Custom Silicon Beyond GPUs—Scaling Intelligence into the AGI Age

The race toward Artificial General Intelligence (AGI) is accelerating, powered by ever-larger and more capable LLMs. But as these models scale, we are hitting the physical limits of general-purpose GPUs. Memory access is now slower and more energy-hungry than computation itself. Power demands are straining datacenter infrastructure. And as models span more chips, interconnect delays—not compute—become the bottleneck.
This talk explores how custom silicon—such as AWS Trainium, Google TPU, Cerebras WSE, and Groq LPU—is designed to overcome these barriers.
By keeping memory closer to compute, optimizing data movement, and delivering higher efficiency per watt, domain-specific accelerators are reshaping the hardware stack around the needs of LLMs.
Attendees will gain a clear framework for understanding why this shift matters, how it breaks past the Memory, Power, and Interconnect walls, and what it means for building a more scalable, cost-effective, and energy-efficient foundation on the road to AGI.
Speaker(s): Asheesh Goja,
Santa Clara University, Santa Clara, California, United States, Virtual: https://events.vtools.ieee.org/m/498701