In brief: Not every language-model layer needs the same compute location. Samsung's application US20260330207A1, published on October 22, assigns tasks to an external processor, a unit near memory and computing logic inside memory. The deciding factor is the number of operations performed per amount of data. The design describes the control mechanism; it does not demonstrate a measured speedup.
Trading data movement for compute capacity
For high computational intensity, a controller supplies data to the external processor, such as a GPU. Processing Near Memory (PNM) handles medium intensity, while Processing In Memory (PIM) handles low intensity. PIM processes data inside DRAM, avoiding transfers outside it, but has little space there for complex arithmetic logic. PNM provides more logic area but must fetch data over the memory channel. The external processor has greater computational resources and additionally requires the transfer path to the host.
Claim 1 seeks this three-level assignment. Claim 9 gives examples: conventional multi-head attention for PIM, and grouped-query attention or expert networks for PNM. Description paragraphs [0101]–[0116] explain corresponding flows during output-token generation. CXL is one possible interface, rather than a requirement for every described implementation.
The application on record
The title is “Heterogeneous Computing Memory Device and Artificial Intelligence Computing System Including the Same.” Samsung Electronics is the applicant; the cover additionally names Seoul National University R&DB Foundation as an assignee. US application 19/446,433 was filed on January 12, 2026, claiming Korean priority dated April 18, 2025 (10-2025-0050877). A1 denotes an application publication. We could not independently confirm the current official prosecution status on October 24.
Pandorex View: Routing must earn its overhead
PIM and PNM are established concepts; Samsung discussed their programming in 2024. The specific publication is new here. It provides no reproducible figures for tokens per second, energy consumption or chip area. Whether the assignment delivers practical gains also depends on model structure, data transfers and control overhead. It does not establish an announced production product.