Hot Chips 2026: Samsung makes LPDDR5X smart with logic unit in memory — LPDDR5X-PIM is 3.01x faster than LPDDR5X in AI inference with 8x the bandwidth

Samsung unveiled LPDDR5X-PIM, a memory with in-memory logic, at Hot Chips 2026. The tech is 2.28x faster than LPDDR5X in AI inference and 3.01x faster in tokens per second. Samsung aims to reduce data movement, lower power consumption, and target server, client, and mobile markets. The company plans to develop LPDDR6X-PIM with JEDEC.

Original reporting
Published Aug 25, 2026, 9:15 PM UTC
Analysis
alphai AI DeskAI-generated
Added to alphai Aug 25, 2026, 9:32 PM UTC. Informational, not investment advice.
How this was made
alphai summarizes source reporting and applies a structured AI analysis for relevance, timing, sentiment and ticker impact. Always verify material claims with the original publisher.
Hot Chips 2026: Samsung makes LPDDR5X smart with logic unit in memory — LPDDR5X-PIM is 3.01x faster than LPDDR5X in AI inference with 8x the bandwidth — source image
Decision brief

The 30-second read

Low
01

Why it matters

The announcement introduces a new memory class that could reduce AI system costs, but market impact depends on adoption by OEMs and accelerator designers.

02

Market read

While the technology is novel, immediate trading relevance is low; investors may watch for future adoption signals.

03

What to watch

Potential supply chain constraints for 561‑ball packages and the need for ecosystem support.

Relevance 4/10Novelty 5/10Timing: after Hot Chips 2026 presentation

Background

Samsung unveiled its first LPDDR5X-PIM memory at Hot Chips 2026, claiming significant AI inference speed and bandwidth gains.

Market effects

Highlights potential shift toward LPDDR5X-PIM in AI edge devices, may spur competition in memory market.

Primarily impacts South Korean semiconductor sector and global AI hardware supply chains.

Limited immediate effect; longer-term relevance for AI accelerator manufacturers.

Counterpoint

Adoption may be slow due to power consumption concerns and competition from HBM solutions.

Key entities

  • Samsung Electronics

    Introduced LPDDR5X-PIM memory with claimed 2.28‑3.01× AI performance improvements.

  • Nvidia

    Mentioned as a user of LPDDR5X in its Vera CPU.

  • Intel

    Uses LPDDR5X in its Crescent Island AI accelerator.

Related articles

$NVDAMed

Nvidia Unveils Custom HBM Design With 30% More Bandwidth Than HBM4E

Nvidia introduced NVHBM, a custom high-bandwidth memory technology with 30% more bandwidth than HBM4E. The design moves the memory controller to the base die, improving performance and power efficiency. Nvidia plans to collaborate with suppliers like Samsung and SK Hynix for manufacturing. The industry expects growing demand for custom HBM solutions.

$NVDAMed

NVHBM Raises Memory Bandwidth by 30% as NVIDIA Seeks to Lead the Custom HBM Market

NVIDIA unveiled NVHBM, a custom HBM base die technology, aiming to lead the custom high-bandwidth memory market. Developed with major memory suppliers and Annapurna Labs, NVHBM integrates the memory controller into the HBM base die, potentially increasing memory bandwidth by 30%, reducing power consumption by 15%, and freeing up 25% more GPU area. NVIDIA aims to shift base die design from memory makers to xPU companies like itself.

$NVDAMed

Nvidia custom 'NVHBM' promises 30% higher bandwidth, 15% lower power than commodity HBM4e — custom base die and PHY will be available to NVLink Fusion partners

Nvidia introduced NVHBM, a custom high-bandwidth memory for AI accelerators, promising 30% higher bandwidth and 15% lower power than HBM4e. It's available to NVLink Fusion partners, with Amazon's Annapurna Labs as the first adopter. The technology aims to improve performance and efficiency for AI workloads, though benefits won't appear in current systems.

$MUMed

South Korea Chip Data: Memory Up 277%, Logic Down; AI Buys Bandwidth, Not Compute

South Korea July 2026 semiconductor export data showed memory chip shipments up 276.9% year over year, while system semiconductor exports fell 0.7%, according to the Ministry of Trade, Industry and Energy. The article links the divergence to AI inference needing higher memory bandwidth. It cites Q2 2026 profits for Samsung and SK Hynix and notes a pending US DRAM supply-restriction class action.

$BANDMed

5 Insightful Analyst Questions From Bandwidth’s Q2 Earnings Call

Bandwidth reported Q2 revenue of $219.9M, up 22.2% year over year and slightly above estimates, with adjusted EPS of $0.37 and adjusted EBITDA of $27.77M. It raised full-year revenue guidance to $905M midpoint but lowered adjusted EPS to $1.75. Management cited AI-enabled voice and messaging demand; guidance and margin stability drove a negative market reaction.

$BANDMed

Bandwidth Inc. Q2 2026 Earnings Call Summary

Strategic Performance Drivers AI is fundamentally shifting enterprise buying decisions toward providers capable of supporting mission-critical, compliant, and high-performance voice interactions. The company secured five new million-dollar-plus customer wins and expansions in Q2, all of which included Maestro or AI services, signaling strong product-market fit.