Jeson and I interview Tapa Ghosh, an engineer, Thiel Fellow, and the founder of Volantis Semi, which raised $5M from Open AI’s Sam Altman to build photonic motherboards for AI chips. Tapa's journey began with crashed drones, leading him to a hardware deep dive and giving a talk Stanford on designing 3D ICs for deep learning as a teenager.
We explore Volantis's approach to solving a key bottleneck in AI: latency and bandwidth in communication between chips. Tapa explains why using light instead of electricity could revolutionize AI infrastructure, and offers insights into the current landscape dominated by NVIDIA.
His perspectives on self-learning, team building, and resurfacing old ideas to solve new problems reveal what it’s like to work on the foundational layer to our AI future. Please enjoy this wide-ranging conversation with Tapa Ghosh.
Timestamps:
(00:00:00) - Tapa’s origin story: How a crashed drone sparked an interest in chip design
(00:05:45) - How to speedrun to the frontiers of knowledge
(00:11:28) - Giving a talk at Stanford: Petascale Deep Learning on a Single Chip
(00:13:27) - Differences between CPUs, GPUs, and ASICs
(00:15:48) - How NVIDIA went from graphics to AI computing
(00:17:37) - Why AI workloads today are communication-bound, not compute-bound
(00:19:32) - Study the history of your field
(00:21:48) - Why Volantis is building a photonic motherboard
(00:25:48) - What he’s doing differently this time around vs. his last startup
(00:32:21) - The landscape of AI chips: NVIDIA, AMD, challengers
(00:38:10) - New problems as we continue to scale AI clusters for training & inference
(00:39:51) - What Moore’s Law teaches us about human psychology & coordination
(00:42:08) - The engineering marvel of semiconductor fabrication
(00:43:56) - Getting ghosted by Intel vs. gifted pineapple cakes by TSMC
(00:48:46) - Other promising opportunities in AI hardware / infra
(00:52:44) - One of Tapa's unresolved rabbit holes
