Wow, Tesla's Next Gen Chip Is NOT What We Thought ⚡️

以下是视频内容的摘要，涵盖了其中提到的所有新闻事项： **特斯拉的AI芯片与硬件开发：** * **AI5 流片 (Tape Out)：** 埃隆·马斯克宣布 AI5 已完成流片，AI6 和 Dojo 3 也正在开发中。并分享了一张 AI5 的设计图。 * **设计：** AI5 的特点是有一个中央系统级芯片 (SOC) 处理器，两侧各有 6 个存储芯片（共 12 个），它们都封装在同一块基板上。这与 AI4 的设计有显著不同。 * **生产：** 感谢台积电 (TSMC) 和三星的支持。预计它将成为有史以来产量最大的AI芯片之一。 * **应用：** AI5 主要为 Optimus 和特斯拉的超级计算机集群中的 AI 边缘计算进行了优化。 * **此前声明与当前：** 埃隆目前的声明澄清，AI5 的主要重点是 Optimus 和超级计算机，这与他早期（3 月 18 日、1 月 18 日）提到的将用于汽车/RoboTaxi 的说法有所不同。主持人推测，Optimus 对 AI5 的即时需求得到了优先考虑，但多余的芯片产能仍有可能用于车辆或数据中心。 * **流片定义：** 意味着设计已完成、锁定并发送给晶圆厂（三星/台积电）。 * **性能：** 据称，单个 AI5 的有效计算能力大约是双 SoC AI4 的五倍。主持人推测，考虑到冗余处理方式，双 AI5 系统的有效计算能力仍将是 AI4 的大约 5 倍。 * **冗余：** 主持人推测，特斯拉仍将在一个更大的 PCB 上使用两个 AI5 封装（SOC、内存、基板）以实现冗余，类似于 AI4。 * **封装 (COWAS)：** AI5 采用晶圆级基板上芯片封装 (COWAS)，这是一种先进的封装技术，将内存集成在基板上主处理器旁边。 * **内存集成：** 这与 AI4 有很大不同，AI4 的内存位于距离处理器更远的较大 PCB 上。AI5 的集成内存意味着更短的连接、更优越的内存延迟、更高的带宽、更好的效率以及整体推理性能。 * **内存容量：** 埃隆和特斯拉表示，AI5 的内存容量将是 AI4 的九倍。AI4 有 16GB 内存，这表明 AI5 可能有 144GB（9x16GB）。根据图像推测，LPDDR5X 模块可能带来 192GB（12 个芯片 x 16GB/芯片）。 * **流片灵活性：** 中央 SOC 的晶体管布局在流片后被锁定，但特斯拉保留了与不同数量、类型和配置的内存搭配的灵活性，以适应不同的使用场景（例如，Optimus 版本与 FSD 的轻量版）。 * **功耗：** 埃隆表示，AI5 的功耗可以降低到接近 250 瓦，这对 Optimus 来说是一个显著的改进。 * **目的：** AI5 是一款用于特斯拉 AI 软件的专用推理芯片，旨在超越市场上任何其他能满足特斯拉需求的产品。它能实现更好的上下文窗口、更快的数据访问、改进的反应时间以及对复杂场景的理解。 * **“闭环”：** X 平台上的 Shinaka 表示，AI5 是将 90 亿英里学习到的物理知识物理结晶到晶体管中，它基于软件学习而设计，并且工厂建在训练数据车辆的制造地。 * **制造地点：** 早期样品和生产可能来自韩国（由 KR-2613 指示），但大规模生产将在台积电和三星位于美国的工厂进行。 * **内存类型决定：** 特斯拉似乎选择了传统的行业标准 DRAM (LPDDR5X)，而非高带宽内存 (HBM)，这可能是为了降低成本、复杂性并简化制造，这也与埃隆早前关于 HBM 并非“必胜之选”的说法一致。 * **AI4 的能力：** * 埃隆表示，AI4 足以实现“远超人类安全水平的 FSD”。 * 特斯拉“非常有信心”AI4 将足以支持无人监督的 FSD，这得益于编译器/运行时重写。 * **硬件 3 车主：** 这被视为一个“非常强有力的轶事证据”，表明无人监督的 FSD（或“精简版”14/15）可能会应用于硬件 3 车辆，尽管市场对此持怀疑态度。主持人认为，由于软件效率提升（量化、蒸馏），这种情况发生的可能性不为零。埃隆已承诺，如果无人监督模型无法运行，将为 HW3 车主提供解决方案。 * **未来芯片 (AI6, AI6.5, AI7, Dojo 3)：** * **AI6：** 解决了 AI5 的设计妥协，拥有新的思路，目标是在相同的半掩模尺寸下，将性能相对于 AI5 真正翻倍，将使用三星在德克萨斯州的 2nm 晶圆厂。预计 2028 年年中实现量产。 * **AI6.5：** 将利用台积电在亚利桑那州的 2nm 工艺进一步提升性能。 * **SRAM 与内存带宽：** AI6/6.5 都将把大约一半的 TRIP AI 加速器专用于 SRAM，从而显著提高 SRAM 缓存中计算的有效内存带宽。 * **AI7 和 Dojo 3：** 被提及用于空间AI计算。 * **TerraFab 与制造：** * 特斯拉为 TerraFab 发布了“介电层沉积高级工艺工程师”和“光刻高级工艺工程师”的招聘信息。 * **介电层沉积：** 涉及三种不同芯片家族的工作：边缘推理处理器 (AI5)、用于轨道卫星的抗辐射芯片以及高带宽内存。 * **光刻：** 证实特斯拉计划使用 ASML 的 EUV 和 DUV 设备进行生产，开发并验证 EUV 和 RFI 抗蚀剂工艺。 * **AI5 流片体验：** 埃隆表示，最棒的部分是与团队合作，比参加派对更有趣。而“最不爽”的部分是为了快速推进而做出设计妥协，但他们最终提前 45 天完成了任务。 **全自动驾驶 (FSD) 软件与安全评分：** * **FSD 在欧洲：** * **瑞典：** 瑞典交通管理局重申，FSD 的欧盟范围批准仅适用于批准后生产的*新*车辆。对现有车辆的更新属于国家法规范围。瑞典尚未收到特斯拉/RDW 提交的国家批准申请（预计 5 月）。欧盟范围的批准尚无时间表。 * **比利时：** 一名内阁官员正在审阅一份信息，表明他们对此表现出积极兴趣。 * **FSD 14.3.1：** * 早期评价积极，并开始向早期体验组之外的用户广泛推送。 * 报告称性能更精细、更流畅，车道偏离和偏好设置有所改善。 * 注意到停车标志行为“大大改善”（果断停车，不犹豫，无二次停车），尽管主持人根据个人经验对此表示怀疑。 * **Grok 集成：** 展示了一个例子，特斯拉中的 Grok 可以设置基于位置的提醒（例如，“当我靠近家时提醒我买狗粮”）。 * **安全评分 V3：** * 一篇博客文章提供了更多信息。 * 基于超过 265 亿英里的驾驶数据，对未来碰撞风险进行更准确的评估。 * 安全因素现在*只*针对未启用 FSD 时的驾驶里程（手动驾驶行为）进行衡量。 * 启用 FSD 时的驾驶里程将自动获得 100 的安全评分，并计入*启用 FSD 时的*整体安全评分计算。 * 夜间驾驶的风险等级已得到细化。 * 可用性因车辆的信息娱乐系统和自动驾驶硬件而异。 * 一个例子显示，前两个月默认安全评分为 90，导致每月 260 美元的保费，如果经常使用 FSD，保费可能会降至每月 180 美元。 * 尚未广泛推出，但这表明特斯拉对 FSD 驾驶行为的信心日益增强。 * 现有保单持有人（加州以外）将在保单续保时升级到 V3，并提供各州具体的适用日期表格。 **市场反应：** * **特斯拉股票：** 今天上涨 7.62% 至 391.95 美元。纳斯达克指数上涨 1.4%，这意味着特斯拉表现超出 5.4 倍。 * **成交量：** 远高于平时，比 30 天平均水平高出约 78%。 * **猜测：** 主持人认为这可能归因于 AI4 足以支持无人监督 FSD 的保证、异常的期权活动，或者仅仅是“有人知道些什么”，但他建议顺其自然享受这份成果。 **赞助商：** * Surfshark：为美国客户新增了一项功能：高达 100 万美元的身份盗窃保险。它还提供替代身份、内容网页拦截器、无限设备支持，并最近完成了一次独立的安全性审计，没有发现任何关键漏洞。

Here's a summary of the video transcription, including every single news item mentioned: **Tesla's AI Chips & Hardware Development:** * **AI5 Tape Out:** Elon Musk announced the taping out of AI5, with AI6 and Dojo 3 also in the works. An image of the AI5 design was shared. * **Design:** AI5 features a central System-on-Chip (SOC) processor with 12 memory dies (six on each side) all packaged together on a substrate. This is a significant departure from AI4. * **Production:** Thanks to Taiwan Semi (TSMC) and Samsung for support. It's expected to be one of the most produced AI chips ever. * **Application:** AI5 is primarily optimized for AI edge compute in Optimus and Tesla's supercomputer clusters. * **Previous Statements vs. Current:** Elon's current statement clarifies AI5's primary focus on Optimus and supercomputers, a departure from earlier mentions (March 18, Jan 18) that it would go into cars/RoboTaxis. The host speculates that Optimus's immediate need for AI5 is prioritized, but excess chip capacity *could* still find its way into vehicles or data centers. * **Tape Out Definition:** Means the design is complete, locked, and shipped to the fabs (Samsung/TSMC). * **Performance:** A single AI5 is said to have around five times the useful compute of a dual SoC AI4. The host speculates a dual AI5 system would still be around 5x the useful compute compared to AI4 due to how redundancy is handled. * **Redundancy:** The host speculates that Tesla will still use two AI5 packages (SOC, memory, substrate) on a larger PCB for redundancy, similar to AI4. * **Packaging (COWAS):** AI5 uses Chip-on-Wafer-on-Substrate (COWAS), an advanced packaging technique, integrating memory right next to the main processor on the substrate. * **Memory Integration:** This is a big departure from AI4, where memory was on the larger PCB further from the processor. AI5's integrated memory means shorter connections, superior memory latency, improved bandwidth, better efficiency, and overall inference performance. * **Memory Capacity:** Elon and Tesla stated AI5 will have nine times the memory capacity of AI4. AI4 had 16GB RAM, suggesting AI5 might have 144GB (9x16GB). Speculation based on the image suggests LPDDR5X modules could lead to 192GB (12 dies x 16GB/die). * **Tape Out Flexibility:** The transistor layout of the central SOC is locked with tape out, but Tesla retains flexibility