--- title: "Samsung's LPDDR5X-PIM at Hot Chips 2026: Memory That Computes" slug: samsung-lpddr5x-pim-hot-chips-2026 category: tech category_label: "Tech" author: "BrainWavePost Staff" date: 2026-08-27 tags: ["Samsung", "LPDDR5X-PIM", "processing-in-memory", "Hot Chips 2026", "on-device AI", "DRAM", "AI inference", "memory bandwidth"] read_time_minutes: 9 canonical_url: https://brainwavepost.com/article/samsung-lpddr5x-pim-hot-chips-2026 source: BrainWavePost --- # Samsung's LPDDR5X-PIM at Hot Chips 2026: Memory That Computes *Tech · 2026-08-27 · BrainWavePost Staff · 9 min read* > At Hot Chips 2026, Samsung showed working silicon for LPDDR5X-PIM — low-power DRAM with multiply-accumulate logic built into the memory banks. The company claims 8x internal bandwidth, 2.28x faster model execution, and 3.01x higher token throughput for on-device AI inference. Here is what the silicon actually does, and what remains unproven. ## What Samsung announced At the Hot Chips 2026 symposium in late August, Samsung Electronics presented its LPDDR5X-PIM — a processing-in-memory variant of standard low-power mobile DRAM that embeds computation logic directly inside the memory chip. Unlike many PIM papers over the years, Samsung showed working silicon, not a simulation or a roadmap slide. [1][2] The headline numbers: by keeping multiply-accumulate (MAC) operations inside the memory banks, LPDDR5X-PIM raises effective peak bandwidth from 76.8 GB/s to 614 GB/s in PIM mode — an 8x jump — because data no longer has to cross the narrow, power-hungry bus to the processor for the operations the in-memory logic can handle. [1][3] In Samsung's preliminary on-device AI inference tests, the device ran models 2.28x faster than conventional LPDDR5X and delivered 3.01x higher token throughput — the metric that matters most for LLM-style workloads on phones and edge devices. [3] ## Why PIM matters for on-device AI AI inference is increasingly memory-bound: for many neural network layers, the processor spends more energy and time shuttling weights back and forth from DRAM than it does doing arithmetic. This 'memory wall' is the central bottleneck for running large models on battery-powered devices. [1][4] Processing-in-memory attacks the wall by moving the arithmetic to where the data already lives. If the matrix multiplication can happen inside the DRAM banks, only the compact results need to travel to the host processor — cutting both latency and the energy cost of data movement, which can dominate total inference power. [4] Samsung's choice of LPDDR — the memory used in virtually every smartphone — rather than server-class HBM is strategic: it targets the fastest-growing segment of AI deployment, where power and cost constraints rule out HBM entirely. [2][5] ## The engineering: minimal changes to standard memory A key claim in Samsung's Hot Chips presentation is that LPDDR5X-PIM minimizes changes to the standard memory system. The in-memory logic is confined to the DRAM die, the device retains the standard LPDDR5X interface, and the host processor communicates with it using conventional memory transactions — meaning system-on-chip designers do not need a new bus protocol to adopt it. [1][2] This is the approach Samsung has pursued since its 2021 HBM-PIM and 2022 LPDDR-PIM research disclosures, where the company argued that PIM adoption depends on staying compatible with existing memory controllers and software stacks. The Hot Chips 2026 device is the working-silicon culmination of that design philosophy. [4] ## The accuracy box: what is solid and what is not - Solid: Samsung demonstrated working LPDDR5X-PIM silicon at Hot Chips 2026 — a real device, not a paper design. [1][2] - Solid: the 8x internal bandwidth figure (76.8 GB/s to 614 GB/s) is a direct architectural consequence of in-bank computation, not a benchmark artifact. [1][3] - Preliminary: the 2.28x model-execution and 3.01x token-throughput figures are vendor-reported, on Samsung's own test workloads; independent benchmarking does not exist yet. [3] - Open question: real-world gains depend on software support. PIM benefits materialize only when frameworks and compilers offload the right operations to the in-memory logic, and that ecosystem is early. [4] - Careful with framing: PIM accelerates the memory-bound portions of inference. Layers that are compute-bound see little benefit, so end-to-end gains will vary by model architecture. [4] ## The competitive context Samsung is positioning LPDDR5X-PIM partly as an answer to soaring HBM costs: as AI datacenter demand pushes high-bandwidth memory prices up, a PIM-enabled low-power DRAM offers device makers a cheaper path to faster on-device inference without HBM-class packaging. [5] SK hynix and others have shown their own PIM research devices (SK hynix demonstrated AiM — Accelerator-in-Memory — at earlier conferences), so Samsung's claim to leadership rests on pairing LPDDR-class power with demonstrated working silicon and standard-interface compatibility. [4] ## What to watch next 1. Whether Samsung moves LPDDR5X-PIM from demonstration silicon to mass production, and on what timeline. [2][5] 2. Whether smartphone SoC vendors and mobile AI frameworks (e.g., on-device LLM runtimes) add explicit PIM offload paths. [4] 3. Independent third-party benchmarks once evaluation hardware reaches reviewers. [3] ## Sources and further reading - [1] Tom's Hardware — 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: 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 - [2] ServeTheHome — Samsung LPDDR5X-PIM at Hot Chips 2026 (25 August 2026): https://www.servethehome.com/samsung-lpddr5x-pim-at-hot-chips-2026/ - [3] Hot Chips 2026 — official program (memory technology sessions, 23–25 August 2026): https://hotchips.org/ - [4] Samsung Semiconductor — HBM-PIM / LPDDR-PIM research background (Hot Chips 33, ISSCC 2022 disclosures): https://semiconductor.samsung.com/newsroom/tech-blog/hbm-pim/ - [5] Business Korea — Samsung Unveils Processing DRAM as HBM Alternative (26 August 2026): https://www.businesskorea.co.kr/news/articleView.html?idxno=275535 --- _Canonical article: [https://brainwavepost.com/article/samsung-lpddr5x-pim-hot-chips-2026](https://brainwavepost.com/article/samsung-lpddr5x-pim-hot-chips-2026) — © BrainWavePost. Educational content; see the article page for full disclaimers._