{
  "id": 3325652,
  "title": "Hot Chips 2026: Samsung makes LPDDR5X smart with logic unit in memory — LPDDR5X-PIM is 3.01x faster than LPDDR5X in AI inference with 8x the bandwidth",
  "url": "https://urgent.news/2026/08/25/hot-chips-2026-samsung-makes-lpddr5x-smart-with-logic-unit-in-memory",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-08-25T18:31:37.000Z",
  "source": {
    "name": "Tom's Hardware",
    "slug": "tom-s-hardware",
    "url": "https://www.tomshardware.com/pc-components/dram/hot-chips-2026-samsung-makes-lpddr5x-smart-with-logic-unit-in-memory-lpddr5x-pim-is-3-01x-faster-than-lpddr5x-in-ai-inference-with-8x-the-bandwidth"
  },
  "original_language": "en",
  "account": "On May 13, Samsung unveiled LPDDR5X-PIM, the first low-power memory technology to incorporate in-memory logic. At the Hot Chips 2026 conference, the company showcased how this innovation significantly enhances AI inference performance. By integrating PIM, or Processing-in-Memory, alongside LPDDR5X memory, Samsung aims to eliminate the processor as a bottleneck and boost computation speed, particularly for AI inference tasks.\n\nThe primary motivation behind adding in-memory processing to LPDDR5X lies in the escalating costs and power demands associated with HBM (High Bandwidth Memory). Memory constitutes the largest portion of AI chip expenses, and its share is expected to grow. LPDDR5X, with its lower power consumption compared to HBM and DDR5, appears to be an ideal candidate for PIM implementation.\n\nDuring the Hot Chips 2026 presentation, Samsung introduced a real product demonstrating LPDDR5X-PIM and validated its concept introduced earlier. The company showcased a layout of how PIM is integrated into LPDDR5X, with each memory bank featuring its own PIM alongside scale and source register files feeding parallel MAC (Multiply-Accumulate) trees. Once calculations are completed, the output is written to a vector register file.\n\nLPDDR5X-PIM operates in two modes: single-bank, emulating traditional DRAM, and multi-bank, leveraging PIM functionality. To overcome the reordering issue in conventional DRAM, Samsung employs Address Align Mode (AAM) to map DRAM addresses to MAC instructions, assigning the Vector Register File (VRF) and Source Register File (SRF) addresses based on the Row Address (RA) and Column Address (CA) respectively.\n\nTo illustrate the memory operation, Samsung provided an example of a MAC operation, assuming weight parameters are already stored in the memory cell. In multi-bank (PIM) mode, 512 bytes of FP8 activation data are broken down into 16, 256-bit packets written into each bank in the Source Register File. A PIMX_RD reads the weight data from the DRAM bank and feeds it into the MAC trees alongside data from the SRF. After the calculation, the output vector from each operation is written into the Vector Register File.\n\nUpon completing the calculation, a PIMX_WR command transfers the output data back to the DRAM bank. Samsung noted that there doesn't have to be a one-to-one relationship between reads and writes, but this example demonstrates a practical approach. With a Vector Register File (VRF) size of 1 kbit, a maximum of four calculations can be stored.",
  "summary": "Samsung detailed the industry's first LPDDR5X-PIM at Hot Chips 2026, adding logic directly to memory to speed up data-intensive workloads like AI inference.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Tom's Hardware",
        "title": "Hot Chips 2026: Micron warns HBM wafer penalty is widening with every generation — AI memory uses 3x more silicon than DDR5, company says memory wall is 'getting worse' as prices rise",
        "url": "https://urgent.news/2026/08/25/hot-chips-2026-micron-warns-hbm-wafer-penalty-is-widening-with-every",
        "published": "2026-08-25T12:19:38.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}