An OpenAI-designed inference accelerator announced with engineering samples in lab use. Broadcom provides implementation and networking; Celestica contributes system integration.
Official event/announcement hero image, not a verified die or package photograph. No image bytes are stored or displayed. Photograph rights and any annotation, crop, measurement or written analysis must be resolved separately.
The announcement describes reducing data movement and balancing compute, memory and networking for language-model inference. It does not disclose the on-chip fabric, memory capacity, process, die area or physical floorplan.
CURATOR INTERPRETATIONThe architectural question is how weights and intermediate state reach compute resources at useful utilization. The published system goal supports this question; it supplies neither a wiring diagram nor reproducible bandwidth measurements. Performance claims remain vendor claims pending a detailed report.
VLSI Korea