ChangXin Memory Technologies has begun handing out HBM3E samples to Chinese AI chip designers and intends to start limited production within two months, according to Reuters, which reported that only one to three usable parts come off each wafer.
Sources put the yield on the stacked memory at 20 to 30 percent. SK hynix, the market leader, runs near 70 percent on comparable parts. Testing has already told the difference: engineers quoted by Reuters rate the Chinese part closer to SK hynix’s previous-generation HBM3 than to its HBM3E.
Two versions have gone to customers, a standard build and a revised one aimed at lifting yield. High-volume memory is where the economics bite. Parts that fail test are scrapped or sent back for rework, so a low yield turns every good stack into a costly one.
The push is not hard to explain. China cannot buy enough leading-edge HBM, and Washington’s December 2024 controls cut off direct supply. Huawei, Cambricon, MetaX and Iluvatar CoreX have already raised prices on existing and next-generation accelerators by 20 to 50 percent as memory costs climbed.
CXMT began developing the technology in 2024. Samsung, which widened HBM4 sales in July and has shipped HBM4E samples, remains a generation ahead. Counterpoint Research puts SK hynix at half of second-quarter HBM revenue, Samsung at 33 percent and Micron at 18 percent.
