Neural Network Inference Acceleration Patents: Who Leads, Gaps 2026
- Filings peaked in 2020 at 81 and have since declined, suggesting the core sparse-accelerator claim space is now well staked out rather than still expanding.
- Recent-year momentum has gone flat across the largest holders, with Intel, Microsoft, NVIDIA and Xilinx all showing zero filings in the latest year in this dataset.
- The most-cited records all center on sparse convolutional accelerators, meaning the highest-traffic prior art is concentrated in one architectural family, not spread evenly.
Filing growth compares 2021 (80 records) with 2024 (70) — 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. Top-5 share is the combined record count of the five largest assignees divided by all 564 records in scope (CR5), not by the ranked leaders only.
What this landscape covers
This dataset tracks patent families combining accelerator architecture terms — neural network accelerator, inference acceleration, AI accelerator — with implementation-level terms in the claims and description: quantization, sparsity, dataflow, systolic array, and energy efficiency. The IPC filter narrows to G06N3, G06F15 and G06F7, which is where hardware-level neural computation claims are classified. The result is a corpus centred on how inference is executed in silicon, not on model architectures or training methods themselves.
Records span 2015 through the mid-2026 cut-off. Because publication typically lags filing by around 18 months, counts for the most recent one to two years are understated and should be read as a floor, not a ceiling.
Filing trend and technology composition
Two views of the same 564-family corpus: how filing activity has moved over time, and which IPC subclasses carry the claim weight.
A 2020 peak followed by a pull-back
Filings rose from 21 in 2017 to a peak of 81 in 2020, held near the midpoint at 62 in 2022, and have since fallen — 13 in the most recent (partial) year. Read the tail end cautiously given publication lag, but the shape from 2020 onward reads as consolidation rather than continued growth.
G06N and G06F dominate the classification mix
G06N (AI-model computing) appears in 537 of 564 records and G06F (electric digital data processing) in 298, confirming this is fundamentally a hardware-execution corpus. Image/video-specific subclasses (G06V, G06T, G06K) and memory/coding subclasses (G11C, H03M) each carry single-digit-to-low-double-digit counts — present as application context, not as the centre of gravity.
Shares are the percentage of the 564 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on Neural Network Inference Acceleration with Eureka
This page is one run against one query. Ask Eureka your own question about neural network inference acceleration and every answer comes back with the patent numbers behind it.
Try EurekaThe most-cited records in this corpus
Systolic-array accelerator for recurrent neural network data processing
Proposed is a data parallel processing method for a recurrent neural network in a neural network accelerator based on a systolic array. A data processing device receives voice data of a user in a predetermined time section, separates it by sentence into voice data units, vectorizes those units into input vectors, and feeds the vectors to a systolic-array-based neural network accelerator for processing.Filed by the Electronics and Telecommunications Research Institute, published 2023-06-08 — illustrates how systolic-array dataflow claims are now being paired with specific application pipelines (voice segmentation) rather than filed as bare architecture claims.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20180046906A1 | Sparse convolutional neural network accelerator | 438 |
| 2 | US20180046900A1 | Sparse convolutional neural network accelerator | 371 |
| 3 | US10891538B2 | Sparse convolutional neural network accelerator | 340 |
| 4 | US20180121796A1 | Flexible neural network accelerator and methods therefor | 316 |
| 5 | US10528864B2 | Sparse convolutional neural network accelerator | 312 |
| 6 | US10860922B2 | Sparse convolutional neural network accelerator | 307 |
| 7 | US20190286973A1 | Hardware accelerated neural network subgraphs | 185 |
| 8 | US20180046916A1 | Sparse convolutional neural network accelerator | 175 |
| 9 | US20190286972A1 | Hardware accelerated neural network subgraphs | 132 |
| 10 | US20190114534A1 | Neural network processing system having multiple processors and a neural network accelerator | 129 |
Citation counts inside a searched corpus favour older filings that have had more time to accumulate citers — treat this as a signal of influence on the field, not of current commercial relevance.
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 numbers say about this field
Three read-throughs from the filing trend, the citation table, and the co-assignee pairs.
The growth phase has passed its peak
Filings climbed steadily to a 2020 high, held near that level through 2022, then declined. Combined with the recent-year momentum data showing several major holders at zero, this reads as a maturing claim space rather than an emerging one — new entrants face denser prior art than the 2017 baseline suggests.
Influence clusters around sparse convolutional accelerators
The five most-cited records in this corpus are all variants on sparse convolutional neural network accelerators, with the leading record cited 438 times. That concentration means anyone designing a sparsity-handling datapath is working against a small number of heavily-cited reference points, not a diffuse literature.
