Robotics & Physical AIAug 30, 2026

Samsung puts compute inside LPDDR5X memory banks — and the software cost is the story

At Hot Chips 2026, covered by Chips and Cheese on 29 August 2026, Samsung detailed an LPDDR5X-9600 part with processing-in-memory blocks inside each of its 16 DRAM banks. Working from internal bandwidth the blocks see 614 GB/s where an external controller sees 76.8 GB/s, and each chip contributes about 1.2 INT8/FP8 TOPS (2.4 with 4-bit weights) — eight chips land near the NPU in Intel Meteor Lake. Chips and Cheese notes the catch: PIM mode needs uncacheable mappings and does not sit well with multithreading, caching or speculative execution, so the memory-side win is paid for on the CPU side.

What it means On-device inference keeps running into memory bandwidth, and this is the clearest public accounting of what moving compute into DRAM actually costs the rest of the machine.

Where it came from Chips and Cheese

Back to the Stream