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 →Filing growth compares 2021 (69 records) with 2024 (35) — 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 1,095 records in scope (CR5), not by the ranked leaders only.
Low precision training formats reduce the bit width used to represent weights, activations and gradients during neural network training — block floating point, shared exponent grouping, and the surrounding machinery of loss scaling, rounding mode selection and accumulation precision that keeps numerically unstable training runs converging. The dataset in scope spans 1,095 published records from 2015 through mid-2026, drawn from filings that combine format-level claims (block floating point, shared exponent) with the supporting numerical controls (dynamic range, loss scaling, accumulation precision, outlier channel handling, rounding mode, mantissa bit width).
Publication lags filing by roughly 18 months, so the most recent one to two years in any trend understate actual filing activity. Family counts, used throughout the ranking below, are the fairer unit for cross-jurisdiction comparison because they neutralise continuation filings and duplicate regional filings of the same invention.
Two views of the same 1,095 records: how filing activity has moved year over year, and which IPC subclasses the claims actually sit in.
Filings rose to a peak of 115 in 2019, then declined across complete-year data to 35 in 2024, a 49% drop from 2021's 69 filings. 2025 and 2026 figures are still filling in under the usual publication lag and should not be read as a continued decline.
G06F (electric digital data processing) and G06N (AI-model computing) anchor the field at 44.4% and 19.2% of the 1,095 records respectively. Four communications-related subclasses — H04B, H04L, H03M and H04N — each cover more than 9% of records, reflecting how format claims extend into transmission and coding hardware, not just training silicon.
Shares are the percentage of the 1,095 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
This page is one run against one query. Ask Eureka your own question about low precision training formats and every answer comes back with the patent numbers behind it.
Try EurekaA system for block floating point computation in a neural network receives a plurality of floating point numbers. An exponent value for an exponent portion of each floating point number is identified and mantissa portions of the floating point numbers are grouped. A shared exponent value of the grouped mantissa portions is selected according to the identified exponent values and then removed from the grouped mantissa portions to define multi-tiered shared exponent block floating point numbers. One or more dot product operations are performed on the grouped mantissa portions of the multi-tiered shared exponent block floating point numbers.Filed by Microsoft Technology Licensing, published 2019-11-14.


| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US5889868A | Optimization methods for the insertion, protection, and detection of digital watermarks in digitized data | 846 |
| 2 | US6522767B1 | Optimization methods for the insertion, protection, and detection of digital watermarks in digitized data | 347 |
| 3 | US20120250740A1 | OFDM signal processing in a base transceiver system | 288 |
| 4 | US5537435A | Transceiver apparatus employing wideband FFT channelizer with output sample timing adjustment and inverse FFT… | 233 |
| 5 | US20130322753A1 | Systems and methods for local tone mapping | 226 |
| 6 | US5839100A | Lossless and loss-limited compression of sampled data signals | 216 |
| 7 | US4760525A | Complex arithmetic vector processor for performing control function, scalar operation, and set-up of vector s… | 215 |
| 8 | US20190340492A1 | Design flow for quantized neural networks | 210 |
| 9 | US5410750A | Interference suppressor for a radio receiver | 210 |
| 10 | US20040012600A1 | Scalable high performance 3d graphics | 197 |
Citation counts favour older filings that have had more time to accumulate citations within the searched corpus — read them as a signal of influence on later filings, not as a ranking of current technical importance.
Each row carries its publication number; clicking a row searches Eureka by that number.
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 →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 →Three patterns worth reading before deciding where to file or challenge.
The five leading assignees combined hold 41.8% of all 1,095 records in scope, rising to 59.9% for the ten leading assignees. That leaves a long tail of single- or low-count filers competing for the remaining claim space — a structure typical of a field where the core numerical techniques were staked out early by a handful of large compute and communications incumbents.
Filing peaked at 115 in 2019 and has trended down since, with the last fully-published years showing a 49% drop from 2021 to 2024. Because publication lags filing by around 18 months, 2025-2026 figures are still incomplete and should not be read as confirming a continued slide — but the multi-year direction through 2024 is real.
Electric digital data processing (G06F) and AI-model computing (G06N) dominate, but transmission subclasses H04B and H04L each exceed 12% of the 1,095 records. This overlap suggests block floating point and shared exponent techniques are being claimed for signal and video pipelines as well as neural network training, widening the prior art a new filer must clear.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to low precision training formats, with the prior art for and against each one.
