Semiconductor News · کورتەی دابینکردن

Near-memory computing moves toward AI inference workloads

LPDDR6-PIM and custom HBM approaches target lower data movement, but software, cost, and thermal barriers remain before broad deployment.

بڵاوکراوەتەوە نوێکراوەتەوە 3 خولەک خوێندنەوە ✓ ئامادەی AI ھاوبەشکردن

بە کورتی — سێ ڕستە

  • LPDDR6-PIM and custom HBM place computation closer to memory to reduce data movement and improve AI inference performance. (S1)
  • Low-batch decode is more suitable for near-memory approaches than large-batch decode or prefill. (S1)
  • Samsung has published architecture and performance benchmarks for an LPDDR6-PIM chip targeting mobile on-device AI. (S1)

وەرگیراوە لە key_takeaways (خانەی ڕاستەقینە، زۆرترین 3 تۆمار). ڕاستی و شیکاری بە جیا نیشانە کراون.

Near-memory computing moves toward AI inference workloads — Semiconductor News news briefEditorial cover for a Semiconductor News brief. Memory motif; headline "Near-memory computing moves toward AI inference workloads". No reported key numbers on this brief, so no number strip is drawn. 1 cited source. Static cover (no animation).SEMICONDUCTOR NEWS · NEWS BRIEFNear-memory computing moves toward AIinference workloadsXAIPART
Cover image generated from data already shown on this page: the category motif, the headline, and reported key numbers. No decorative illustration; no third-party image.

لە وەڵامێکی AIەوە هاتوویت — ئەمە ئەسڵی پشتڕاستکراوەیە

لەوانەیە پوختەیەکی AI بنەمای داواکارییەکانی لەدەستدابێت. پێشبینییە ڕەسەنەکە، بەشەکانی تووشبوو و هەموو ئاماژەکان لەم پەڕەیەدان، هەریەکەیان بەرەو سەرچاوەکەی دەگەڕێتەوە.

پێشبینی XAIPART

ئەم کورتەیە پێشبینییەکی هەڵوەشێنراوەی تێدایە.

بەشەکانی تووشبوو

ھیچ گونجانێک لە کاتالۆگدا بۆ ئەم چیرۆکە نییە — ھیچ پیشان نادرێت لە جیاتی دەرھێنانی گونجان.

Reported by Semiconductor Engineering on . Summarized and analyzed by XAIPART Editorial.

What happened

Edge AI inference is encouraging chip and system designers to place more computation near memory, reducing transfers between processors and data stores while addressing tight power limits.S1 Candidate approaches include processing near memory, called PNM, processing in memory, called PIM, and compute-in-memory architectures that perform calculations within a memory array.S1 Processing in the same package through custom HBM or on the same die through LPDDR6-PIM can shorten data paths and improve performance.S1

The article distinguishes hardware compute-in-memory from the separate enterprise-software meaning of IMC, which runs databases and analytics from RAM rather than disk.S1 In hardware, analog multiply-accumulate operations can execute on SRAM or ReRAM bitlines, while digital PIM places logic beside DRAM banks.S1 PNM instead places logic on a separate die within the memory package, including proposed custom HBM4 and HBM4E base dies.S1 Expedera cautioned that near-memory designs work well for low-batch decode, but large-batch decode and prefill make activation communication more difficult.S1

Adoption remains constrained by software, manufacturing, cost, and thermal issues.S1 Compilers, operating systems, memory controllers, and processors need protocols and scheduling support for offloaded operations, while combining logic and DRAM processes can increase manufacturing expense.S1 Samsung has published the architecture and performance benchmarks for an LPDDR6-PIM chip targeting low-power mobile on-device AI, and JEDEC work on the standard is nearing completion.S1 Thermal design is also critical because stacking memory near logic concentrates heat and can create thermo-mechanical stress.S1

MetricReported valueSource
Expected in-memory application horizonThree to five years, maybe around 2030S1
Example PIM logic placementEvery 10 columns or so of memoryS1
Historical period covered by programming modelsOver the last 40 yearsS1

Background timeline

  1. - Semiconductor Engineering published its analysis of memory architectures for AI inference.S1
  2. around 2030 - Ayar Labs projected a possible application horizon for in-memory systems, while describing the technology as not yet ready for broad commercialization.S1

