PIM AI Accelerator Patents: Who Leads, Where the Gaps Are 2026
- Flat, not falling. filings held at 8 in both 2021 and 2024 — a 0% span, not a decline, once the 2025-26 publication lag is discounted.
- A concentrated leader, then a real gap. the top-ranked assignee holds 25 records against 13 at fifth place and 2 at tenth, across 12 ranked companies total.
- Compute and memory claims dominate together. 66.3% of the 95 records touch G06F digital processing and 42.1% touch G06N AI-model computing, with G11C memory classes present in 29.5%.
Filing growth compares 2021 (8 records) with 2024 (8) — a three-year span. 2024 is the most recent year we treat as complete: publication lags filing by roughly 18 months, so 2025 onwards are still filling in and any growth rate that ends there would understate the field.
What this landscape covers
This dataset tracks 95 published records filed against processing-in-memory (PIM) architectures paired explicitly with AI accelerator or neural-processor language — a narrower cut than general in-memory computing, and narrower still than AI hardware overall. The search string requires both a memory-side term (“processing in memory”, “PIM accelerator”, “in memory computing”, “memory neural accelerator”) and an accelerator-side term (“AI accelerator”, “neural processor”, “machine learning hardware”) to co-occur in title, abstract or claims, which is why the volume is modest relative to AI hardware patenting broadly.
Filings begin appearing from 2017, build through the early 2020s, and peak at 40 in 2025 before the current year's partial count. Because publication typically lags filing by about 18 months, the apparent softening after 2025 is an artifact of the data cut-off rather than a real pullback — read the last one to two years as incomplete, not declining.
Let an AI agent run this analysis on your own technology
Pick a task. Every answer cites the patents behind it.
Filing trend and technology composition
Two views of the same 95-record set: how filing volume moved year over year, and which IPC subclasses the claims actually sit in.
Filings held flat into the most recent complete year
From 2017's zero baseline, filings rose through the early 2020s and reached a peak of 40 in 2025. The comparable, complete-year figure to trust is 2021 to 2024: 8 filings in each year, a 0% span. Do not read 2025's higher count or 2026's partial count as acceleration or as decline — both are still filling in.
Digital processing and AI-model classes lead; memory-manufacture classes are marginal
G06F (electric digital data processing) appears in 66.3% of the 95 records and G06N (AI-model computing) in 42.1%, confirming that most filings frame the invention as a computing or AI method first. G11C (static and digital memories) appears in 29.5%, the clearest memory-hardware signal. Optical computing (G06E), semiconductor devices (H01L), memory device manufacture (H10B) and H10W each register a single record — thin enough to treat as edge cases rather than established sub-fields.
Shares are the percentage of the 95 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on AI Accelerators — Processing-in-Memory AI Acceleration Patent Landscape with Eureka
This page is one run against one query. Ask Eureka your own question about ai accelerators — processing-in-memory ai acceleration patent landscape and every answer comes back with the patent numbers behind it.
Try EurekaThe most-cited records and one to watch
US20260140878A1 — Artificial intelligence accelerator having computing units heterogeneously integrated with memory dies
The filing describes a 3D-integrated AI accelerator: a processing block with parallel processing cores sits laterally alongside a memory block over a common substrate, with a logic base die interposed vertically between substrate and memory block, and electrical connections through the substrate heterogeneously linking the two.Filed by Adeia Semiconductor Technologies LLC, published 2026-05-21.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20210263865A1 | Operation method of an accelerator and system including the same | 12 |
| 2 | US20210065767A1 | Memory with artificial intelligence mode | 9 |
| 3 | US11004500B2 | Memory with artificial intelligence mode | 7 |
| 4 | US20250202840A1 | Abstraction of Compute Nodes and Accelerators in a Switched CXL Interconnect | 6 |
| 5 | US20210263870A1 | Accelerator, operation method of the accelerator, and an apparatus including the accelerator | 6 |
| 6 | US20210064246A1 | Artificial intelligence accelerator | 6 |
| 7 | US20250202839A1 | Elastic Multi-Directional Resource Augmentation in a Switched CXL Fabric | 5 |
| 8 | US20250199980A1 | Translating Between CXL.mem and CXL.cache Read Transactions | 5 |
| 9 | US20230176739A1 | Artificial intelligence accelerator | 5 |
| 10 | US20210065754A1 | Activation functions for artificial intelligence operations | 5 |
Citation counts inside this searched corpus favour older publications that have had more time to accumulate citations — treat them as a signal of influence on the field, not of current commercial weight.
Each row carries its publication number; clicking a row searches Eureka by that number.
Put your own technology through the same analysis
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →What the filing pattern tells you
Three read-outs from the ranking, the trend and the citation table that matter for a freedom-to-operate or whitespace call.
One leader, then a real drop-off
The top-ranked assignee holds 25 records; by fifth place that is down to 13, and tenth place holds just 2. With only 12 companies in the ranked set, this is a landscape with a clear leader and a long tail rather than even competition across many filers.
Flat over the last complete three-year span
Filing counts moved from 8 in 2021 to 8 in 2024 — 0% change over that span. The 2025 peak of 40 and the 2026 partial count sit outside the comparable window and should not be read as a trend break in either direction.
