Hybrid Memory Cube

Hybrid Memory Cube (HMC) is a 3D-stacked DRAM architecture that uses through-silicon vias (TSVs) and a high-speed serialized interface to deliver dramatically higher bandwidth and energy efficiency than conventional DDR memory — developed by Micron and the Hybrid Memory Cube Consortium, HMC pioneered the concept of intelligent memory with a logic base die that manages memory access, error correction, and protocol conversion, influencing the design of HBM and CXL-attached memory while targeting networking, high-performance computing, and data-intensive applications.

HMC: Vault-Partitioned Stack, Serial Links Instead of a Wide Bus every vault owns a private TSV column through all DRAM tiers; the logic base die exposes only narrow SerDes links Cross-Section — 4 Vaults, 4 DRAM Tiers tier 4 tier 3 tier 2 tier 1 vault 0 vault 1 vault 2 vault 3 Logic Base Die vault ctrl vault ctrl crossbar SerDes Each vault is independent end-to-end — its own TSV bus and controller — so 16–32 vaults run truly parallel memory operations at once. Why serial, not parallel Short internal TSVs carry the wide bus; the external link only needs a few fast serial lanes. Serial Link Fan-Out LOGIC DIE link 0 · 16 lanes link 1 · 16 lanes link 2 · 16 lanes link 3 · 16 lanes Host / Switch PCB-trace reach — not interposer-bound up to 30 Gbps/lane 320 GB/s (Gen2) ~3.7 pJ/bit 70% below DDR3 What this trades away vs. HBM HBM: 1024-bit parallel bus, short interposer reach, higher peak bandwidth (819 GB/s, HBM3) per stack. HMC: narrow serial links, long PCB reach, lower peak bandwidth, but placeable anywhere on board. HBM's simpler interposer integration won the AI/HPC market; HMC is now discontinued.

What Is HMC?

  • Definition: A 3D-stacked DRAM technology where 4-8 DRAM dies are vertically stacked on a logic base die using TSVs, with the logic die providing a high-speed serialized interface (up to 30 Gbps per lane) rather than the wide parallel interface used by DDR or HBM — enabling long-reach, high-bandwidth memory connections over PCB traces.
  • Serialized Interface: Unlike HBM's 1024-bit parallel interface that requires an interposer, HMC uses narrow, high-speed serial links (16 lanes per link, up to 4 links per device) — allowing HMC to be placed anywhere on a PCB, not just adjacent to the processor.
  • Vault Architecture: HMC organizes memory into 16-32 independent "vaults," each spanning all DRAM layers with its own TSV bus and vault controller in the logic die — enabling massive internal parallelism with 16-32 simultaneous memory operations.
  • Logic Base Die: The bottom die in the HMC stack is a logic chip (not DRAM) that contains memory controllers, SerDes transceivers, crossbar switch, error correction, and power management — making HMC a "smart memory" that offloads protocol handling from the host processor.

Why HMC Matters

  • Bandwidth Revolution: HMC Gen2 delivered 320 GB/s per device — 15× the bandwidth of DDR3 and 8× DDR4 at the time of introduction, demonstrating that 3D stacking could fundamentally change the memory bandwidth equation.
  • Energy Efficiency: HMC achieved ~3.7 pJ/bit — 70% lower energy per bit than DDR3, primarily because the short TSV connections within the stack consume far less energy than driving signals across long PCB traces.
  • Architecture Influence: HMC's vault architecture and logic base die concept directly influenced HBM's channel architecture and Samsung's Processing-in-Memory (PIM) designs — the idea of putting intelligence at the memory became a major research direction.
  • Network Memory: HMC's serialized interface enabled memory to be placed at the end of a high-speed link rather than directly adjacent to the processor — a concept that evolved into CXL-attached memory and memory pooling architectures.

HMC Specifications

ParameterHMC Gen1HMC Gen2
Capacity2-4 GB4-8 GB
Bandwidth160 GB/s320 GB/s
Links4 (16 lanes each)4 (16 lanes each)
Lane Speed10-15 Gbps28-30 Gbps
Vaults1632
Stack Height4-8 DRAM dies + logic4-8 DRAM dies + logic
Power~11W~11W
Energy/bit~5 pJ/bit~3.7 pJ/bit

HMC vs. HBM vs. DDR

FeatureHMCHBMDDR5
InterfaceSerial (30 Gbps/lane)Parallel (1024-bit)Parallel (64-bit)
PlacementAnywhere on PCBOn interposer (adjacent)DIMM slot
BW/Device320 GB/s819 GB/s (HBM3)51.2 GB/s
IntelligenceLogic base dieMinimal logicNone
ReachLong (PCB traces)Short (interposer)Medium (DIMM)
MarketNiche (networking)Mainstream (AI/HPC)Mainstream (general)
StatusDiscontinuedActive developmentActive development

HMC is the visionary 3D memory architecture that proved intelligent stacked memory was possible — pioneering the vault architecture, logic base die, and serialized memory interface concepts that influenced HBM, CXL-attached memory, and processing-in-memory designs, even though HBM's simpler integration with GPU interposers ultimately captured the high-bandwidth memory market.

Take hybrid memory cube further

Ask the copilot about this term, or have our engineers assess it against your process.