XAIPART take

Technical impact

XAIPART analysis: Design teams should evaluate near-memory compute stage by stage rather than treating it as a universal accelerator replacement. Low-batch decode appears better suited to these architectures than prefill or large-batch decoding, where activation traffic becomes harder to control. XAIPART analysis

Market impact

XAIPART analysis: LPDDR6-PIM offers an earlier path for power-sensitive mobile and fixed edge systems than more flexible custom HBM base dies. However, mixed logic and DRAM manufacturing can raise unit costs, while thermal constraints may limit how much compute can be placed beneath memory. XAIPART analysis

Action

XAIPART analysis: Buyers should request workload-specific benchmarks covering decode, prefill, batch size, energy per operation, thermal performance, and compiler support. Engineering teams should compare conventional DRAM against LPDDR6-PIM and custom HBM options, including the operational work required for scheduling and system integration. XAIPART analysis

XAIPART Forecast

XAIPART analysis XAIPART expects that within 12 months, at least 2 of Samsung, SK hynix, and Micron will disclose production-availability or volume-shipment evidence for LPDDR6-PIM, custom HBM4/HBM4E, or HBM-based near-memory products; if fewer than 2 provide such evidence, this forecast is wrong. The evidence should specify component availability, customer applications, or shipment volume rather than development samples alone.

Facts are as reported by the cited source. The design-in, sourcing and forward-looking statements above are XAIPART's analysis, not the source's claims. The article relies on a single source and expert commentary, and it does not disclose production pricing, lead times, qualification schedules, or comparative benchmark results for the named PIM and PNM designs. XAIPART analysis

Corrections: none at publication. This brief will be updated with a timestamp if the source corrects its report.

جێگرەوەکان و چرخەی ژیان

جێگرەوە تۆمارکراوەکان — مەودای کاتالۆگ، نەک ڕیزبەندی بازاڕ
بەشپەیوەندیخولگەی ژیانبنەما
— ھیچ بەدیلێک لە تۆماردا نییە پێویستی بە داتا ھەیە — پێویستی پێیەتی match_type پێش ئەوەی هیچ نیشانەیەکی پێشوو دەربکەوێت

نیشانەکانی «پێشوو» و «بەدیل» تەنھا لە جۆرێکی گونجانی تۆمارکراو، دۆخی چرخەی ژیان، یان دەقی سەرچاوەوە دێن. لەو شوێنەی ھیچ گونجانێک تۆمار نەکراوە، ھیچ نیشانەیەک پیشان نادرێت. لیستەکانی خۆمان ئاماژەیەکی بازاڕ نین.

چی چاودێری بکەیت

پێشبینی ڕەتکراوە — سنوور، ماوە و دۆخی ئێستا. پێشبینییە ڕەتکراوەکان بە دیار دەمێننەوە.
داواکاریسنوورلەلایەندۆخ
XAIPART expects that within 12 months, at least 2 of Samsung, SK hynix, and Micron will disclose production-availability or volume-shipment evidence for LPDDR6-PIM, custom HBM4/HBM4E, or HBM-based near-memory products fewer than 2 provide such evidence within 12 months Pending

ئەو پێشبینییانەی هەڵە دەردەچن بە دیار دەمێننەوە و نیشان دەدرێن. ئێمە لەناویان نابەین.

دیمەنی کێبڕکێ

  • 27949 لیستەی زیندوو لەم پۆلەدا لەسەر XAIPART
  • 113 بەرهەمهێنەرانی نوێنەرایەتیکراو
  • 48 بەرهەمهێنەرانی ئاماژەپێکراو لە ئەرشیفی هەواڵەکانمان لە ٣٠ ڕۆژی ڕابردوودا

ژمارەکان تەنها کاتالۆگی XAIPART نیشان دەدەن — پشکی بازاڕ نین.

بەشەکانی تووشبوو

ھیچ گونجانێک لە کاتالۆگدا نییە. ئەم هەواڵە بۆ هیچ پارچەیەک کە ئێستا لە XAIPART لیستکراوە ناگونجێت. ئێمە هاوتایی نادۆزینەوە، و لیستێکی شوێنگرەوە نیشان نادەین.

