The bandwidth wall is the problem that no matter how fast the processor gets, it stalls waiting for data from memory. One radical answer is to stop moving the data so far - put some computation right in the memory subsystem. Samsung's 2021 HBM grant claims a version of that.
The grant US11069400B1 (issued July 20, 2021; SAMSUNG ELECTRONICS CO., LTD.; CPC G11C 11/4096 DRAM access plus G06F 9/3001 and G06F 9/30101 - instruction and arithmetic-logic codes) claims a high-bandwidth memory and system. The presence of G06F 9/3001 arithmetic-processing classifications inside a memory patent is the signal: this HBM is meant to compute, not just store.
Do the bandwidth math on why this is attractive. If a workload reads enormous arrays just to perform a simple reduction, the energy and time go into moving bytes, not computing on them. Performing the reduction inside the memory means only the small result crosses the expensive interface - you spend bandwidth on answers, not raw data.
Samsung is both a memory leader and a logic foundry, which makes it uniquely positioned to fuse the two. A 2021 grant blending DRAM-access IP with arithmetic-logic IP is exactly what you would expect from a company that can build the memory and the compute in the same fab.
This is conceptual kin to the tiered-memory ideas elsewhere in the sector: both accept that bandwidth is scarce and try to use it more sparingly. Where tiering routes hot data to fast memory, processing-near-memory avoids the trip entirely for some operations.
For the reader, the through-line is that the bandwidth wall is being fought on multiple fronts - wider stacks, tiering, and compute-in-memory. Samsung's 2021 grant stakes out the last of these in the HBM context, well before processing-in-memory became a common pitch.