Compute framing outweighs pure memory framing
Two-thirds of records touch G06F and over four in ten touch G06N, both ahead of G11C memory classification at 29.5%. Most applicants are claiming the accelerator or the computing method, with the memory element as a supporting structure rather than the headline.
US-centred filing with a modest international spread
United States filings account for 72 of the 95 records in scope, with Europe (EPO) at 14, WIPO/PCT at 6, Germany at 2 and Austria at 1. Anyone assessing regional freedom to operate should treat the US as the primary battleground and the EPO route as the secondary check.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to ai accelerators — processing-in-memory ai acceleration patent landscape, with the prior art for and against each one.
Who is filing, and what is still open
The ranked set is small — 12 companies — which makes both the leader's position and the under-claimed branches easier to see than in a crowded field.
A single assignee sets the pace
The top-ranked company holds 25 of the records in the ranking, well ahead of the rest of the field. That volume alone does not confirm technical breadth — it confirms claim-space occupation, which is a different thing to check when scoping a design-around.
A viable second tier exists
Fifth place in the ranking holds 13 records, meaningfully below the leader but well above the tail. This tier is where most freedom-to-operate work in this space will actually collide with active portfolios, since the leader's claims are already well mapped.
Recent-year counts have thinned across the board
Multiple ranked assignees show sharp latest-year declines, several down to zero filings year over year. Given the roughly 18-month publication lag, this looks more like incomplete recent data than a real pullback — but it means the latest-year snapshot understates who is still active.
| Assignee | Recent year | YoY |
|---|---|---|
| UniFabrix Ltd | 3 | -87% |
| Micron Technology Inc | 0 | -100% |
| Samsung Electronics Co., Ltd. | 0 | -100% |
| Lodestar Licensing Group LLC | 0 | -100% |
| Seoul National University R&DB Foundation | 0 | — |
| Renesas Electronics Corporation | 0 | -100% |
| Macronix International Co., Ltd. | 0 | — |
| ADEIA SEMICON TECH LLC | 0 | -100% |
Where to take this analysis
The landscape points to a concentrated leader, a thin second tier and several under-claimed structural branches. The next step is usually specific to your position in the field.
Scope a design-around against the leader
With 25 records concentrated in one assignee's portfolio, a design-around needs claim-by-claim mapping rather than a keyword search — start with the most-cited records in this set.
Explore claims in EurekaCheck the under-claimed branches before filing
Optical PIM paths, heterogeneous die integration and CXL-switched interconnects all register at low single-digit record counts. Confirm they are genuinely open, not just thinly indexed, before drafting.
Run a whitespace search in EurekaTrack the second-tier filers
The gap between the leader and fifth place is where most active competitive collisions will happen. Watching this tier closely is more useful than watching the leader alone.
Set up assignee alerts in EurekaCommon questions on PIM AI accelerator patents
This dataset identifies 95 published records that explicitly combine processing-in-memory language with AI accelerator or neural-processor language in the title, abstract or claims. That is a narrow, precisely-defined slice — broader searches on in-memory computing or AI hardware separately would return far more. Treat 95 as the count for this specific technical intersection, not for PIM or AI hardware overall.
The ranking covers 12 companies, with the top-ranked assignee holding 25 records against 13 at fifth place and 2 at tenth. That pattern shows a clear leader followed by a real drop-off rather than even competition. It is a small ranked field, so a company's absence from it does not mean it is inactive — it may simply file outside this search's exact term combination.
Over the last comparable complete-year span, filings held flat: 8 in 2021 and 8 in 2024, a 0% change. The 2025 count of 40 and the partial 2026 count look higher and lower respectively, but publication lags filing by roughly 18 months, so recent years are still filling in. Treat the 2021-2024 comparison as the reliable read, not the raw year-by-year chart.
Two-thirds of the 95 records (66.3%) touch G06F, electric digital data processing, and 42.1% touch G06N, AI-model computing — both ahead of G11C memory classification at 29.5%. This means most applicants are framing their claims around the computing method or accelerator architecture, with the memory element as a supporting structure. Classes overlap since a record can carry several, so these shares are not mutually exclusive.
IPC classes with only one record each — optical computing, semiconductor device structure, memory device manufacture and a residual H10W class — are thin enough to warrant a closer look rather than an assumption of prior art density. Structural branches like heterogeneous die integration and CXL-switched interconnects also register lightly against the compute-heavy core of G06F and G06N filings. A freedom-to-operate search focused on these narrower branches is more useful than a broad keyword sweep across the whole 95-record set.
Research AI Accelerators — Processing-in-Memory AI Acceleration Patent Landscape in depth with Eureka
Go past this page: query the whole ai accelerators — processing-in-memory ai acceleration patent landscape corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries inherent limitations: recent filings (typically the most recent 18–24 months) are under-counted due to standard publication lag; counts may be reported at either a patent-family or a patent-record basis and are not always directly comparable; classification, applicant-name, and citation data may contain errors, duplicates, or omissions; and the underlying search query defines and constrains the scope shown. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, or validity search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. Any patent, commercial, or strategic decision should be verified independently and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company’s registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.