗隧道里摸索的工程师们递来了一支强光手电。但问题是,这束光到底能照亮多远?

80
Hot 热度
70
Quality 质量
70
Impact 影响力

Analysis 深度分析

The spotlight of capital markets is almost entirely focused on chip manufacturing and AI computing power, as if building factories faster and stacking more computing power could solve all problems. But the reality is that design complexity is increasing exponentially, while verification remains the "slowest sailor" dragging the entire ship forward. It is highly experience-dependent, repetitive, and tedious—like a desperate arms race where engineers constantly write test cases to keep up with the complexity of chip designs. Traditional EDA tools are precise rulers, but they don’t think; general-purpose large models are knowledgeable chatbots, but they don’t understand the "grammar" and "hidden rules" of chips. What the industry needs is a seasoned craftsman who speaks the jargon and can roll up their sleeves to get the job done.

Zhiwei Chuangxin’s entry point is exactly this verification stage, which industry giants consider a "non-core profit area" yet everyone complains about. Based on publicly available information, their approach is pragmatic: rather than building a grand, all-encompassing platform, they focus on four of the most frequent and time-consuming pain points—test generation, assertion generation, reference model construction, and debugging. This "scalpel-like" precision is far more reliable than grand narratives claiming to do everything. Their "AI large model + data flywheel" logic also holds: AI generates initial content, simulation tools verify correctness, and results are fed back into the model for learning. This creates a closed loop that theoretically becomes more accurate with use.

However, the biggest challenge has never been the technical concept, but the real-world data barrier. Verification code and test platforms are among the core intellectual properties of chip companies and cannot be made public. Public datasets are like textbook exercises—far removed from real industrial-level complexity. Wang Xi said, "Data is the real moat," a statement that is half confident, half resigned. How Zhiwei Chuangxin acquires or generates enough "dirty" and realistic training data will directly determine whether ChatDV becomes a powerful tool for solving real problems or just a demo spinning idly on simple modules. Backed by the EDA National Innovation Center and universities, they may gain access to data from some national projects within regulatory frameworks, but bridging the trust gap with broad commercial client data remains a challenge.

From a commercialization perspective, they have wisely chosen to partner with companies like CEC, X-EPIC, and Thundersoft. These firms have verification needs and a degree of technical openness, making them ideal pilot scenarios. Their multiple business models—including licenses, integrated machines, and design services—aim to cover clients of different scales. However, it’s crucial to recognize that when integrating AI tools deeply into rigorous chip design workflows, clients’ primary concern is not whether it can work, but whether they dare to trust it. A single misjudgment leading to a tape-out failure can result in losses of tens of millions. Therefore, while ChatDV’s 89% automatic repair rate and several-fold improvement in accuracy are impressive, its reliability and explainability under extreme conditions will be key to convincing chip giants to pay. They need to understand why the model offers a particular recommendation, not just that it provides one.

Wang Xi’s team boasts an impressive background, combining academic origins with engineering capabilities. The Turing Award lab heritage and support from national platforms give the company a strong starting point and technical credibility. But this is also a double-edged sword. Academic DNA often pursues technical perfection, whereas the chip business world prioritizes stability, control, and risk aversion. Balancing cutting-edge exploration with engineering reliability is a challenge every team transitioning from academia to the market must face.

Their ultimate vision— "one-click chip generation," or even "chip AGI"—is extremely ambitious and exciting. But the current ChatDV is more like a solid foundationstone. Moving from module-level verification automation to IP-level and then subsystem-level is a long road. They claim to have discovered unknown bugs in classic RISC-V processors, which is undoubtedly a strong proof of concept. But chip design is a tightly interconnected systematic engineering effort. The coordination between front-end verification tools and back-end physical design tools, as well as alignment with foundry processes, still requires a long journey. Zhiwei Chuangxin states that it complements traditional EDA vendors, which is wise. However, will boundaries remain clear in the future? When the verification agent becomes powerful enough, might it encroach into the domains of design or synthesis?

Overall, Zhiwei Chuangxin has pinpointed a real and acute pain point in the semiconductor industry. Their technical roadmap is logical, and the team is capable. But they are attempting to introduce a highly innovative, data-intelligence-dependent variable into an industry that is deeply conservative and risk-averse. This requires time, patience, and above all, a few successful and scalable commercial cases to break the industry’s wait-and-see mindset. The "handicraft workshop" era of chip verification may indeed be coming to an end, but whether AI can become the key to unlocking automation—rather than just another complex module needing endless verification—will be answered by the market with its characteristic ruthlessness and fairness. The fuse of this efficiency revolution has just been lit.

