Samsung Put a Processor Inside Your Phone’s RAM

Every time your phone runs an AI feature on-device, most of the energy goes into hauling data out of memory, over to the processor, doing a bit of math, and hauling it back. Again and again. The math is cheap. The commute kills your battery.

Samsung's answer, shown at the Hot Chips conference this week, is to stop commuting. LPDDR5X-PIM puts small compute units inside the memory chip, so the calculations happen where the data already lives. It's the first time this has been done with LPDDR, the low-power memory that goes into phones and laptops instead of data center racks.

The numbers Samsung showed

Running Meta's Llama 3.1 8B model on an edge AI accelerator, Samsung measured 81.3 tokens per second with LPDDR5X-PIM against 27 tokens per second with conventional LPDDR5X. Task completion time dropped from 12.3 seconds to 5.4 seconds. Effective bandwidth in PIM mode reaches 614 GB/s, about eight times what standard LPDDR5X-9600 manages.

An on-device assistant that currently answers at a crawl would answer at a pace you'd tolerate, without shipping your data to a server first.

LPDDR5X-PIM also uses the same JEDEC standard 561-ball package as regular LPDDR5X, so a device maker can drop it in without redesigning the board. Technologies that require everyone downstream to re-engineer their hardware tend to die quietly. This one doesn't need that.

What Samsung is not claiming

Samsung called this a complement to HBM rather than a replacement. HBM4 delivers up to 2TB/s per stack, more than triple what PIM mode does, and it remains the standard for large-scale training and high-throughput cloud inference.

There's also an open question about accuracy. An attendee at the session noted that LPDDR5X-PIM's outputs differ slightly from standard LPDDR5X. Samsung confirmed the discrepancy, said optimization work is ongoing, and expressed confidence that fixing it won't cost the performance gains. The size of that gap has not been made public, so the benchmark numbers are preliminary.

Why it matters beyond the spec sheet

Memory is currently eating the tech industry. HBM's share of AI chip component cost climbed from 52% in early 2024 to 63% by the end of 2025, and the knock-on effects have been showing up in your shopping cart all year. Apple just raised the Mac mini's starting price to $899 and pointed at component costs. Phone makers are dealing with the same squeeze.

Anything that makes AI inference possible on cheaper, lower-power memory takes pressure off that system. JEDEC previewed PIM features for the upcoming LPDDR6 standard back in April, which suggests this is heading toward being normal rather than staying a Samsung science project.

You won't buy a phone advertising PIM memory anytime soon. The gains show up as on-device AI that runs fast enough to use without thinking about it.