ئەمە لە کوێی چیرۆکەکەدایە

Event chain · as of 2026-11-12T20:29:42+08:00 · nodes link to published articles on this site.
Event chain — samsung+semiconductor-newsChain nodes: 2026-10-07 Samsung Foundry is reportedly set to fabricate Anthropic's custom AI accelerator on its 2 nm SF2 node, with Broadcom handling design. The deal could shift AI ASIC sourcing away from TSMC.; 2026-10-07 Samsung Electro-Mechanics signs a 290 bln-won MLCC supply deal for AI servers running Jan-Dec 2027, lifting its 2026 long-term contract tally to 3.9 trillion won.; 2026-10-08 Samsung's HBM4E reportedly passed NVIDIA Rubin Ultra testing at up to 16 Gbps/pin and 4 TB/s per stack, signaling a rebound after the HBM3E rejection.; 2026-10-11 Korean industry projections name Samsung and SK hynix as prime beneficiaries of a 20x DRAM demand surge from humanoid robots by 2030. Engineering and sourcing implications inside.; 2026-11-05 Samsung's reported Taylor Fab 2 allocation could tighten advanced HBM base-die capacity, although volumes, pricing and lead times remain undisclosed.; 2026-11-12 LPDDR6-PIM and custom HBM approaches target lower data movement, but software, cost, and thermal barriers remain before broad deployment.. As of 2026-11-12T20:29:42+08:00. Open forecast: Samsung Foundry is reportedly set to fabricate Anthropic's custom AI acc — pending Open forecast: Samsung Electro-Mechanics signs a 290 bln-won MLCC supply deal for AI se — pending.2026-10-07Samsung Foundry isour brief ↗2026-10-07Samsung Electro-Meour brief ↗2026-10-08Samsung's HBM4E reour brief ↗2026-10-11Korean industry prour brief ↗2026-11-05Samsung's reportedour brief ↗2026-11-12LPDDR6-PIM and cusour brief ↗this briefOpen forecast: Samsung Foundry is reportedly set to fabricate Anthropic's custom AI acc — pendingOpen forecast: Samsung Electro-Mechanics signs a 290 bln-won MLCC supply deal for AI se — pending
  1. samsung-foundry-eyes-anthropic-ai-asic-production-on-sf2-2-nm-node2026-10-07
  2. samsung-electro-mechanics-books-216m-ai-server-mlcc-deal-what-it-means-for-2027-server-boms2026-10-07
  3. samsung-hbm4e-clears-nvidia-rubin-ultra-validation-what-it-means-for-ai-accelerator-sourcing2026-10-08
  4. humanoid-robot-dram-demand-seen-rising-20-fold-by-2030-what-it-means-for-samsung-and-sk-hynix-boms2026-10-11
  5. what-samsung-s-taylor-fab-2-booking-means-for-2-nm-hbm-sourcing2026-11-05 · shortage
  6. near-memory-computing-moves-toward-ai-inference-workloads2026-11-12
  1. Samsung Foundry is reportedly set to fabricate Anthropic's custom AI accelerator on its 2 nm SF2 node, with Broadcom handling design. The deal could shift AI ASIC sourcing away from TSMC. trade
  2. Samsung Electro-Mechanics signs a 290 bln-won MLCC supply deal for AI servers running Jan-Dec 2027, lifting its 2026 long-term contract tally to 3.9 trillion won.
  3. Samsung's HBM4E reportedly passed NVIDIA Rubin Ultra testing at up to 16 Gbps/pin and 4 TB/s per stack, signaling a rebound after the HBM3E rejection. trade
  4. Korean industry projections name Samsung and SK hynix as prime beneficiaries of a 20x DRAM demand surge from humanoid robots by 2030. Engineering and sourcing implications inside.
  5. Samsung's reported Taylor Fab 2 allocation could tighten advanced HBM base-die capacity, although volumes, pricing and lead times remain undisclosed. shortage trade
  6. LPDDR6-PIM and custom HBM approaches target lower data movement, but software, cost, and thermal barriers remain before broad deployment. trade

زنجیرەکە تەنها لە کورتە بڵاوکراوەکان دروست دەکرێت — گرێیەک کە کورتەی نییە کێشراوە نییە.

