原文标题： Exclusive: Nvidia working on new AI chip for China that outperforms the H20, sources say 原文连结：https://tinyurl.com/w7r85aur 发布时间：August 19, 20252:54 PM GMT+8 记者署名：Liam Mo and Fanny Potkin 原文内容： BEIJING/SINGAPORE, Aug 19 (Reuters) - Nvidia (NVDA.O) is developing a new AI chip for China based on its latest Blackwell architecture that will be more powerful than the H20 model it is currently allowed to sell there, two people briefed on the matter said. 北京/新加坡，8月19日（路透社）——两位知情人士表示，辉达（Nvidia，股票代码 NVDA.O）正在为中国开发一款基於最新 Blackwell 架构的新 AI 晶片，其性能将超过目 前允许在中国销售的 H20 型号。 U.S. President Donald Trump last week opened the door to the possibility of more advanced Nvidia chips being sold in China. But the sources noted U.S. regulatory approval is far from guaranteed amid deep-seated fears in Washington about giving China too much access to U.S. artificial intelligence technology. 美国总统川普上周暗示，可能允许更先进的辉达晶片在中国销售。但消息人士指出，由於 华府对於中国过度获取美国 AI 技术的忧虑根深蒂固，美国监管批准仍远未有保证。 The new chip, tentatively known as the B30A, will use a single-die design that is likely to deliver half the raw computing power of the more sophisticated dual-die configuration in Nvidia's flagship B300 accelerator card, the sources said. 消息人士称，这款暂定名为 B30A 的新晶片，将采用单晶片（single-die）设计，计算效 能大约是辉达旗舰产品 B300 加速卡双晶片（dual-die）架构的一半。 A single-die design is when all the main parts of an integrated circuit are made on one continuous piece of silicon rather than split across multiple dies . 单晶片设计指的是所有主要的集成电路元件都制作在一整块矽片上，而不是分散在多个晶 片上。 The new chip would have high-bandwidth memory and Nvidia's NVLink technology for fast data transmission between processors, features that are also in the H20 - a chip based on the company's older Hopper architecture. 这款新晶片将具备高频宽记忆体和辉达的 NVLink 技术，以实现处理器之间的高速数据传 输，这些功能也存在於基於较旧 Hopper 架构的 H20 晶片上。 The chip's specifications are not completely finalised but Nvidia hopes to deliver samples to Chinese clients for testing as early as next month, said the sources who were not authorised to speak to media and declined to be identified. 消息人士表示，这款晶片的规格尚未完全确定，但辉达希望最早能在下月向中国客户提供 样品进行测试。这些人士未获授权公开发言，因此要求匿名。 Nvidia said in a statement: "We evaluate a variety of products for our roadmap , so that we can be prepared to compete to the extent that governments allow." 辉达在声明中表示：「我们会评估多种产品纳入研发蓝图，以便在政府允许的范围内保持 竞争力。」 "Everything we offer is with the full approval of the applicable authorities and designed solely for beneficial commercial use," it said. 「我们提供的所有产品均经过相关主管机关完全批准，并仅为正当商业用途而设计。」 The U.S. Department of Commerce did not respond to a Reuters request for comment. 美国商务部未回应路透社的置评请求。 FLASHPOINT 争议焦点 The extent to which China, which generated 13% of Nvidia's revenue in the past financial year, can have access to cutting-edge AI chips is one of the biggest flashpoints in U.S.-Sino trade tensions. 中国在上一财年占辉达营收的 13%，其能获取多少尖端 AI 晶片，是美中贸易紧张局势中 的最大争议之一。 Nvidia only received permission in July to recommence sales of the H20. It was developed specifically for China after export restrictions were put in place in 2023, but company was abruptly ordered to stop sales in April. 辉达直到今年 7 月才获准恢复 H20 的销售。该晶片原本是因 2023 年出口限制而专为中 国设计，但公司在 4 月突然被勒令停止销售。 