• Skip to main content
  • Skip to search
  • Skip to footer
Cadence Home
  • This search text may be transcribed, used, stored, or accessed by our third-party service providers per our Cookie Policy and Privacy Policy.

  1. Blogs
  2. Verification
  3. Memory Innovation Has Evolved Beyond Bandwidth: Next-Gen…
RS202505133435
RS202505133435

Community Member

Blog Activity
Options
  • Subscribe by email
  • More
  • Cancel
Verification IP
ddr5
Functional Verification
VIP
HBM
DDR
Lpddr6
simulation
verification

Memory Innovation Has Evolved Beyond Bandwidth: Next-Gen DDR, HBM, and LPDDR

15 Nov 2026 • 5 minute read


AI systems are exposing different memory bottlenecks at different points in the architecture. Server CPUs need capacity and low-latency access. AI accelerators need bandwidth. Power-sensitive systems need to move data more efficiently. No single memory technology solves all three.

Recent standards such as DDR5, LPDDR6, and HBM4 have advanced memory performance across a wide range of applications. At the same time, the industry is already working on next-generation technologies such as LPDDR6 PIM, DDR6, and HBM5 to support the growing demands of future AI and compute workloads.

The discussion is no longer centered solely on moving more data. The challenge is doing so efficiently, reliably, and sustainably.

Memory technologies evolve as each generation exposes new system bottlenecks. DDR, LPDDR, HBM, and PIM directions reflect a shift from one-size-fits-all memory scaling to workload-specific memory architectures.

One AI System, Different Memory Problems

Different memory technologies address different parts of the AI system.

Technology

Primary Challenge Addressed

DDR5

Scaling capacity and system bandwidth for CPU-centric workloads while maintaining DIMM-based deployment models

LPDDR6

Improving bandwidth-per-watt and memory efficiency in power-constrained systems

HBM4

Delivering very high bandwidth close to AI accelerator compute

LPDDR6 PIM

Reducing memory-to-compute data movement by enabling processing closer to stored data

DDR6

Extending DDR subsystem scaling through higher bandwidth, improved parallelism, and continued support for high-density memory systems

HBM5

Increasing accelerator-attached memory bandwidth and capacity while improving power and integration efficiency

These technologies are not interchangeable. They address different bottlenecks within the same AI platform.

As AI workloads scale, system architects are increasingly balancing bandwidth, capacity, power efficiency, latency, reliability, and data movement rather than optimizing a single metric.

More Bandwidth Creates New Constraints

Higher bandwidth and larger memory footprints are essential for scaling AI systems, but they also create new system constraints. As more data moves between the compute block and memory, designers must manage higher power demand, thermal density, package complexity, signal integrity risk, and reliability exposure across the memory subsystem.

HBM4 is designed to provide high-bandwidth, energy-efficient memory for AI accelerators and other data-intensive computing systems. By placing stacked memory close to the processor, it helps improve data access, increase memory capacity and throughput, and reduce the energy required to move data. Its adoption also brings greater focus to package integration, thermal management, signal integrity, reliability, and end-to-end subsystem verification.

This is why memory innovation is becoming more targeted. CPU-centric systems need capacity and latency efficiency; accelerators need local high-bandwidth memory, power-sensitive designs need lower energy per transfer, and AI pipelines need architectures that reduce unnecessary data movement.

The goal is shifting from simply increasing bandwidth to delivering the right bandwidth, in the right location, with the right balance of efficiency, reliability, and scalability.

Processing Is Moving Closer to Data

LPDDR6 PIM(Processing-in-Memory) extends the memory-efficiency trend by moving selected processing closer to where data is stored. Rather than sending every operation back to a host processor or accelerator, PIM architecture is intended to support targeted in-memory or near-memory operations that reduce repeated data transfers, improve effective bandwidth use, and lower energy per operation.

While LPDDR6 PIM details continue to mature through standardization, its value is best described as a technical direction rather than a fixed implementation. The key objective is to improve inference-class workload efficiency by reducing memory-to-compute traffic while preserving the low-power advantages of LPDDR-based designs.

This also expands the verification scope. New command behavior, execution ordering, memory access rules, power-state interactions, metadata or reliability handling, and error scenarios must be verified together with standard LPDDR protocol activity so that the processing capability does not compromise memory correctness.

The Growing Complexity of DDR and DIMM Verification

DDR5 gave mainstream server and compute platforms a major step forward in transfer rates, capacity, and power efficiency. Its sub-channel architecture also improves access efficiency for AI, analytics, and high-core-count CPU workloads.

DDR5 is commonly deployed through DIMMs, which combine multiple DRAM devices in standardized modules to provide scalable system capacity. Depending on the platform and module type, DIMMs may also include capabilities for identification, power management, signal integrity, thermal monitoring, and reliability.

As workload demands continue to grow, DDR5-based subsystems are increasingly tested against limits in bandwidth, capacity, signal integrity, and power delivery. The transition toward DDR6 should therefore be viewed as a broader effort to sustain memory scaling across the full subsystem, rather than simply as a speed upgrade.

This evolution also increases verification complexity across the controller, PHY, DRAM devices, and DIMM.

Training, calibration, power transitions, reliability, availability, and serviceability features, and module behavior may each operate correctly in isolation.

The greater challenge is validating that these behaviors continue to work together under realistic workload pressure.

DDR subsystem verification must therefore move beyond interface compliance and stress the complete subsystem with concurrent refresh activity, frequency and power transitions, thermal limits, error recovery, and bandwidth contention.

What Memory Verification Must Cover

  • Protocol and timing correctness
  • Training and calibration flows
  • Power-state transitions and recovery
  • Reliability and error-handling mechanisms
  • Controller, PHY, and subsystem interactions
  • System-level performance behavior

This is where accurate memory models and verification IP become essential for reducing risk before silicon.

Cadence addresses these challenges through its Memory Models and Verification IP portfolio, with capabilities for protocol verification, timing validation, training and calibration flows, subsystem verification, performance analysis, and support for current and emerging memory technologies.

Cadence Memory Models are used by more than 500 customers and support 10,000 memories spanning 100 memory interface types and 85 memory manufacturers.

What’s Next for Memory Verification?

The progression from DDR5 to DDR6, from HBM4 to HBM5, and the emergence of LPDDR6 PIM points to a common trend: increasing system complexity in pursuit of greater efficiency and performance.

Cadence is actively developing verification solutions for current and next-generation memory technologies, including DDR5 MRDIMM, LPDDR6, and next generations of HBMs, DDRs, and LPDDRs. Verification is moving from interface compliance toward validating complete memory subsystems, including controllers, PHYs, memory devices, modules, and system-level behavior. For more information on memory models, please refer to our Simulation VIP memory page.

For further information, reach out to talk_to_vip_expert@cadence.com.

This blog has been authored by Ujash Poshiya and Rahul Shiv Shankar.

Disclaimer: JEDEC standards are subject to change during and after the development process, including disapproval by the JEDEC Board of Directors.

© 2026 Cadence Design Systems, Inc. All Rights Reserved.

  • Terms of Use
  • Privacy
  • Cookie Policy
  • US Trademarks
  • Do Not Sell or Share My Personal Information