to pair it with different amounts, types, and configurations of memory for different use cases (e.g., Optimus vs. a lighter version for FSD). * **Power Consumption:** Elon said AI5's power consumption could be brought down closer to 250 watts, a significant improvement for Optimus. * **Purpose:** AI5 is a specialized inference chip for Tesla AI software, designed to outperform anything else available for Tesla's needs. It enables better context windows, faster data access, improved reaction times, and understanding of complex scenes. * **"The Loop is Closed":** Shinaka on X stated AI5 is the physical crystallization of 9 billion miles of learned physics into transistors, designed from software learnings, with the factory being built where training data cars are made. * **Manufacturing Location:** Early samples and production might be from Korea (indicated by KR-2613), but high-volume production will occur at TSMC and Samsung factories in the US. * **Memory Type Decision:** Tesla appears to have opted for conventional, industry-standard DRAM (LPDDR5X) instead of High Bandwidth Memory (HBM), likely to reduce cost, complexity, and ease manufacturing, aligning with Elon's earlier statements about HBM not being a "slam dunk." * **AI4 Capability:** * Elon stated AI4 is enough to achieve "much better than human safety for FSD." * Tesla is "pretty confident" AI4 will be sufficient for unsupervised FSD, due to compiler/runtime rewriting. * **Hardware 3 Owners:** This is seen as a "very strong anecdotal piece of evidence" that unsupervised FSD (or a "Lite" version 14/15) could come to Hardware 3 vehicles, despite market skepticism. The host believes there's a non-zero chance due to software efficiency gains (quantization, distillation). Elon has committed to taking care of HW3 owners if unsupervised models can't run. * **Future Chips (AI6, AI6.5, AI7, Dojo 3):** * **AI6:** Addresses AI5 design concessions, has new ideas, aims for a true doubling of performance over AI5 in the same half-reticle size, using Samsung's 2nm fab in Texas. Expected mid-2028 for volume production. * **AI6.5:** Will further improve performance using TSMC's 2nm in Arizona. * **SRAM & Memory Bandwidth:** Both AI6/6.5 will have around half of their TRIP AI accelerators dedicated to SRAM, increasing effective memory bandwidth significantly for calculations in SRAM cache. * **AI7 and Dojo 3:** Mentioned for space-based AI compute. * **TerraFab & Manufacturing:** * Tesla has job postings for "senior process engineer for dielectric deposition" and "senior process engineer for lithography" for the TerraFab. * **Dielectric Deposition:** Involves working on three distinct chip families: edge inference processors (AI5), space-hardened chips for orbital satellites, and high bandwidth memory. * **Lithography:** Confirms Tesla's plan to use both EUV and DUV machines from ASML for production, developing and qualifying