News

SK Hynix proposes custom HBM with 'StreamDQ' to offload AI inference tasks from GPUs, using TSMC foundry for base die

Wednesday, January 14, 2026 at 05:02 AM

SK Hynix is developing custom HBM technology, including 'StreamDQ,' which offloads dequantization tasks from GPUs to HBM. This aims to reduce AI inference bottlenecks and improve LLM processing speed by up to 7x. The company is proposing this technology to major customers like NVIDIA and is utilizing TSMC's advanced foundry processes for the HBM base die. This technology is expected to be integrated into HBM4E and future HBM generations, allowing customers to customize the HBM base die with desired functions.

Related Companies

TSMC
TSMC
2330
TW
Nvidia
Nvidia
NVDA
US
SK Hynix
SK Hynix
000660
KR