The video highlights Broadcom’s critical role in designing and packaging future AI GPUs, showcasing a massive multi-die chip called “the beast” that integrates up to 16 compute dies and 16 HBM stacks for unprecedented performance, expected to be deployed by leading AI companies by 2028. It also discusses the challenges of memory technology, Broadcom’s collaboration with foundries, and ongoing advancements toward next-generation chips using cutting-edge process nodes like 2 nanometers.
The video discusses Broadcom’s role as a major contract backend design services company in the semiconductor industry, particularly for AI hardware. Broadcom works with big tech companies like Meta, Google, and Amazon, helping them turn chip designs into actual silicon by managing complex technical details such as timing and latency. The company does not manufacture chips themselves but collaborates closely with foundries like TSMC and Samsung to produce high-performance chips. Broadcom also provides valuable intellectual property (IP) such as high-speed interconnects and memory controllers to support their clients’ chip designs.
The presenter highlights a large chip called “the beast,” shown by Broadcom’s president Charlie Kowas at a conference. This chip represents the future of GPU and AI hardware packaging, featuring multiple compute dies and numerous high-bandwidth memory (HBM) stacks. Broadcom envisions three stages of chip development: current chips with a central compute die and several HBMs, a next-generation chip with multiple compute and IO dies surrounded by more HBMs, and finally, the “beast” with up to 16 compute dies and 16 HBM stacks integrated into a massive package with optical interfaces.
The “beast” chip is expected to be deployed by leading AI model companies such as OpenAI and Anthropic, with Broadcom handling backend design and working with foundries and data center partners for large-scale deployment. The chip’s size is enormous, around 180x210 millimeters, significantly larger than typical chips today. This scale allows for unprecedented compute power and memory bandwidth, essential for training and running advanced AI models. Broadcom anticipates shipping 15 million units of this chip by 2028, marking a significant milestone in AI hardware development.
Broadcom is not the only player in this space; other companies like Marvell, MediaTek, Alchip, Qualcomm, and Intel also offer chip design services. However, Broadcom stands out due to its extensive experience and portfolio of IP, which helps clients optimize chip performance and efficiency. The main challenge moving forward is the availability of high-bandwidth memory or alternative memory technologies that can meet the demands of these large, complex chips. The industry is actively exploring solutions beyond HBM to support future designs.
Finally, Broadcom is already working on next-generation chips using cutting-edge process nodes like 2 nanometers, exemplified by projects such as Fujitsu’s Monaka chip. These developments promise even more powerful and efficient AI hardware in the near future. The video encourages viewers interested in the semiconductor industry and AI hardware to follow these advancements closely, as they will shape the future of computing. The presenter also promotes a merchandise shop for fans of the content.