在即将于明日举行的2024年全球GPU技术大会（GTC）主题演讲中，英伟达预计将揭开其新一代GPU架构——Blackwell的神秘面纱。据行业消息来源XpeaGPU爆料，英伟达的Blackwell系列GPU将采用先进的封装技术，以提升性能和能效。
据透露，首款B100 GPU将基于台积电的CoWoS-L（晶圆基片芯片）2.5D封装技术，该技术通过芯片堆叠实现更高的处理能力。B100将搭载两个计算芯片，与8个8-Hi HBM3e显存堆栈相连，总显存容量达到192GB。这一配置与AMD的Instinct MI300 GPU相媲美，后者同样拥有192GB的显存，采用8个HBM3芯片。
此外，XpeaGPU还暗示了一款更高端的Blackwell GPU——B200。B200将利用12-Hi设计以实现更大的显存容量，达到288GB。尽管具体是HBM3e还是HBM4技术尚未明确，但这一升级无疑将为高性能计算和数据中心应用提供更为强大的支持。
英伟达的黄仁勋CEO预计将在这次GTC大会上详细阐述Blackwell架构的创新点和潜在应用，这无疑将吸引全球科技和计算领域的广泛关注。随着GPU技术的持续发展，英伟达的新一代产品有望在人工智能、机器学习和高性能计算等多个领域树立新的标杆。
英语如下：
News Title: “NVIDIA GTC 2024 Preview: Jensen Huang Expected to Unveil 192GB B100 GPU, Challenging AMD’s Memory Peak”
Keywords: NVIDIA GPU, Blackwell architecture, HBM3e memory
News Content: In the keynote address at the upcoming 2024 Global GPU Technology Conference (GTC) tomorrow, NVIDIA is anticipated to lift the veil on its next-generation GPU architecture, Blackwell. According to insider sources at XpeaGPU, NVIDIA’s Blackwell series GPUs will leverage advanced packaging technology to enhance performance and efficiency.
Leaked information suggests that the initial B100 GPU will be based on TSMC’s CoWoS-L (Chip-on-Wafer-on-Substrate) 2.5D packaging technology, enabling higher processing power through chip stacking. The B100 will feature two compute chips interconnected with eight 8-Hi HBM3e memory stacks, aggregating to a total memory capacity of 192GB. This matches AMD’s Instinct MI300 GPU, which also boasts 192GB of memory using eight HBM3 chips.
Furthermore, XpeaGPU hints at a higher-end Blackwell GPU, the B200, which will leverage a 12-Hi design for an even larger memory capacity of 288GB. While it is unclear whether it will employ HBM3e or HBM4 technology, this upgrade is poised to provide stronger support for high-performance computing and data center applications.
NVIDIA CEO Jensen Huang is expected to delve into the innovations and potential applications of the Blackwell architecture during this GTC event, attracting global attention from the tech and computing sectors. As GPU technology continues to evolve, NVIDIA’s new offerings are likely to set new benchmarks in artificial intelligence, machine learning, and high-performance computing.
【来源】https://www.ithome.com/0/756/310.htm
Views: 1