The distinction that matters: PIM vs CIM
The memory-IDM PIM landscape (verified June 2026)
- Sk Hynix — AiM (Accelerator-in-Memory). GDDR6-AiM (claimed up to 16× AI speedup, ~80% lower power for LLMs), the AiMX accelerator card, plus LPDDR6-PIM and CXL-CMM. Positioning itself as “full-stack AI memory creator.” The most product-shaped IDM PIM effort.
- Samsung Electronics — HBM-PIM (Aquabolt-XL). The PIM research leader: HBM2-PIM (32 processing units between bank pairs, ~2.5× system gain, >60% energy cut on a Xilinx accelerator), LPDDR5-PIM, and AXDIMM (DDR5 acceleration DIMM). PIM roadmap spans HBM3 / DDR5 / GDDR6 / LPDDR5.
- Micron Technology — lighter on public PIM. Focused on winning the HBM4 socket; explores 4F2 / CBA / 3D DRAM but has not pushed a branded PIM product the way the Korean two have.
- Upmem — the commercial DRAM-PIM startup (French): general-purpose processors placed in standard DRAM DIMMs. The one independent company that actually ships DRAM-PIM hardware.
- Neuroblade (Israel): “computational memory” SPU, pivoted toward data-analytics acceleration.
The honest caveat
PIM has been “almost productised” at every Hot Chips since 2021. The binding question for the whole category — IDM PIM and startup CIM alike — is not “does it work in a demo” (it does) but “does a host system design it in at volume.” No one has cleared that bar yet.