Zoom Link: https://upenn.zoom.us/j/99074346805
Meeting ID: 990 7434 6805
Passcode: 519615

The impact of artificial intelligence on society is undeniable, and the race to train and deploy ever larger, more capable models shows no signs of slowing. The explosive growth of AI has been driven by a virtuous cycle of progress in algorithms, data, and hardware. Yet at current trajectories, the energy required to train and serve future frontier models would reach unsustainable planetary scales within decades. In other words, we need exponential improvements across all three pillars to meet the growing and evolving demands of AI.

We stand at the cusp of a paradigm shift in computing. AI workloads are fundamentally different from the human-written software of the past. Traditional computing focused on maximizing single-thread performance of general-purpose CPUs, leading to decades of remarkable innovations in logic design and the rise of the fabless-foundry ecosystem. Over the same period, memory technologies like DRAM and Flash evolved along a separate axis — one focused almost entirely on capacity and cost, cementing the commodity-memory model. AI upends this model of computing. Its success depends less on raw compute and more on moving and storing vast amounts of data efficiently. The meteoric rise of GPUs, TPUs, NPUs, and, more generally, XPUs, illustrates that AI computing requires specialization, parallelism, and bandwidth. High-Bandwidth Memory (HBM) has provided a temporary boost, but it is a short-term fix to a long-term architectural problem: the widening gap between logic and memory.

This talk provides an overview of three recent papers that explore designing computing systems focused on the memory-centered needs of today’s AI models.

Sustaining AI’s progress will require a memory-driven computing revolution — from new device technologies and architectures to new design ecosystems and even new business models. Just as the fabless-foundry model unleashed innovation in logic, AI now needs a similar platform for memory. Such a platform will enable energy-efficient AI systems, custom specialized memory architectures, and a new era of sustainable computing innovation.