Samsung's LPDDR5X-PIM: 8x Memory Bandwidth, but a Software Nightmare

Samsung's Processing-in-Memory (PIM)

At Hot Chips 2026, Samsung detailed its Processing-in-Memory (PIM) implementation in LPDDR5X DRAM, placing MAC units inside each of the chip's 16 banks to exploit internal bandwidth. This yields 614 GB/s, eight times the standard 76.8 GB/s, and up to 2.4 TOPS per chip. While it works with standard memory controllers via special row addresses, the design forces software into a corner: PIM mode disables caching, prefetching, and out-of-order execution, and complicates multitasking. The article explores these challenges and suggests hardware changes to make in-memory compute practical.