在半导体行业被AI狂热裹挟着一路狂飙时,一个最基础却又最令人头疼的环节,正成为整个产业链条上那根绷得最紧、也最脆弱的弦。那就是芯片验证。首次流片成功率低至14%,一个模块的验证工作要耗费工程师近两个月的人力——这些不是危言耸听,而是行业现状。于是,智维创芯带着“全球首个数字芯片验证大模型智能体”ChatDV来了,宣称要将效率提升10倍以上。这听起来像是给在黑暗隧道里摸索的工程师们递来了一支强光手电。但问题是,这束光到底能照亮多远?

资本市场的聚光灯几乎全部打在芯片制造和AI算力上,仿佛只要厂建得够快、算力堆得够高,一切问题迎刃而解。但现实是,设计复杂度指数级上升,而验证环节依然是那个拖着整艘巨轮前进的“最慢的水手”。它高度依赖经验、重复且枯燥,像一场绝望的军备竞赛:工程师不断编写测试用例去追赶芯片设计的复杂度。传统EDA工具是精密的尺子,但它不会思考;通用大模型是个博学的聊天家,但它不懂芯片的“语法”和“潜规则”。行业需要的,是一个既懂行话、又能下场干活的“老师傅”。

智维创芯的切入点,正是这个被巨头们视为“非核心利润区”却又人人喊痛的验证环节。从公开信息看,他们的打法很务实:不搞大而全的通用平台,而是聚焦于测试生成、断言生成、参考模型构建和调试这四个最高频、最耗时的点。这种“手术刀式”的精准打击,比宣称能做一切的宏大叙事靠谱得多。其“AI大模型+数据飞轮”的逻辑也成立:用AI生成初始内容,用仿真工具验证对错,再把结果反馈给模型学习。这形成了一个闭环,理论上能越用越准。

然而,最大的挑战从来不在技术构想,而在于现实的数据壁垒。芯片公司的验证代码、测试平台,是它们最核心的知识产权之一,不可能公开。公开数据集如同教材习题,离真实的工业级复杂场景相去甚远。王翕说“数据是真正的护城河”,这话半是自信,半是无奈。智维创芯如何获取或生成足够多、足够“脏”、足够真实的训练数据,将直接决定ChatDV是解决实际问题的利器,还是只在简单模块上空转的Demo。他们背靠EDA国创中心和高校,或许能在合规框架下接触到一些国家级项目的数据,但这与广泛的商业客户数据之间,还隔着一道商业信任的鸿沟。

从商业落地看,他们选择了与中电科、芯华章、清微智能等公司合作,这是聪明的。这些公司既有验证需求,也有一定的技术包容度,是理想的试点场景。提供的License、一体机、设计服务等多重商业模式,也试图覆盖不同体量的客户。但必须警惕的是,将AI工具深度集成到严肃的芯片设计流程中,客户最大的顾虑从来不是“能不能”,而是“敢不敢”。一次误判导致的流片失败,损失是千万级的。因此,ChatDV的89%自动修复率、提升数倍的准确率,这些数据固然亮眼,但其在极端情况下的可靠性和可解释性,将是说服芯片巨头们掏钱买单的关键。他们需要知道模型为什么会给出某个建议,而不仅仅是它给出了一个建议。

王翕团队的背景堪称豪华,学术出身兼具工程化能力,图灵奖实验室的基因、国家级平台的支持,给了这家公司很高的起点和技术可信度。但这同样是一把双刃剑。学术基因容易追求技术的极致与完美,而芯片商业世界更看重稳定、可控和风险规避。如何平衡前沿探索与工程可靠性,是每一个从高校走向市场的团队必须面对的课题。

其终极愿景——“芯片一键生成”,甚至“芯片AGI”,极其宏大,也很性感。但眼前的ChatDV,更像是一块坚实的基石。从模块级验证自动化,到IP级,再到子系统级,这条路很长。他们声称已发现经典RISC-V处理器中的未知bug,这无疑是个有力的证明。但芯片设计是一个环环相扣的系统工程,前端验证工具与后端物理设计工具的协同、与芯片代工厂工艺的磨合,还有漫长的路要走。智维创芯称与传统EDA厂商是互补关系,这很明智。但未来边界是否清晰?当验证智能体足够强大时,是否会反过来侵入设计或综合的领域?

总的来说,智维创芯切中了半导体产业一个真实而尖锐的痛点,其技术路线有逻辑,团队有实力。但他们正试图在一个高度保守、风险厌恶的行业中,植入一个高度创新、依赖数据智能的变量。这需要时间,需要耐心,更需要拿一两个成功的、可大规模复制的商业案例,来打破行业的观望心态。芯片验证的“手工作坊”时代或许真的该结束了,但AI能否成为那把打开自动化大门的钥匙,而不只是另一个需要被反复验证的复杂模块,市场会给出最冷酷也最公正的答案。这场效率革命,才刚刚点燃引信。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

芯片 芯片 Agent Agent 大模型 大模型 产品发布 产品发布 融资 融资
Share: 分享到: