Home Up

Home Teaching Glossary ARM Processors Supplements Prof issues About

DRAM

DRAM, dynamic random access memory has provided bulk of high-speed random access memory in mainstream computers since the late 1970s. The access time of DRAM makes it well-suited to use as main store, although it is slower than processors which means that cache memory is required to bridge the gap between processor and DRAM performance.

DRAM is read/write random access memory and requires a more complex processor interface than static random access memory. Very simple systems (e.g., controllers and embedded systems)) may use static memory to avoid the cost and complexity of DRAM controller. DRAM stores data as a charge on a capacitor which leaks away in several milliseconds. Consequently, dynamic memory must be continually refreshed to re-write data before it is lost. This operation is invisible to the user.

DRAM History

The world’s first DRAM memory device was created by Robert H. Dennard, an IBM Fellow in 1966 when he and his team were investigating the properties of field-effect transistors. In 1968 Dennard was awarded a patent for a single-transistor DRAM cell.

The first commercial DRAM chip was developed by a new semiconductor company that had been formed two years earlier by Noyce and Moore who left Fairchild Semiconductors. The new company was called Intel and it initially developed a static RAM using bipolar logic and then a static RAM using MOS logic in 1969. In 1970 Intel announced a 1 Kbit (that’s 256 bytes) DRAM memory component called the 1103. This was the world’s first commercial DRAM component and by 1972 it was the best selling memory chip in the world and was rapidly replacing existing magnetic memory systems (ferrite cores).

Another step along DRAM’s path towards domination of the memory market was the introduction of Mostek’s MK4096 DRAM in 1973. Mostek was founded in 1969 by former employees of Texas Instruments. The MK4096 was a 4 Kbit chip that introduced address multiplexing. Static RAMs, both then and now, use a full n-bit address bus to access 2n one of words in the memory array. Mostek’s innovation was to multiplex the address bus into a row and column address and then feed row and column addresses to the DRAM sequentially. The memory array was treated as a square; first the row address was entered and captured to select a row, and then the column address was entered to select a column. The cell (or word) at the row and column intersection is then accessed. Multiplexing an address between row and column halves the number of address pins at the cost of speed and complexity. By multiplexing the address the MK4096 was able to fit in a 16-pin package which was both less bulky and less expensive than the 22-pin package required by a non-multiplexed DRAM of the same size. By 1976 Mostek was producing a 16 Kb DRAM with an improved silicon technology.

Over the years, DRAM has evolved. However, DRAM is still essentially the same as it always was; it’s a memory cell that stores a single bit as a charge on a leaky capacitor. Some aspects have changed such as the bit density per chip, the way in which interfacing and refreshing is carried out, and the way in which DRAM is accessed. The access time of DRAM has improved over the years, but the speedup of DRAM over decades is woefully small compared with the way in which processors have advanced.

The term Memory Wall has been coined by Bill Wulf and Sally McKee in 1994 to describe the mismatch between processor performance and memory performance; for example, average memory access times were projected to be 2 cycles in 2000 and 100 cycles in 2010. This represents a factor of 50 in the relative reduction of memory performance. In practice, the situation did not turn out to be quite so bad. Improvements in cache design and performance have helped to hide DRAM’s limitations, and the Memory Wall puts too much emphasis on access time (latency) rather than throughput (bandwidth). Although access time has not increased as much as we would like, the way in which data is moved into and out of DRAMs has permitted a much higher bandwidth. The river doesn’t flow any faster, but at least its wider..

Although the volume of DRAM production has been increasing steadily, the number of DRAM manufacturers has been in decline for the past 20 years. DRAM production is immensely expensive (the cost of fabrication plants) and very competitive. To a great extent, production has moved from West to East and most of the volume manufacturers are now in Asia. In the 1980s major DRAM producers were:

AMD          Eurotechni      Hitachi         Intersil

AMI           Fairchild         IBM               Samsung

AT&T         Fuijitsu            Infineon      Winbond

Elpida       Hynix               Intel              Zilog

By 2009 the major DRAM manufactures were

Elpida      PowerChip

Hynix       Promos

Micron    Samsung

Nanya     Winbond

The DRAM Memory Cell

Let’s look at a DRAM memory cell. The classic crossed-gate static memory cell uses six transistors. One version of the DRAM cell uses three transistors providing twice the bit density per chip. The figure belowillustrates a three transistor cell. Two-bit lines are used to access the cell, Bitline 1 and BitLine 2 (one is used in writing and one in reading the cell).

A write operation is performed by asserting the WriteLine and BitLine 1 simultaneously to charge or discharge the capacitor. A charge is transferred to or from the capacitor via transistor T1.

A read operation is performed by sensing the voltage on the capacitor via transistors T2 and T3. The read is carried out by asserting the ReadLine and then sensing the voltage on BitLine2.

Essentially, this three-transistor circuit consists of a one transistor switch to charge/discharge the capacitor and a two-transistor sense amplifier to detect its state. The read operation is non-destructive; that is, the voltage ion the capacitor is not modified by a read. The output of the read circuit is inverted so that the data must be inverted again to read its corrector value.

The three-transistor DRAM cell is very attractive and relatively easy to construct. Unfortunately, it is only half the size of a six-transistor static memory cell and, therefore, does not achieve the maximum efficiency in terms of the size of a cell. Three-transistor DRAM cells were used in first generation DRAMs.

The following figure illustrates the one-transistor memory cell. This is, of course, six times more space-efficient than the static memory cell.

As in the case of the three-transistor cell, the BitLine and WriteLine can be asserted to write data into the cell by charging or discharging the cell’s capacitor Cs. Reading the cell is slightly more complicated.

When transistor T1 conducts, the storage capacitor Cs is connected to the bit line. You might think that the voltage on the bit line is the same as that originally stored on the capacitor. However, the bit line has its own capacitance to ground that is distributed along the bit line and its connections. Consequently, the two capacitors CB and CS form a voltage divisor. The voltage that is sensed is proportional to CS/(CS + CB).

Reading the data from a one-transistor cell destroys the data and the original data has to be rewritten in a refreshing operation.

MORE TO FOLLOW