Skip to main content
Be Bitwise

DDR Evolution

Memory Technology · Hardware Foundations track

10 min read

In this module

From SDR to DDR5

The DRAM cell hasn't really changed in decades. Still one transistor, still one capacitor. What's changed dramatically is the interface between DRAM and the memory controller. Each generation of DDR (Double Data Rate) memory pushes bandwidth higher, drops power, and bolts on features, all on top of essentially the same underlying cell.

Tracing DDR's evolution explains why memory specs look the way they do, what those numbers on the DIMM sticker actually mean, and why DDR5 is a real architectural shift rather than another frequency bump.

The Prefetch Architecture

The trick behind DDR bandwidth scaling is prefetching: pulling more bits out of the DRAM array internally per clock cycle than you ship out over each external pin.

SDR SDRAM: 1-bit prefetch. One bit per pin per clock cycle. The internal array clock and the external I/O clock run at the same frequency.

DDR: 2-bit prefetch. The internal array runs at frequency F, but two bits are fetched per array access and transmitted on both the rising and falling edges of the external clock. This doubles the data rate without increasing the array speed.

DDR2: 4-bit prefetch. Four bits fetched internally per array access, requiring the I/O bus to run at 2x the array frequency and still transferring on both clock edges.

DDR3: 8-bit prefetch. Eight bits per internal access.

DDR4: 8-bit prefetch (same as DDR3), but with higher I/O frequencies.

DDR5: 16-bit prefetch. Sixteen bits per internal access, doubling the prefetch from DDR4. This is combined with a burst length of 16 (up from 8 in DDR4).

Pick a generation and a speed grade below and watch the array rate: it hovers between about 100 and 400 MHz from PC133 to DDR5-6400 while the pins climb nearly fifty-fold. Then compare CAS latency in nanoseconds across the generations, and switch to DDR5 to see why a 32-bit subchannel with BL16 lands on exactly one 64-byte cache line.

DDR prefetch calculator

DDR4 prefetch: 8n, as DDR3; bank groups let accesses to different groups overlap so the bus can run faster.

Data rate

3200 MT/s

0.313 ns per bit on each pin

I/O bus clock

1600 MHz

MT/s ÷ 2: data moves on both edges. Cycle time 0.625 ns

Array (core) rate

400 MHz

MT/s ÷ 8n prefetch

Burst

BL8 = 64 bytes

8 beats x 64 bits. Burst length 8, or a 4-beat burst chop

Peak bandwidth

25.6 GB/s

Per 64-bit DIMM: MT/s x 64 bits ÷ 8

CAS latency

13.75 ns

CL22 x 0.625 ns cycle

Each array access on a pin delivers 8 bits, so the array only has to keep up with 400 million accesses a second while the pins run at 3200 MT/s, and the I/O clock runs 4 times faster than the array.

Fastest listed grade of each generation

GradePins (MT/s)Array (MHz)Latency
PC133
133
133
CL3 = 22.5 ns
DDR-400
400
200
CL3 = 15 ns
DDR2-1066
1067
267
CL7 = 13.12 ns
DDR3-2133
2133
267
CL14 = 13.12 ns
DDR4-3200
3200
400
CL22 = 13.75 ns
DDR5-6400
6400
400
CL52 = 16.25 ns

Both bars share one scale. The pins got about 48 times faster across these generations; the array rate stayed between 133 and 400 MHz, and latency in nanoseconds barely moved.

Grades are common JEDEC bins; latencies are typical values for each grade, and real modules vary.

Why Not Just Increase Array Speed?

The DRAM core has barely sped up in decades: a DDR4-3200 module runs its array at 400 MHz and relies on 8n prefetch to fill 3200 MT/s on the pins. Charge sharing, sense amplification, and restore are pinned down by capacitor physics and transistor characteristics, and there's no clever way around them. Prefetch sidesteps the problem by going wider internally: read more bits in parallel from the slow array, then serialise them onto the fast I/O bus. Each DDR generation trades a wider internal bus for higher external bandwidth.

