Samsung PIM technology
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Samsung unveiled LPDDR5X-PIM at Hot Chips 2026, embedding AI compute in DRAM. It triples AI inference speed, targeting on-device AI in PCs and mobile.
Background
Samsung has been developing Processing-in-Memory (PIM) for years, starting with HBM-PIM in 2021. At Hot Chips 2026, they detailed LPDDR5X-PIM, which integrates compute units into low-power DRAM to reduce data movement between memory and processor, a major bottleneck for AI workloads. This shift reflects the growing demand for efficient on-device AI as large language models become smaller and run locally on devices.
How it unfolded
- 2021Samsung introduced PIM technology for HBM (Aquabolt-XL HBM2-PIM proof of concept).
- Aug 25, 2026Samsung presented detailed specifications of LPDDR5X-PIM at Hot Chips 2026 at Stanford University, showcasing working silicon.
- Aug 26, 2026Reports emerged that Samsung's AI PC chipset 'GAIA' may incorporate LPDDR5X-PIM, with mass production possible as early as 2027.
Who’s saying what
- Samsung
- Samsung claims LPDDR5X-PIM delivers up to 3x AI inference throughput and 2x processing speed compared to conventional LPDDR5X, with full compatibility with existing systems.
- Analysts
- Analysts note that PIM faces technical and ecosystem challenges, but GAIA could be a key test of its commercial potential.
- Competitors
- SK hynix and Micron are also preparing PIM technology for next-generation LPDDR6, indicating intensifying competition.
Still unverified
The claim that GAIA will be the first PIM commercialization in AI PCs and mass production by 2027 is based on sources cited by ZDNet and is not officially confirmed by Samsung.