Academic-institute pairing outweighs corporate co-filing
The strongest co-assignee pair in the dataset — Zhejiang University and Zhejiang Lab — outnumbers any corporate pairing observed, including Intel's pairings with individual named inventors. Joint university-lab filing appears to be a more active collaboration channel here than cross-company alliances.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to neural network inference acceleration, with the prior art for and against each one.
Who is filing, and who has slowed down
Recent-year momentum diverges sharply from historical citation leadership: the accelerator pioneers that built the most-cited prior art are not the ones filing now.
SambaNova Systems, Inc. is the only holder still filing
Among the tracked assignees, SambaNova Systems, Inc. (SambaNova Systems) is the only one with an active filing in the latest year captured. Every other major holder listed — Intel, Microsoft, NVIDIA, Xilinx, ETRI — shows zero filings in the same window.
Intel's filing pace has stopped, not slowed
Intel's co-assignee activity (pairings with individually named inventors, at 5 and 4 occurrences) points to a period of active internal collaboration that has not carried into the latest year, which shows a full year-over-year drop to zero.
A dense academic pairing sits outside the corporate momentum picture
Zhejiang University and Zhejiang Lab form the strongest co-assignee pair in the dataset, independent of the corporate names showing momentum declines — suggesting research-institute output continues even as commercial filers pause.
| Assignee | Recent year | YoY |
|---|---|---|
| SambaNova Systems, Inc. | 1 | — |
| Intel Corporation | 0 | -100% |
| Microsoft Technology Licensing, LLC | 0 | -100% |
| NVIDIA Corporation | 0 | -100% |
| Xilinx, Inc. | 0 | — |
| Electronics and Telecommunications Research Institute (ETRI) | 0 | — |
| Samsung Electronics Co., Ltd. | 0 | — |
| Korea Advanced Institute of Science and Technology (KAIST) | 0 | — |
Where to take this
The filing and citation data point to a maturing core and a thinner periphery. Two directions follow from that.
Check the periphery before the core
G11C, H03M, G06V and G06T overlaps carry single-digit-to-low-double-digit counts against a 537-record G06N core. That gap is either genuine white space or an unclaimed application niche — worth a targeted search before committing a filing strategy to the dense sparse-accelerator core.
Explore white space in EurekaTrack momentum, not just historical leadership
The assignees with the most-cited records are not the ones filing now. A monitoring approach built on recent-year momentum will surface active filers like SambaNova earlier than one built purely on citation counts.
Set up momentum tracking in EurekaCommon questions on this landscape
In this dataset, it is a patent family whose title or abstract references accelerator architecture — terms like neural network accelerator, inference acceleration, or AI accelerator — combined with an implementation-level term in the claims or description such as quantization, sparsity, dataflow, systolic array, or energy efficiency. The IPC filter further restricts the set to G06N3, G06F15 and G06F7, the classification areas covering AI-model computing and digital data processing hardware. This combination targets hardware-execution claims specifically, rather than model training methods or software-only inference optimizations.
Filings rose steadily from 21 in 2017 to a peak of 81 in 2020, held near 62 at the 2022 midpoint, and have declined since, with 13 recorded in the most recent partial year. Some of that recent-year drop is a publication-lag artifact — filings typically take about 18 months to publish, so the last one to two years understate true activity. But the pattern from 2020 onward, combined with several major holders showing zero filings in the latest year, is consistent with a claim space that grew quickly and has since consolidated.
The most-cited historical records in this corpus are dominated by sparse convolutional neural network accelerator filings, and named assignees tracked for recent momentum include Intel, Microsoft, NVIDIA, Xilinx, SambaNova Systems and the Electronics and Telecommunications Research Institute. However, citation leadership and current filing momentum diverge sharply: most of the historically cited leaders show zero filings in the latest tracked year, while SambaNova Systems is the one assignee still actively filing. Any competitive read should separate 'who built the foundational claims' from 'who is filing now.'
The IPC composition shows a heavy concentration in G06N (537 of 564 records) and G06F (298), with much thinner coverage in G06V, G06T, G11C, H03M and G06K — each in the single-to-low-double digits. That gap suggests under-claimed intersections: memory-addressing schemes for sparse weights, coding-domain compression for accelerator I/O, and energy-efficiency scheduling for systolic arrays all sit adjacent to the dense core without being heavily claimed themselves. These are worth a targeted prior-art check before assuming the broader field is closed.
Citation counts inside any searched patent corpus are structurally biased toward older filings, which have simply had more years to accumulate citers. In this dataset, the top five most-cited records are all several years old and cluster around sparse convolutional accelerator architectures. That makes citation count a reasonable proxy for historical influence on the field, but a poor one for judging which technology is commercially important today — recent filings by definition cannot yet have accumulated comparable citation counts.
Research Neural Network Inference Acceleration in depth with Eureka
Go past this page: query the whole neural network inference acceleration 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.