OpenAI Reveals Jalapeño, Its First Custom Inference Chip, With Up to 4.1x Performance Gains

OpenAI has shared first benchmark results for Jalapeño, its first in-house inference chip, claiming up to 1.9x more throughput per watt and up to 4.1x higher performance on interactive workloads versus commercial systems.