DDR3, DDR4, and DDR5 Compared

SpecificationDDR3DDR4DDR5
Supply voltage1.5V (1.35V for DDR3L)1.2V1.1V
Data rates800-2133 MT/s1600-3200 MT/s3200-8800+ MT/s
Prefetch8n8n16n
Burst length8816
Banks816 (4 bank groups x 4)32 (8 bank groups x 4)
Channels per DIMM1 x 64-bit1 x 64-bit2 x 32-bit
On-die ECCNoNoYes
Voltage regulatorOn motherboardOn motherboardOn DIMM (PMIC)
Max DIMM capacity16 GB (typical)64 GB (typical)256 GB (typical)

Voltage Scaling

Each generation steps the I/O signalling voltage down: 1.5V (DDR3), 1.2V (DDR4), 1.1V (DDR5). Lower voltage cuts power consumption quadratically (dynamic power scales with V squared) and unlocks higher frequencies by shrinking signal swing time. The catch: smaller voltage swings make the receiver's job harder, so signal integrity techniques have to get fancier with each generation.

Bank Group Architecture

DDR4 brought bank groups. Banks inside the same group share some internal circuitry and carry a minimum timing constraint between accesses (tCCD_L, the long column-to-column delay). Hop across to a different bank group and you can use the shorter delay, tCCD_S. So interleaving accesses across groups beats hammering one group over and over.

DDR5 doubles the bank count to 32 (8 groups of 4 banks each), giving modern multi-core workloads more headroom for running many independent memory streams in parallel.

DDR5 Innovations

DDR5 isn't just a frequency bump. There are real architectural changes underneath.

Dual Sub-Channels

The biggest DDR5 change is splitting each DIMM into two independent 32-bit sub-channels. Each sub-channel runs its own command/address bus, its own bank groups, and minds its own business.

That doubles the number of independent commands the controller can issue per unit time. For workloads sprinkled with many small, independent accesses (multi-threaded server applications, mostly), two 32-bit channels deliver substantially better effective bandwidth than one 64-bit channel, even though the raw data width is identical.

On-Die ECC

DDR5 ships with on-die error correction (ODECC) baked into the DRAM chip. Every internal 128-bit transfer carries extra parity bits, letting the DRAM catch and fix single-bit errors before the data ever leaves the chip.

That's a different beast from system-level ECC, which the memory controller computes using extra DRAM chips on the DIMM. On-die ECC lets DRAM manufacturers push to higher density and thinner process geometries, where individual cell reliability is shakier, by quietly papering over errors as they happen.

On-Die ECC Is Not a Replacement for System ECC

On-die ECC fixes errors inside the DRAM chip before anything reaches the controller. System-level ECC, on ECC DIMMs, protects data once it's on the bus and inside the controller. Different failure modes: on-die ECC handles internal cell failures and retention errors, while system ECC catches transmission errors, multi-bit chip failures, and gives you broader coverage. Server and enterprise systems still demand ECC DIMMs on top of on-die ECC.

Decision Feedback Equalisation (DFE)

Above 4800 MT/s, signal integrity becomes a real headache. Inter-symbol interference (ISI), where the voltage from one bit transition smears into the next, will corrupt your data. DDR5 receivers fight back with Decision Feedback Equalisation, a trick lifted from high-speed networking. DFE looks at the bits it has already received, predicts the ISI those bits will cause on the current bit, and subtracts it out. The signal gets cleaned up without amplifying noise along with it.

Power Management IC (PMIC)

DDR5 yanks the voltage regulator off the motherboard and bolts it onto the DIMM itself. The PMIC (Power Management Integrated Circuit) takes the input the motherboard hands over (5V on desktop and laptop DIMMs, 12V on server RDIMMs) and generates the precise 1.1V supply locally, right next to the DRAM chips. Less distance means less droop, finer-grained power management, and a simpler motherboard.

Memory Controllers

The memory controller is the bridge between the processor and DRAM. On modern chips it lives right on the CPU die. Intel moved it on-die with Nehalem, AMD with Opteron, retiring the old northbridge chip.

Controller Responsibilities

  • Command scheduling: converting processor read/write requests into the precise sequence of DRAM commands (ACT, RD, WR, PRE, REF) while obeying all timing constraints.
  • Request reordering: rearranging pending requests to maximise row buffer hits and bank-level parallelism.
  • Refresh management: interleaving refresh commands with normal traffic to minimise stalls.
  • Power management: placing idle ranks in low-power states (self-refresh, power-down mode) to save energy.
  • Error correction: computing and checking ECC on every access (for ECC-enabled systems).

