A PC and a phone can both have 16 GB of memory, yet their specifications often say DDR and LPDDR respectively. Graphics cards use GDDR, while many AI accelerators use HBM. What accounts for these different names?

Capacity describes how much data the memory can hold. Understanding the memory family also requires looking at its connection to the processor, data throughput, power consumption and physical integration.

Four branches of DRAM #

DDR, LPDDR, GDDR and HBM all use DRAM storage cells that require refreshing. Each family has developed different interfaces, electrical requirements and system organizations.

DDR stands for Double Data Rate. In this article, it refers to general-purpose main-memory families such as DDR4 and DDR5, widely used in PCs and servers. LPDDR adds Low Power to the name, reflecting its emphasis on energy efficiency in phones, thin laptops, automotive systems and some server platforms.

The G in GDDR stands for Graphics. It serves GPUs and other processors with high-throughput memory traffic. HBM means High Bandwidth Memory, combining stacked DRAM with a wide data interface for AI acceleration and high-performance computing.

Four DRAM families and their typical roles
DDR, LPDDR, GDDR and HBM emphasize different system requirements. Their applications can overlap.
Four DRAM families and their typical roles
DDR, LPDDR, GDDR and HBM emphasize different system requirements. Their applications can overlap.

Each family also has generations, including DDR4 and DDR5, or GDDR6 and GDDR7. DDR can also describe double-edge data transfer. Connecting a memory device to a processor requires support for the specific interface in the controller, PHY and physical design.

Start with an x8 device to calculate bandwidth #

Capacity measures how much data memory holds. Bandwidth measures data transferred per second. Latency measures the waiting time for an access. Two 16 GB systems can have different bandwidths, and equal peak bandwidth does not determine access latency.

An x8 device has eight data I/O lines, typically labeled DQ0 through DQ7. During one read transfer, each line carries one bit, giving eight bits, or one byte, across the interface. Clock, command, address and power pins are excluded from this data width.

Eight DQ lines transfer one byte from an x8 DRAM device, yielding 4.8 GB/s at DDR5-4800
An x8 DDR5-4800 device transfers one byte per transfer, giving a theoretical bandwidth of 4.8 GB/s.
Eight DQ lines transfer one byte from an x8 DRAM device, yielding 4.8 GB/s at DDR5-4800
An x8 DDR5-4800 device transfers one byte per transfer, giving a theoretical bandwidth of 4.8 GB/s.

DDR5-4800 specifies 4,800 MT/s, or 4.8 billion transfers per second. At one byte per transfer, the result is 4.8 billion bytes per second, or 4.8 GB/s. The stated transfer rate already accounts for double-edge operation and can be used directly.

Theoretical bandwidth in GB/s = data width in bits × per-pin data rate in Gb/s ÷ 8.

These are decimal rate units. One Gb/s is 10⁹ bits per second, and one GB/s is 10⁹ bytes per second. Refresh, row operations, request queues and access patterns affect the bandwidth an application actually achieves.

DDR and LPDDR emphasize different main-memory requirements #

DDR commonly connects to PCs and servers through DIMMs, enabling expansion within the platform's supported configurations. A DDR5 DIMM has two independent 32-bit data subchannels, totaling 64 data bits when additional ECC check bits are excluded.

At DDR5-4800, one subchannel provides 32 × 4.8 ÷ 8 = 19.2 GB/s. Both subchannels operating together provide 38.4 GB/s per DIMM. Ranks on the same subchannel share its data lines. Adding ranks changes capacity and scheduling opportunities, while the interface peak remains governed by data width and rate.

LPDDR gives active and idle energy consumption greater emphasis while continuing to increase data rates. Phones need to process imaging, video and AI workloads within tight battery and thermal budgets, making this a common application for LPDDR.

For an example LPDDR5 interface with 32 independent data lines running at 6.4 Gb/s per pin, theoretical bandwidth is 25.6 GB/s. With 64 independent lines at the same rate, it is 51.2 GB/s. The LPDDR5 name and capacity alone leave the system bandwidth unspecified.

Physical installation is a separate consideration. LPDDR often uses soldered or stacked-package implementations, and replaceable modules such as LPCAMM2 also exist. Upgrade options and supported capacity depend on the platform and module design.

How GDDR and HBM feed a compute chip #

GDDR typically uses discrete packages arranged around a GPU and connected through PCB traces. High per-pin data rates and multiple parallel interfaces support the heavy memory traffic of graphics and compute workloads.

For example, a 32-bit GDDR6 device operating at 24 Gb/s provides 96 GB/s. Eight devices, each using an independent 32-bit connection, form a 256-bit interface delivering 768 GB/s. Device count, organization and operating rate depend on the graphics-card configuration.

GDDR connects through PCB traces, while stacked HBM connects to a compute die within a common package
Typical connections. The HBM example uses an interposer. Device counts and stack heights are illustrative.
GDDR connects through PCB traces, while stacked HBM connects to a compute die within a common package
Typical connections. The HBM example uses an interposer. Device counts and stack heights are illustrative.

HBM stacks multiple DRAM dies vertically and uses through-silicon vias, or TSVs, for electrical connections between layers. A typical system places the HBM stack and compute die within one package, using an interposer or another high-density interconnect structure.

Stacking adds capacity within a limited footprint. The wide interface supplies many parallel data paths. An example HBM3 stack with 1,024 data bits and a per-pin rate of 6.4 Gb/s provides 819.2 GB/s. Those 1,024 bits are distributed across multiple channels, with individual accesses governed by channel organization and access granularity.

The table substitutes data width and per-pin rate into the formula, with results in GB/s.

Calculation scopeFormulaBandwidth
x8 DDR5-4800 device8 × 4.8 ÷ 84.8
DDR5-4800 subchannel32 × 4.8 ÷ 819.2
Complete DDR5-4800 DIMM64 × 4.8 ÷ 838.4
Example LPDDR5 interface32 × 6.4 ÷ 825.6
Example GDDR6 device32 × 24 ÷ 896
Example HBM3 stack1,024 × 6.4 ÷ 8819.2

The rows describe different scopes, from individual devices to subchannels, modules and stacks. For a system comparison, first establish the processor's complete memory interface, then consider sustained bandwidth, capacity and latency. HBM integration also requires coordinated package, power and thermal design, with its configuration established during platform development.