Trump said last week he might allow Nvidia to sell a scaled-down version of its next-generation chip in China after announcing an unprecedented deal that will see Nvidia and rival AMD give the U.S. government 15% of revenue from sales of some advanced chips in China. 川普上周表示，在宣布一项前所未有的协议後，他可能允许辉达在中国销售下一代晶片的 缩减版。根据协议，辉达和竞争对手 AMD 将把部分先进晶片在中国的销售收入的 15% 上 缴美国政府。 A new Nvidia chip for China might have "30% to 50% off", he suggested in an apparent reference to the chip's computing power, adding that the H20 was " obsolete". 川普暗示，针对中国的新辉达晶片性能可能会「打 7 折到 5 折」，显然是指运算效能， 并补充说 H20 已经「过时」。 U.S. legislators, both Democratic and Republican, have worried that access to even scaled-down versions of flagship AI chips will impede U.S. efforts to maintain its lead in artificial intelligence. 美国国会两党议员均担心，即使是缩减版的旗舰 AI 晶片，也会削弱美国保持 AI 领先地 位的努力。 But Nvidia and others argue that it is important to retain Chinese interest in its chips - which work with Nvidia's software tools - so that developers do not completely switch over to offerings from rivals like Huawei. 但辉达等公司则主张，必须维持中国对其晶片的兴趣，因为这些晶片需搭配辉达软体工具 使用，避免中国开发者完全转向 华为 等竞争对手。 Huawei has made great strides in chip development, with its latest models said to be on par with Nvidia in some aspects like computing power, though analysts say it lags in key areas such as software ecosystem support and memory bandwidth capabilities. 华为在晶片开发上已取得重大进展，其最新型号据称在运算效能等部分方面可与辉达媲美 ，但分析师指出，华为在 软体生态系统支持 和 记忆体频宽能力 等关键领域仍落後。 Complicating Nvidia's efforts to retain market share in China, Chinese state media have also in recent weeks alleged that the U.S firm's chips could pose security risks, and authorities have cautioned Chinese tech firms about purchasing the H20. Nvidia says its chips carry no backdoor risks. 使辉达在中国维持市占率的努力更加复杂的是，近期中国官媒指控美国公司的晶片可能存 在安全风险，并提醒中国科技公司谨慎购买 H20。辉达则强调其晶片不存在「後门」风险 。 Nvidia is also preparing to start delivering a separate new China-specific chip based on its Blackwell architecture and designed primarily for AI inference tasks, according to two other people familiar with those plans. 另有两位知情人士透露，辉达还计划推出另一款基於 Blackwell 架构、专为 AI 推理任 务 设计的中国特供晶片。 Reuters reported in May that this chip, currently dubbed the RTX6000D, will sell for less than the H20, reflecting weaker specifications and simpler manufacturing requirements. 路透社 5 月曾报导，这款暂名 RTX6000D 的晶片，售价将低於 H20，因为规格更弱、制 造要求也更简单。 The chip is designed to fall under thresholds set by the U.S. government. It uses conventional GDDR memory and features memory bandwidth of 1,398 gigabytes per second, just below the 1.4 terabyte threshold established by restrictions introduced in April that led to the initial H20 ban. 这款晶片被设计在美国政府的门槛以下。它采用传统 GDDR 记忆体，记忆体频宽为 1398 GB/s，略低於美国 4 月新规定设定的 1.4 TB/s 限制，而该限制曾导致 H20 最初被禁售 。 Nvidia is set to deliver small batches of RTX6000D to Chinese clients in September, said one of the people. 其中一位消息人士表示，辉达计划於 9 月向中国客户交付少量 RTX6000D。 心得/评论： Nvidia 新晶片 B30A： 基於最新 Blackwell 架构，单晶片设计，性能优於 H20。 预计效能约为旗舰 B300 双晶片的一半。 辉达希望 9 月就能向中国客户送样测试。 美中博弈背景： 美国限制 AI 晶片出口，中国占 Nvidia 收入 13%，市场太大无法放弃。 川普开放可能性，但附带条件：美国政府抽成 15%，且性能需缩水 30–50%。 国会两党担忧中国获取技术会削弱美国 AI 优势。 市场挑战： 华为晶片在部分性能追上 Nvidia，但在生态与记忆体频宽仍落後。 中国官媒开始批评 Nvidia 晶片可能存在「安全风险」，影响市场信任度。 另一款中国特供 RTX6000D： 基於 Blackwell，规格较低，售价比 H20 更便宜。 预设低於美国出口管制门槛（记忆体频宽限制）。 预计 9 月少量交付中国。 --

※ 发信站: 批踢踢实业坊(ptt.cc), 来自: 150.117.197.145 (台湾) ※ 文章网址: https://webptt.com/cn.aspx?n=bbs/Stock/M.1755594999.A.A84.html