英特尔将为 Nvidia 封装部分 AI-GPU

据台湾《联合报》(United Daily News),Nvidia 可能会将其部分数据中心 GPU 的封装工作外包给英特尔。目前,台积电(TSMC)通过其 CoWoS 封装技术完全承担了这项任务,但这家台湾芯片制造商无法满足所有需求。Nvidia 并不会完全转向英特尔,而只会将其生产的一小部分产品外包给这家美国芯片制造商。这将从今年第二季度开始实施。台积电仍将是 Nvidia 的主要供应商。英特尔将为 Nvidia 每个月加工约 5000 个晶圆,这将满足 Nvidia 总需求的约 10%。这可能意味着每月生产 30 万个 Nvidia H100 数据中心 GPU。

消息来源称,Intel Foundry Services 将不会自己生产 GPU,而只会使用先进的封装技术进行封装。目前,Nvidia 的许多用于人工智能的数据中心 GPU(包括高端的 A100、H100 和 H200)都是使用台积电的 CoWoS 技术进行封装的。CoWoS 是一种 2.5D 封装技术,可以将不同芯片连接在硅中间层上。这样,芯片可以更加紧密地排列在一起。包括 Nvidia 在内的高性能计算芯片制造商使用 CoWoS 技术将其 GPU 芯片与高速 HBM 内存连接起来。英特尔提供了基于中间层的 Foveros 封装技术,尽管该技术与台积电的技术不完全相同。Nvidia 可能需要调整和验证其设计,然后才能利用英特尔的封装能力进行大规模生产。

台积电此前承认,目前无法满足对先进封装的需求,即将独立芯片芯片组组合成一个可工作整体的需求。这是由于对用于训练 AI 模型的数据中心 GPU 的需求很高。该公司去年已表示,将加快提升其封装能力的努力。

如今,英特尔的 Foundry Services 部门还为其他公司提供芯片制造和封装服务,而不仅仅是为自家产品提供生产能力。该美国芯片制造商最近在新墨西哥州开设了一家新的封装工厂。这是英特尔首个大规模使用 Foveros 技术进行芯片封装的工厂。该公司还在俄勒冈州拥有封装能力,并正在马来西亚建设一家先进的封装工厂。


According to Taiwan’s United Daily News, Nvidia may outsource the packaging work of some of its data center GPUs to Intel. Currently, Taiwan Semiconductor Manufacturing Company (TSMC) handles this task entirely through its CoWoS packaging technology, but the Taiwanese chip manufacturer is unable to meet all the demands. Nvidia will not completely shift to Intel but will only outsource a small portion of its production to the American chip manufacturer. This is expected to be implemented starting from the second quarter of this year. TSMC will remain Nvidia’s main supplier. According to the report, Intel will process approximately 5,000 wafers per month for Nvidia, fulfilling about 10% of Nvidia’s total demand. This could potentially mean the production of 300,000 Nvidia H100 data center GPUs per month.

Sources suggest that Intel Foundry Services will not produce GPUs themselves but will utilize advanced packaging technology for the task. Currently, many of Nvidia’s data center GPUs used for artificial intelligence, including the high-end A100, H100, and H200, are packaged using TSMC’s CoWoS technology. CoWoS is a 2.5D packaging technology that allows different chips to be connected on a silicon interposer, enabling tighter integration of the chips. High-performance computing chip manufacturers, including Nvidia, use CoWoS technology to connect their GPU chips with high-speed HBM memory. Intel offers Foveros packaging technology based on interposers, although this technology is not identical to TSMC’s. Nvidia may need to adjust and validate their designs before utilizing Intel’s packaging capabilities for large-scale production.

Previously, TSMC acknowledged that it is currently unable to meet the demand for advanced packaging, which involves assembling individual chip dies into a functional whole. This is due to the high demand for data center GPUs used for training AI models. The company expressed its commitment to accelerating its efforts to enhance its packaging capabilities last year.

Intel’s Foundry Services division currently provides chip manufacturing and packaging services not only for their own products but also for other companies. The American chip manufacturer recently opened a new packaging facility in New Mexico, which is their first large-scale factory utilizing Foveros technology for chip packaging. The company also possesses packaging capabilities in Oregon and is constructing an advanced packaging facility in Malaysia.

英特尔将为Nvidia封装部分AI-GPU
(0)
Harley Norris的头像Harley Norris

发表回复

您的电子邮箱地址不会被公开。 必填项已用 * 标注