EUV and RFI resist processes. * **AI5 Tape Out Experience:** Elon said the best part was working with the team, more fun than parties. The "least awesome" part was making design concessions to move fast, but they finished 45 days ahead of schedule. **Full Self-Driving (FSD) Software & Safety Score:** * **FSD in Europe:** * **Sweden:** The Swedish Transport Administration reiterated that EU-wide approval for FSD will only apply to *new* vehicles produced after approval. Updating existing vehicles falls under national regulations. No application from Tesla/RDW has been received yet for national approval in Sweden (expected May). No timeline for EU-wide approval. * **Belgium:** A cabinet official is reviewing a message, indicating proactive interest. * **FSD 14.3.1:** * Early reviews are strong, and it's starting to go wide beyond the early access group. * Reports of polished, smoother performance, improved lane bias and preference changes. * "Much improved" stop sign behavior (commits, no hesitation, no double stops) was noted, though the host expresses skepticism based on personal experience. * **Grok Integration:** An example was shown where Grok in a Tesla can set location-based reminders (e.g., "remind me to pick up dog food when I get close to home"). * **Safety Score V3:** * A blog post provides more information. * More accurate assessment of future collision risk, based on over 26.5 billion miles of driving. * Safety factors are now measured *only* against miles driven without FSD engaged (manual driving behavior). * Miles driven with FSD engaged automatically receive a safety score of 100, and are included in the overall safety score calculation *with FSD*. * Risk levels for late-night driving have been refined. * Availability varies by vehicle's infotainment system and self-driving hardware. * An example showed a default safety score of 90 used for the first two months, leading to a $260/month premium, which could drop to $180/month if FSD is used often. * Not yet widely available, but shows Tesla's growing confidence in FSD driving behavior. * Existing policyholders (outside California) will be upgraded to V3 at policy renewal, with a state-specific table provided for relevant dates. **Market Reaction:** * **Tesla Stock:** Up 7.62% to $391.95 today. The NDX was up 1.4%, meaning Tesla outperformed by 5.4x. * **Volume:** Much higher than usual, about 78% above the 30-day average. * **Speculation:** The host suggests this could be due to reassurance that AI4 will be enough for unsupervised FSD, unusual options activity, or simply "somebody knows something," but advises enjoying it for what it is. **Sponsor:** * Surfshark: Added a new feature for US customers: up to $1 million in identity theft coverage. Also offers alternative ID, content web blocker, unlimited devices, and recently completed an independent security audit with zero critical vulnerabilities.