In this dataset it means filings that combine GPU or GPGPU architecture language with memory-integration terms such as HBM integration, unified memory, or on-package memory, classified under G06F13/16, H01L25/065 or G06F12. That combination requires the claim to treat memory placement or addressing as a core architectural element rather than a downstream implementation choice. Filings that mention memory only in passing, without tying it to GPU structure, fall outside this corpus.

The corpus shows filings rising from 11 in 2017 to a peak of 37 in 2020, then falling to 6 by the 2022 midpoint. This pattern is typical of a claim space that was actively contested during a specific hardware generation and has since settled, with fewer parties filing broad foundational claims and more filing narrower, defensive ones. Publication lag also means the 2025-2026 figures are undercounted and should not yet be read as confirmation of continued decline.

NVIDIA shows the most recent-year filing activity in this corpus, holding steady with 0% year-over-year change, while Intel's latest-year activity dropped -100% YoY to zero. Microsoft and AMD both show zero filings in the latest year tracked. Co-assignee data also shows NVIDIA's strongest collaborative pairs are with individual named inventors rather than other companies, suggesting concentrated in-house inventor teams.

Sub-areas with thin representation relative to the core unified-memory and HBM-integration claims include dynamic memory reconfiguration for mixed inference and graphics workloads, error detection in embedded compute-adjacent memory, and cross-die HBM stacking address translation. These sit adjacent to the most-cited records in the corpus but have not attracted the same filing density, which is worth checking before assuming a claim direction is already crowded. A prior-art search focused specifically on these narrower branches is more useful than a broad unified-memory search at this stage.

Not necessarily. Citation counts inside a fixed corpus accumulate over time, so older filings such as the 2004-priority multithreaded-kernel record naturally collect more citations simply by having been available longer for later filers to reference. A high count signals influence on subsequent filings, not current commercial relevance or enforceability. Recent filings with few citations may still be more relevant to a present-day freedom-to-operate question.