RAM Details

Memory

We try to avoid giving history lectures in this website, but on the subject of memory, a brief review is essential to understanding the terminology. One of the basic innovations that made digital computers possible, Random Access Memory (RAM) allows the CPU to retrieve information stored at a specific memory address without having to read through all the memory to find it. Contrast this with a tape drive, where the whole tape may have to be wound by the read head to reach the location of the desired information. Even with the relatively fast hard drive, the read head must physically move, as much as tens of millimeters, and wait for the disk to spin until the information is under the head.

The access times of drives are measure in milliseconds, or thousandths of a second; the access time of RAM is measured in nanoseconds, or billionths of a second. When it comes to locating a single byte of data at a random location, memory outperforms other storage media by a factor of hundreds of thousands. If this wasn't the case, the super-fast CPUs would have no purpose, because they would spend all their time waiting for new instructions and data to work on.

There are two basic types of RAM in use: Static RAM (SRAM) and Dynamic RAM (DRAM). Both types of memory forget everything if the power is tuned off, but SRAM doesn't require the constant refreshing the DRAM does, ergo the names, "Static" and "Dynamic." SPRAM requires four or five times as many transistors to implement as DRAM, as it actually traps each bit of information in a structure called a flip-flop. DRAM stores a bit as a temporary charge on the leg of a single transistor, but this decays away so rapidly that it must be reread and refreshed many times per second. SRAM is used as cache memory on CPUs and in other applications, but always in relatively small amounts because of its increased power and real estate demands. DRAM is used for the main PC memory and has been since the original IBM PC was introduced 20 years ago.


Fast Page Memory (FPM) was the first big performance enhancement to DRAM, which had previously treated each new memory transaction like a surprise invitation. FPM made it faster to access data in the same memory "page" although the term "row" offers a better representation of what really goes on. When a new data bit is to come from the same matrix row as the previous bit, the memory controller needs only to increment the column location and the same row address will be used, saving an address transaction.

Extended Data Out (EDO) DRAM shortens the recovery time between sequential DRAM reads, offering about a 20 percent performance boost in overall memory throughput. EDO was backward compatible, meaning it would function in systems that were designed to support FPM RAM, albeit without any performance increase. Burst EDO (BEDO) was the next level of enhancement in which a series, or burst, of bytes from memory could be transferred to the CPU in a single request. If the CPU actually required data from these subsequent locations, an operation has been saved, and if not, nothing has been lost.

Synchronous DRAM (SDRAM) can really boost memory bandwidth through synchronization with the system clock. This eliminates a large number of timing delays, which can result in wait states on the part of the CPU (i.e., idle time). The motherboard must be designed to support SDRAM, which is not backward compatible to EDO or FPM. Early SDRAM modules were 5V devices, but the current modules require 3.3V. Fortunately, memory module designers got together earl on and came up with a standard system of notches in the contact edge of memory modules, which prevents them from being installed in the wrong type of memory socket. SDRAM was originally available at 66 MHz, but 100 MHz (PC100) and 133 MHz (PC133) devices so followed. The current speed champion for SDRAM is PC150 (150MHz) and is popular with overclockers.

Double Data Rate (DDR) SDRAM, simply known as DDR, is the next step after SDRAM. DDR can effectively double the throughput of earlier SDRAM by transferring data on both the rising and falling edges of the bus clock. The Athlon was the first CPU to take advantage of DDR, which it currently supports with a 266 MHz FSB, and the Pentium 4 is expected to follow suit. The Duron supports DDR at 200 MHz. DDR modules are known both by their MHz rating and PC nomenclature similar to the earlier SDRAM modules. 200 MHz DDR is PC1600,266 MHz DDR is PC2100, and 300 MHz DDR is PC2400. The motherboard must explicitly support DDR for it to be used.

256 MB DDR DIMM

The RAM used in the IBM PC came in the form of one bit-wide chips in Dual Inline packages ,or Dip Chips. As memory chips shrank and capacity grew, they were mounted on small circuit boards called Single Inline Memory Modules (SIMMs), first 8 bits wide (1byte) and then 32 bits wide (4 bytes). As chips continued to shrink and capacity and memory buses continued to grow, the SIMM width was doubled, giving us DIMM. The current DIMM modules are 64 bits wide, the same as the memory bus in PC systems not using RDRAM, so they can be installed singly or in any multiple.


RAMBUS or RDRAM, technology represents a departure from the step-by-step evolution of RAM we have presented to this point. RDRAM Inline Memory Modules (RIMMs) are only 16 bits wide, but they make up in speed what they lack in width. Dual-channel motherboards operate RIMMs in pairs, allowing an effective width of 32 bits at speeds of up to 800 MHz. Not surprisingly, RDRAM modules run awfully hot, and they currently cost about twice as much as DDR modules of the same capacity.

RIMMs must be installed in banks of two, and unused banks must be filled with special dummies called Continuity RIMMs (CRIMMS) for electrical signal continuity. The 800 MHz RIMM is designated PC800; there are also slower versions designated PC700 and PC600, although they are primarily used in mass-market systems.

Both DDR and RIMM modules are available with the Error Correction Code (ECC) enhancement. ECC memory can correct single-bit errors on the fly and catch multiple-bit errors, unlike the earlier parity error checking, which couldn't correct any errors or identify two-bit flips. ECC memory must be supported by the motherboard, unless it used ECC onboard technology; and the price premium for ECC memory has almost disappeared with plummeting memory price.