Mark Saroufim

Kernels and Chips: Cursor, NVIDIA, and Meta Researchers on GPU Performance | YC Paper Club

Stuart Sul, John, Francois Chaubard, Jon Saad-Falcon, Mark Saroufim, Misha Smelyanskiy, Brennan Shacklett

The event synthesizes critical industry shifts toward specialized ASICs and disaggregated architectures that separate training and inference workloads to maximize intelligence per watt. Technical deep dives highlight breakthroughs in multi-GPU kernel optimization and local inference, demonstrating that consumer accelerators can now deliver nearly 90% of frontier model utility while reducing energy costs by 70%. These innovations collectively address systemic bottlenecks in network communication and latency, establishing a roadmap for distributed, cost-efficient AI ecosystems that minimize reliance on centralized cloud resources.