PC memory may use the following access methods:
Asynchronous DRAM , which was used in all PCs until the late 1990s, uses a window of fixed minimum duration to determine when operations may occur. If the CPU has transferred data while a window is open, and if a subsequent clock cycle occurs while that window remains open, the CPU cannot transfer additional data until the next window opens, thereby wasting that clock cycle. Asynchronous operation forces the CPU to conform to a fixed schedule for transferring data, rather than doing so whenever it wishes. Asynchronous DRAM is available in the following types:
FPM was commonly used on 486 and earlier systems, and may be installed in early Pentium systems. FPM is not supported by recent chipsets. Although you can migrate FPM DRAM from an old Socket 5 or Socket 7 system to a newer Socket 7 system, it is good for little else. You may be able to install surplus FPM DRAM in your laser printer.
EDO, also sometimes called Hyper Page Mode DRAM, is marginally faster than FPM, is still available in all common package types, and was commonly installed on new systems until late 1998. EDO DRAM now costs so much that it is often less expensive to replace the existing motherboard, processor, and memory with current products than to buy EDO memory. That said, you may be able to upgrade the memory in an EDO-based system economically. Many EDO-based systems can also use SDRAM DIMMs, which are faster, much less expensive, and much more readily available. To upgrade the memory in such systems, replace the existing EDO memory with compatible SDRAM DIMMs.
BEDO slightly improved upon EDO, but was inferior to SDRAM, which was introduced at about the same time, and so never became popular. If you have a BEDO-based system, follow the upgrade advice given for an EDO-based system.
All forms of asynchronous DRAM are now obsolete. Although asynchronous DRAM is still available, it costs so much per megabyte that it never makes sense to buy it. For example, in July 2003, the price per megabyte of asynchronous DRAM SIMMs was five to 25 times that of SDRAM DIMMs, depending on capacity and type.
Synchronous DRAM , also called SDRAM, shares a common clock reference with the CPU. No window is needed because the CPU and memory are slaved together, allowing the CPU to transfer data to and from memory whenever it wishes to do so, instead of requiring the CPU to await an arbitrary window. For links to formal SDRAM standards, see the Intel SDRAM Specifications page at http://developer.intel.com/technology/memory/pcsdram/spec/index.htm.
SDRAM takes one of the following forms:
Ordinary SDRAM, sometimes called JEDEC SDRAM or PC66 SDRAM to differentiate it from PC100 SDRAM and PC133 SDRAM. PC66 SDRAM was formerly less expensive than PC100 or PC133 SDRAM, but as the price of those faster variants declined to near that of PC66 SDRAM, the demand for PC66 SDRAM plummeted. PC66 SDRAM is now hard to find and may cost more than PC100 or PC133 SDRAM. Because PC133 SDRAM can be used on nearly any system running a 133 MHz or slower FSB, buying JEDEC SDRAM never makes sense, even for systems that run memory at 66 MHz. PC66 SDRAM salvaged from an older system can be used in any system that runs a 66 MHz FSB, including those running older-model Celeron or Pentium II processors.
SDRAM that complies with the Intel PC100 specification, and is rated for use on a 100 MHz FSB. Like PC66 SDRAM, PC100 SDRAM is obsolescent, and may actually cost more than faster PC133 SDRAM.
SDRAM that complies with the Intel PC133 specification, and is rated for use on a 133 MHz FSB. PC133 SDRAM costs little or no more than PC100 SDRAM, operates properly?although only at the lower speed?in nearly all systems designed to use PC66 or PC100 SDRAM, and is usually the best choice when you're buying SDRAM memory, even for a 66 MHz or 100 MHz FSB system. PC133 SDRAM is commonly available in two variants which vary only in CAS latency. CAS-3 PC133 SDRAM is the more common form, and is what you will receive if you do not specify otherwise. CAS-2 PC133 SDRAM has lower latency, is therefore slightly faster in a motherboard that can take advantage of it, and costs only a few cents more per megabyte.
PC133 SDRAM is obsolescent. Even entry-level systems now use some variant of DDR-SDRAM, described in the next section, so PC133 SDRAM is now useful only to upgrade older systems. In those systems, PC133 SDRAM provides memory throughput that is well matched to the bandwidth of the processor.
