NVIDIA: chiplets traded single-threaded speed for core counts, agentic AI forces reversal

NVIDIA's Vera team told Wccftech that x86's shift to chiplets for higher core counts hurt single-threaded performance and IPC. Agentic AI needs a fast CPU-GPU loop, so Vera uses a monolithic compute die, custom LPDDR5X memory and second-gen Scalable Coherent Fabric.

• Vera combines high single-threaded performance with high throughput, NVIDIA says
• A monolithic compute die removes latency from inter-core communication
• Custom LPDDR5X memory subsystem delivers high bandwidth
• Agents must finish sequential CPU work fast to keep expensive GPUs utilized