فشاری تێچوو بەپێی ڕەهەند

Source: FRED · Generated · reference output z and composite are algorithm outputs, not a fetch snapshot

Cost pressure by dimension — composite stateDiverging bars for three dimensions around a zero midpoint. Input Cost has no reading; Supply Demand z 0.0276, neutral; Fx Energy z -0.0087, neutral. Composite index 0.1, band neutral.0Input costno readingSupply/demandz +0.0276 · neutral48 dFX & energyz −0.0087 · neutral5 dcomposite +0.1 · neutral−100 · −40 · −15 · 0 · +15 · +40 · +100
Underlying values — copyable for verification.
DimensionFRED series IDUnitLagAs of
Input costPCOPPUSDM, PALUMUSDM, PNICKUSDMUSD/mt52 d
Supply/demandPCU334413334413, IPG3344Sindex(1982=100) / index(2017=100)48 d
FX & energyDEXCHUS, DCOILBRENTEU, DHHNGSPrate / USD/bbl / USD/MMBtu5 d

No free authoritative interface for: gold · silver · rare earth · polysilicon · lithium carbonate · market share — listed rather than approximated.

ئەمە چۆن بەرهەم هێنرا

بەڵگەی پرۆسە — ھێنان، ڕاستی سەرچاوە، مۆدێل و دەروازە
شوێنپێی پشکنین بۆ ئەم کورتەنامەیە
Source fetched Semiconductor Engineering · semiengineering.com · 34,402 chars · fetch status success
Source sanity (S1) ok — passed before any model call (no s1_ block recorded)
Model / prompt MiniMax-M3.1-Flash · prompt Blog Agent - News Summary (v2.8-final) · 15,266 tokens · 22.3s · run completed
Gate verdict pass · 6/6 applicable checks passed · spec v2.9 · C10 ratio 0.002226 · C10 rounds 0
Time contract 2 <time datetime> values in this brief — 0 violations (+08:00 or date-only)

هەر ڕیزێکی سەرەوە لەلایەن پرۆسێسەکەوە بەرهەم دەهێنرێت، نەک بە دەست نووسرابێت.

نمرەی سەروەرێتی

ئەنجامی دەروازەی دیاریکراو — PHP ی خاوێن، هیچ مۆدێلێک لە بڕیاردانەکەدا بەشدار نییە
شیکردنەوەی نمرەی دەروازە 6 پشکنینەکانی جێبەجێکراو، 6دەرپەڕیوە: 100 لەسەد. No hold items or mandatory failures. Gate spec: v2.9 (S0 + S1 + C1–C20) · Source affinity (C20): 0.7746. پشکنینەکانی جێبەجێکراو تێپەڕێنراون 6 / 6 = 100% · بڕیار: pass No hold items or mandatory failures Gate spec: v2.9 (S0 + S1 + C1–C20) · Source affinity (C20): 0.7746 دەروازەکە بڕیار لە بڵاوکردنەوە دەدات. نمرەی وردبینی سەربەخۆ تەنها ڕاوێژییە.
هەڵەیەکت دۆزیوەتەوە؟ پێمان بڵێ چی هەڵەیە — ڕاستکردنەوەکان لەگەڵ دەقی ڕەسەن بڵاودەکرێنەوە و تۆمار دەکرێن. بە نهێنی وتارەکان دەستکاری ناکەین.

بۆچوونەکان

یەکەم کەس بە بۆچوون بنووسە.

سەرچاوەکان

  1. Rethinking Memory For AI Inference — https://semiengineering.com/rethinking-memory-for-ai-inference/ · rss · سەردانکراوە

ڕاستییەکان بە S1/S2 نیشان دەدرێن و بۆ سەرچاوەی بەستەردراو دەگەڕێنەوە. ئەو دەقانەی نیشانکراون شیکاری بیرکردنەوەی خۆی XAIPARTن، جیاکراوە لە ڕاستییەکانی ڕاپۆرتکراو. بەستەرەکانی سەرچاوە شوێن دەکرێن نەک nofollow — دانانی سەرچاوە بە مەبەستە.