Double Data Rate SDRAM (DDR-SDRAM) doubles the amount of data transferred per clock cycle, and thereby effectively doubles peak memory bandwidth. DDR-SDRAM is an evolutionary improvement of standard SDRAM, which is now sometimes called Single Data Rate SDRAM or SDR-SDRAM to differentiate it. Because DDR-SDRAM costs essentially the same to produce as SDR-SDRAM, it sells for about the same price.
The chips used to produce DDR-SDRAM DIMMs are named for their operating speed. For example, 100 MHz chips are double-pumped to 200 MHz, and so are called PC200 (or DDR200) chips. (We use the "PCxxx" speed nomenclature rather than the "DDRxxx" nomenclature for first generation DDR memory, although "DDRxxx" is also commonly used.) Similarly, chips that operate at 133 MHz are called PC266 chips, those that operate at 166 MHz are called PC333 chips, and those that operate at 200 MHz are called PC400 chips. (In fact, only PC200 and PC266 are formal standards, although memory makers produce so-called PC333 and PC400 chips based on de facto standards.)
Unlike SDR-SDRAM DIMMs, which are designated by their chip speeds, DDR-SDRAM DIMMs are designated by their throughput. Their data path is 64 bits (8 bytes) wide. So, for example, a DDR-SDRAM DIMM that uses PC200 chips transfers 8 bytes 200 million times per second, for a total throughput of 1,600 million bytes/second and is called a PC1600 DIMM. Similarly, DDR-SDRAM DIMMs that use PC266 chips are labeled PC2100, and those that use PC333 chips are labeled PC2700. Based on chip speed, PC2700 modules are actually PC2667, but no one uses that name. Some early DIMMs built with PC333 chips were labeled PC2600, but all memory makers now use the PC2700 designation for obvious reasons. DDR-SDRAM DIMMs that use PC400 chips are labeled PC3200.
PC2700 DDR-SDRAM is now the dominant mainstream memory technology, although the April 2003 introduction of the Intel 875P chipset has kickstarted the market for PC3200 memory. All mainstream AMD and Intel processors and chipsets now support PC2700 or faster DDR-SDRAM, and there is little reason to choose any slower type of memory. PC1600 DDR-SDRAM was economically obsoleted by mid-2002, when the price of PC2100 memory fell to the same level as PC1600. By early 2003, the same fate befell PC2100 memory, as the price of PC2700 memory fell to PC2100 levels.
By early 2003, the original DDR-SDRAM technology was fast approaching its limits. As AMD and Intel transition to higher FSB speeds, DDR-SDRAM will be hard-pressed to keep pace. Current DDR technology tops out at PC3200. Dual-channel DDR chipsets using PC3200 memory limit peak throughput to 6400 MB/s. That is sufficient for now, but as FSB speeds increase from 400 MHz to 533 MHz, 800 MHz, and beyond, even dual-channel DDR-SDRAM will be challenged to keep up with increases in processor bandwidth.
The long-term solution is DDR-II SDRAM. DDR-II incorporates a series of evolutionary improvements on DDR-I technology, including increased performance and bandwidth, reduced cost, lower power consumption, and improved packaging. DDR-II basic device timing and page size are compatible with DDR-I, and because the DDR-II command set is a superset of the DDR-I command set, a DDR-II controller can also control DDR-I modules.
DDR-II DIMMs use a new 232-pin connector, and it is likely that DDR-I modules will be produced with that connector to facilitate the transition from DDR-I to DDR-II modules. DDR-II chips will initially ship in DDR400 and DDR533 variants, which will be used to produce PC3200 and PC4300 DDR-II DIMMs. We expect that DDR-II will eventually be produced in DDR600, DDR667, and DDR800 variants, which will be used to produce PC4800, PC5300, and PC6400 modules, respectively.
Although some high-performance video cards currently use DDR-II, we do not expect DDR-II to become standard in desktop PCs until late 2004 or 2005. We think PC2700 DDR-I modules in single- and dual-channel configurations will remain the standard until DDR-II chipsets ship in volume in 2004. DDR-II will at first be used in high-end systems, and will migrate to midrange and entry-level systems throughout 2004 and 2005.