Bank-Level Parallelism

The controller leans on the fact that banks operate independently. While one bank is activating a row (tRCD latency), another can be slinging data, and a third can be precharging. By interleaving requests across banks and bank groups, the controller hides each bank's latency behind parallel work, approaching peak bandwidth even though every bank in isolation is sluggish.

ECC Memory

Error-Correcting Code (ECC) memory adds extra DRAM chips to each DIMM to store parity information alongside data.

SECDED

The standard ECC scheme is SECDED: Single Error Correct, Double Error Detect. For every 64 bits of data, you store 8 extra bits of ECC (a Hamming code with an additional parity bit). That gets you:

  • Correction of any single-bit error in the 64-bit data word
  • Detection (no correction) of any two-bit error

SECDED catches the overwhelming majority of memory errors in practice. Single-bit errors, courtesy of cosmic ray strikes, alpha particle decay, and flaky cells, get silently corrected and you never know. Two-bit errors raise a machine check exception, handing the problem to the operating system, which typically kills the affected process or kicks off a restart.

Here's the same scheme shrunk to a 4-bit nibble so you can follow every bit: three Hamming parity bits plus one overall parity bit. The syndrome points straight at a single flipped bit; the overall parity is what tells one error apart from two.

Try itFlip three bits at once and see what SECDED does with an error it was never designed for.
C (GCC 9.2.0)

Chipkill and Advanced ECC

Server-class systems run sturdier schemes:

  • Chipkill (IBM) / SDDC (Intel): can correct the complete failure of an entire DRAM chip on a DIMM. Protects against chip-level manufacturing defects and outright hard failures.
  • Rank-level ECC: spreads ECC across a whole rank for even higher fault tolerance, paid for with more DRAM chips and a more elaborate controller.
Key Concept
Memory Reliability in Practice

Google's large-scale study of DRAM errors in data centres found that roughly one-third of machines experienced at least one correctable memory error per year, and the error rate rose with utilisation, module age and a module's own error history, while temperature mattered far less than expected. ECC memory is essential in servers not because errors are rare, but because they are common enough that undetected corruption would be a significant reliability problem at scale.

Channel Configuration and Bandwidth

A system's total memory bandwidth is determined by:

c
Bandwidth = data_rate x bus_width x channels

For a dual-channel DDR5-4800 system:

c
Bandwidth = 4800 MT/s x 8 bytes (64 bits) x 2 channels = 76.8 GB/s

The 64 bits per channel in that formula are the two 32-bit sub-channels of each DDR5 DIMM counted together. So a motherboard with two DIMM slots and one DDR5 DIMM per slot gives the memory controller four independent 32-bit sub-channels, but the total bandwidth comes out the same.

DIMMs aren't the only shape DRAM comes in. Phones and thin laptops use LPDDR (LPDDR5 and 5X), soldered beside or on top of the processor with narrow 16-bit channels and aggressive power-saving states. Graphics cards use GDDR (GDDR6 and 7), which trades capacity and power for very high per-pin rates. And HBM, which you'll meet in the exercise below, stacks DRAM dies on a silicon interposer next to the processor for a very wide bus. On servers, CXL memory expanders add another tier: DDR attached over a PCIe-based link, with more latency than local DIMMs.

Modern server processors (AMD EPYC, Intel Xeon) push 8 or 12 memory channels with multiple DIMMs per channel, hitting aggregate bandwidths of 400-600 GB/s per socket. That kind of bandwidth is what data-hungry workloads (machine learning, scientific computing, in-memory databases) actually live or die on.

Try a few with the formula before you open the answer: dual-channel DDR5-5600, a 12-channel server running DDR5-4800 with one DIMM per channel, and a single HBM3 stack (a 1024-bit interface at 6.4 Gb/s per pin).

Show the answer
  • DDR5-5600 dual channel: 5600 × 8 × 2 = 89.6 GB/s.
  • 12 channels of DDR5-4800: 4800 × 8 × 12 = 460.8 GB/s.
  • One HBM3 stack: 6.4 Gb/s × 1024 bits ÷ 8 = 819.2 GB/s, from one package beside the processor. That's why GPUs and AI accelerators put up with HBM's cost.

These are peak figures. A STREAM-style measurement on a real machine usually lands well short, because refresh, bank conflicts and read/write turnarounds all eat into the bus time.