SK Hynix touts up to 5.15x inference gains with custom HBM
<p class="P1" data-start="401" data-end="695">SK Hynix is pushing computation deeper into high-bandwidth memory (HBM), saying a custom HBM architecture that places compute functions in the base die could improve large language model inference performance by up to 5.15 times as data movement becomes a growing constraint on computing performance.
Read original article ↗
Related Articles
SK Hynix’s $4B HBM Project Targets U.S. Chipmaking Gap
SK Hynix’s $4B Indiana HBM fab aims to plug America’s AI chip packaging hole and build a Purdue-backed supply chain by 2
SK Hynix explores Intel as second source for HBM base dies
SK Hynix is reportedly considering adding Intel Foundry as a second manufacturing source for high-bandwidth memory (HBM)
SK Hynix explores Japanese expansion to meet global AI chip demand
To capture surging global demand for memory chips driven by artificial intelligence (AI), SK Hynix is actively exploring