Samsung unveils LPDDR5X-PIM at Hot Chips 2026, 3.01x faster AI inference
This digest was compiled by AI from multiple sources — links to the originals are below.
Samsung introduced the industry's first LPDDR5X-PIM memory at Hot Chips 2026, adding in-memory logic to the low-power memory standard. The technology delivers 3.01x faster AI inference and 8x the bandwidth compared to standard LPDDR5X. Samsung also revealed plans for LPDDR6X-PIM, with an initial JEDEC specification expected this year.
Key Facts
- Samsung's LPDDR5X-PIM is 3.01x faster than standard LPDDR5X in AI inference tasks.
- The technology provides 8x the bandwidth of standard LPDDR5X.
- Samsung first demonstrated PIM technology in 2021 and introduced HBM-PIM in 2023.
- Samsung expects an initial LPDDR6X-PIM specification from JEDEC this year.
In-Memory Processing Architecture
Samsung's LPDDR5X-PIM integrates a small logic unit alongside DRAM cells, enabling basic calculations directly in memory. Each memory bank has its own PIM unit, an advancement over HBM-PIM where Samsung had to cut banks to fit the logic. The memory bank, scale register file, and source register file feed parallel MAC trees, with output written to a vector register file. Samsung's LPDDR5X operates in two modes: single-bank (traditional DRAM) or multi-bank (PIM), with commands switching between standard read/write and PIM read/write. Samsung uses Address Align Mode (AAM) to handle reordering challenges, mapping DRAM addresses to MAC instructions.
Performance and Market Context
In inference tasks, Samsung says its LPDDR5X-PIM is 2.28x faster than standard LPDDR5X. The headline figure of 3.01x faster AI inference with 8x bandwidth was presented at Hot Chips 2026. Samsung opened its presentation noting that memory makes up the bulk of AI chip costs and its share continues to grow. Micron warned at Hot Chips that HBM wafer demand is worsening, underscoring the cost pressures driving PIM adoption.
7 sources
Samsung unveils LPDDR5X-PIM at Hot Chips 2026, 3.01x faster AI inference
