Week endingFriday, July 10, 2026Covers July 4 to July 10, 2026 · 8 items
Money and capacity led the week. Samsung gave record preliminary second quarter results, SK hynix began trading in the United States, and Micron said it is raising its planned US investment, with a New York construction milestone and a wafer supply deal with GlobalWafers.
On the product side Samsung said its PCIe 6.0 enterprise SSD, the PM1763, is in mass production. Three preprints tied memory design to AI serving, covering dequantization inside HBM, high bandwidth flash with a tile friendly layout, and SSD backed memory tables for language models.
Samsung says the PM1763 is in mass production, a PCIe 6.0 enterprise drive aimed at AI and HPC servers. It adds that the drive is built for secure, isolated workloads. Published speeds are Samsung's own figures.
Samsung ·
Samsung's preliminary second quarter figures show its largest quarterly operating profit, which it ties to AI server memory demand and higher DRAM prices. The full report with division detail is due at the end of the month.
Samsung ·
Micron says it is expanding its US investment plans through 2035 because of AI driven memory demand. It also marked the first concrete pour at its Clay, New York site, which moves from early site work to above ground construction.
Micron ·
Micron announced strategic financing for a GlobalWafers wafer plant in Sherman, Texas, paired with a long term agreement for access to raw silicon wafer capacity. The deal still needs definitive agreements and regulatory approvals.
SK hynix says its ADRs started trading on Nasdaq, a move meant to widen its US investor base as the largest supplier of high bandwidth memory. The new shares are to be listed in Korea later in July.
Large access granularity in high bandwidth flash makes tiled matrix work fetch far more data than it needs. TileLens changes the layout so requests line up with compute tiles. Results come from a GPU simulator, and the paper is accepted to HPCA 2027.
arXiv ·
TF-Engram builds phrase memory offline and keeps large tables on SSD, using prefetching to hide the slower tier during decoding. The authors say it cuts GPU memory demand and improves downstream scores on a small model.
arXiv ·
StreamDQ puts small dequantization blocks in the base die of high bandwidth memory so quantized weights are expanded on the memory side. The gains are from the authors' evaluation, not shipping hardware.
arXiv ·