TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

SDRAM Memory Systems: Architecture Overview and Design Verification

mg
by
0
Share Facebook X LinkedIn WhatsApp Email

DRAM (dynamic random access memory) is attractive to designers because it offers a broad range of performance and is used in a wide variety of memory system designs, from computers to embedded systems. In 2010 we pointed readers to a TechOnline DRAM primer that covered DRAM concepts, likely future developments and how verification improves memory designs. That document is no longer online, so this post now covers the same ground itself, using a DDR3 datasheet for the numbers.

How a DRAM stores a bit

Each DRAM bit is a tiny capacitor with one access transistor. A charged capacitor is a 1, an empty one is a 0. Because the cell is so small, DRAM packs far more bits into a given area than SRAM, which needs six transistors per bit. The price is that the charge leaks away, so every cell must be read and rewritten, or refreshed, many times a second. That is what "dynamic" means.

How the memory is organized

Cells sit in a grid of rows and columns, and a device holds several independent grids called banks. To read data, the controller first opens a row in a bank (the ACTIVATE command), which copies the whole row into sense amplifiers. It then reads or writes columns from that open row. Before it can open a different row in the same bank, it must close the current one (PRECHARGE). Accesses to an already open row are fast; accesses that need a new row are slower. Good memory controllers reorder requests to take advantage of open rows and to spread work across banks.

What "synchronous" and "DDR" add

SDRAM (synchronous DRAM) runs from a clock, so commands and data line up with clock edges and the device can be pipelined. DDR (double data rate) SDRAM moves data on both the rising and falling edge of the clock, doubling the transfer rate for a given clock. Each DDR generation since has raised the data rate, lowered the supply voltage and added features that help keep high-speed signals clean.

Why verification matters

A DRAM interface only works if the controller, the board and the memory all meet the timing and electrical rules in the JEDEC standard and the device datasheet. As data rates rise and voltage swings shrink, small layout or timing errors can cause intermittent failures that are hard to reproduce. Design verification therefore covers two layers: protocol and timing (are the commands legal and correctly spaced?) and signal integrity (are the eyes open, is the strobe correctly aligned to the data?).

Deep dive: the numbers and details

This part uses a real DDR3 device to put numbers on the concepts, with worked examples for refresh overhead, latency and bandwidth. Skip it if the overview is all you need.

A reference device: Micron 1 Gb DDR3

All figures in this section come from Micron's 1Gb x4, x8, x16 DDR3 SDRAM datasheet (Rev. L, March 2013), a copy hosted for a university course.

  • Supply: VDD = VDDQ = 1.5 V plus or minus 0.075 V.
  • 8n-bit prefetch architecture and 8 internal banks.
  • Fixed burst length of 8, with a burst chop of 4 selectable in the mode register or on the fly.
  • Differential clock and differential bidirectional data strobe.
  • Nominal and dynamic on-die termination (ODT), write leveling and output driver calibration, all features aimed at signal integrity.
  • Refresh: 64 ms with 8,192 refresh cycles from 0 to 85 °C; 32 ms from 85 to 95 °C.
  • Page size: 1 KB for the x4 and x8 parts, 2 KB for the x16 part.

Speed grades and core timing

Speed gradeData rate (MT/s)tRCD-tRP-CL (clocks)tRCD = tRP = CL (ns)
-187E10667-7-713.1
-15E13339-9-913.5
-125160011-11-1113.75
-107186613-13-1313.91

The pattern is the key insight: the clock count rises with each speed grade, but the time in nanoseconds barely moves. The DRAM core, the cell array and sense amplifiers, does not get much faster; the interface does.

Worked example 1: refresh overhead

The datasheet gives the average refresh interval tREFI = 64 ms / 8,192 = 7.8 µs, and a REFRESH command period tRFC of at least 110 ns for the 1 Gb device. While a refresh runs, the device cannot serve reads or writes, so the fraction of time lost is about 110 ns / 7,800 ns = 1.4%. For the 8 Gb density the same datasheet lists tRFC = 350 ns, which would be 350 / 7,800 = 4.5%. Above 85 °C, tREFI halves to 3.9 µs and these overheads double. This is why refresh cost grows with density and temperature.

Worked example 2: latency for a random read

For the DDR3-1600 grade (-125, tCK = 1.25 ns), a read to a closed row in an idle bank needs ACTIVATE, then a READ after tRCD, then data after CL:

  • tRCD + CL = 13.75 ns + 13.75 ns = 27.5 ns to the first data.
  • If a different row is open in that bank, add tRP = 13.75 ns to close it first: 41.25 ns.
  • The minimum time between ACTIVATE commands to the same bank, tRC, is 48.75 ns for this grade.

Worked example 3: peak bandwidth

A x16 DDR3-1600 device moves 16 bits on each of 1,600 million transfers per second: 1,600 x 10^6 x 2 bytes = 3.2 GB/s. Four such devices on a 64-bit channel give 12.8 GB/s. One BL8 burst on the 64-bit bus delivers 64 bytes, which is a typical processor cache line, in four clock cycles (5 ns).

Comparing examples 2 and 3 shows why bank interleaving matters: one random access takes tens of nanoseconds, but a burst only 5 ns, so the controller must keep several banks busy to approach peak bandwidth.

What verification checks

  • Command timing: every spacing rule (tRCD, tRP, tRAS, tRC, tRFC, tREFI and many more) is respected by the controller under all traffic patterns. Protocol checkers in simulation and on a logic analyzer catch violations.
  • Initialization and training: mode registers, ODT settings, write leveling and read training complete correctly at each supported speed.
  • Signal integrity: data eyes at the DRAM and the controller meet setup and hold windows, with margin for crosstalk, reflections and power supply noise.
  • Corners: refresh rate and timing are checked at temperature extremes, since the datasheet doubles the refresh rate above 85 °C.
Share Facebook X LinkedIn WhatsApp Email

Post a Comment

0 Comments

Your comments will be moderated before it appears here.

Post a Comment (0)

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!