The leading assignee sits well above the rest of the ranked field, and recent-year filing activity has thinned out even among the incumbents.
The leading assignee holds 171 records, more than triple the fifth-place count of 47 and well above tenth place at 27. That gap at the top is the clearest structural feature of this dataset — it is not an evenly distributed field.
Several of the most active historical filers show sharp year-over-year declines in the latest tracked year, with some down to zero and one down 80% year over year. Given the publication lag, this looks more like incomplete recent data than an actual retreat from the technology.
Only ten co-assignee pairs appear in the dataset, and the strongest pairing appears 19 times — modest compared with the leading solo filer's 171 records. Most activity in this field is filed by a single assignee rather than through joint ventures or cross-licensing structures.
| Assignee | Recent year | YoY |
|---|---|---|
| Intel Corp | 1 | -80% |
| NVIDIA Corp | 1 | — |
| Microsoft Technology Licensing, LLC | 0 | -100% |
| Telefonaktiebolaget LM Ericsson (publ) | 0 | -100% |
| Wistaria Trading Inc | 0 | — |
| Apple Inc | 0 | — |
| Ericsson Inc | 0 | — |
| Dolby Laboratories Licensing Corp | 0 | — |
Two directions depend on whether you are scoping freedom-to-operate or looking for filing opportunity.
The concentration at the top of this field means a new filing in block floating point or shared exponent grouping is likely to sit close to existing claims from the leading assignees. Reading the independent claims of the leading filer's family is a faster first step than a full landscape search.
Explore claims in EurekaRounding-mode selection, mixed mantissa scheduling and outlier channel isolation show thinner concentration than the core block floating point claims. These are worth monitoring for new entrants before they consolidate the way the core format claims have.
Set up monitoring in EurekaIn this dataset it means claims combining a reduced-precision numerical representation — typically block floating point or shared exponent grouping — with the supporting numerical controls that keep training stable, such as loss scaling, accumulation precision, rounding mode selection or dynamic range management. A filing that only claims a generic quantization scheme without these specific mechanisms usually falls outside this scope. The search string behind this landscape requires both a format term and a numerical-control term to appear together, which is why the corpus centres on training-stability techniques rather than general model compression.
The assignee ranking behind this landscape covers 100 companies, with one clear leader holding 171 records — more than three times the fifth-place count of 47. The top five assignees combined account for 41.8% of all 1,095 records in scope, and the top ten account for 59.9%. Beyond that group, filing activity thins into a long tail of companies with single-digit record counts, so competitive attention should focus on the leading names and watch for new entrants rather than assuming an even spread.
Filing peaked in 2019 at 115 records and has declined since across complete-year data, with 2021's 69 filings dropping to 35 by 2024 — a 49% fall over that span. Because publication lags actual filing by roughly 18 months, the 2025 and 2026 figures in any dataset are still incomplete and will rise as more records publish, so it is too early to call the current trajectory a continued decline. The safest reading is that filing has cooled from its 2019 peak but the last one to two years remain an open question.
This Microsoft Technology Licensing filing, published 2019-11-14, covers block floating point computation using shared exponents: identifying an exponent value for each floating point number in a group, grouping the mantissa portions, selecting a shared exponent for the group, and performing dot product operations on the resulting shared-exponent block floating point numbers. The claims are specifically about the shared-exponent grouping and dot-product mechanics rather than block floating point in general, so implementations using a different grouping or scaling scheme may sit outside its scope — that boundary is worth checking claim-by-claim rather than assuming coverage from the abstract alone.
The technology composition shows the core claims concentrated in G06F (electric digital data processing, 44.4% of records) and G06N (AI-model computing, 19.2%), with less crowded overlap into transmission and coding subclasses like H03M and H04W. Within that spread, mechanisms like rounding-mode selection for gradient accumulation, mixed mantissa bit-width scheduling across network layers, and outlier channel isolation for activation quantization show thinner concentration at the top than the core block floating point and shared exponent claims. These are reasonable areas to search before filing, though thinner concentration is not proof of an open path — it reflects the class-level spread in this corpus, not a freedom-to-operate clearance.
Go past this page: query the whole low precision training formats 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.
Help us improve
Something wrong, out of date, or missing on this page? Tell us and it goes straight to the team that maintains it.
This page