On the theory that if twice as fast is good, four times as fast must be better, the QBM Alliance is developing Quad Band Memory (QBM), sometimes called Quad Data Rate SDRAM (QDR-SDRAM). The QBM Alliance roster boasts many second- and third-tier companies, including Acer Laboratories, Acuid, Avant Technology, CST, Denali Software, Integrated Circuit Systems, Kentron Technologies, Netlist, Peripheral Enhancements, PNY Technologies, SiS, SiSoft, ST Microelectronics, Terarecon, and VIA Technologies. Alas, the key first-tier chipset/motherboard companies?notably Intel and AMD?are not QBM Alliance members. Neither have major memory companies such as Crucial/Micron, Kingston, or Samsung chosen to join the QBM Alliance. Without real support from those companies, we don't think the QBM Alliance can hope to establish a viable standard.
QBM is based on DDR-I technology, but quad-pumps rather than double-pumps the data channel. Although we have been wrong before, we expect QBM to fail for both technical and marketing reasons. Technically, QBM offers little real advantage over dual-channel PC2700 or PC3200 DDR-SDRAM, which is already widely supported by chipsets for Intel and AMD processors. That means memory makers have no incentive to produce yet another type of module that would sell in relatively small numbers, making it difficult to recoup their startup costs. From a marketing standpoint, QBM is almost doomed from the start. At best, QBM will garner support from second- and third-tier companies such as VIA Technologies and Kentron, which means that QBM will be perceived by consumers as a second-rate solution. Meanwhile, Intel and AMD will continue backing DDR-I and DDR-II, leaving only scraps for QBM.
SDRAM uses separate address, control, and data busses, each with many lines. Managing these wide parallel busses limits performance. Protocol-based DRAM instead uses a narrow, very fast channel with protocols that manage address, control, and data information. Rambus RDRAM, a proprietary RAM standard developed jointly by Intel and Rambus, is the sole surviving type of protocol-based RAM.
There are three types of Rambus memory, called Base Rambus, Concurrent Rambus, and Direct Rambus. The first two are obsolescent, and are used only in devices such as game consoles. All Rambus memory used in PCs is Direct Rambus memory. RDRAM is available in four speeds, designated PC600, PC700, PC800, and PC1066, although only PC800 and PC1066 are used in current systems. As with DDR-SDRAM, RDRAM modules are named according to their throughput, but with a difference. RDRAM uses a 16-bit or 18-bit data path (versus 64-bit for SDRAM) to transfer two bytes at a time. Accordingly, PC600 RDRAM provides peak throughput of 1200 million bytes/second, PC700 1400 million bytes/second, PC800 1600 million bytes/second, and PC1066 2133 million bytes/second.
In theory, then, it appears that PC800 RDRAM matches the throughput of PC1600 DDR-SDRAM, and PC1066 RDRAM the throughput of PC2100 DDR-SDRAM. In fact, that is true only if you are considering peak throughput. In the real world, RDRAM provides higher sustained throughput because it is more efficient than SDRAM in typical applications. Whereas SDRAM efficiencies are in the 40% to 70% range, RDRAM efficiency is about 80%. Accordingly, PC800 RDRAM might deliver sustained throughput of 1280 million bytes/second, whereas PC1600 DDR-SDRAM delivers much less, and even PC2100 DDR-SDRAM may not be able to match the actual sustained throughput of PC800 RDRAM.
On that basis, RDRAM might seem the better choice, but that is seldom true for several reasons. First, the throughput advantage of RDRAM is unrealized in most applications. Although modern processors such as the Pentium 4 can in theory use very wide memory bandwidths, in practice few applications require more memory bandwidth than PC1600 DDR-SDRAM provides, let alone PC2100, PC2700, or PC3200 DDR-SDRAM. Second, RDRAM typically costs significantly more than DDR-SDRAM. Third, throughput is only one aspect of memory performance. At least as important as throughput is latency?the time that elapses from requesting data from memory until the memory begins delivering that data. Despite the arguments of Rambus to the contrary, real-world RDRAM implementations exhibit high latency. What's worse is that RDRAM latency is cumulative. That is, with SDRAM, latency is a property of the memory chips themselves and remains the same regardless of the number of DIMMs installed in the system. With RDRAM, installing additional memory modules increases latency linearly. Not surprisingly, all the memory performance comparisons that we have seen from Rambus are based on using one RDRAM module per channel.
One can argue theory all day, of course, but the simple fact is that in our experience Rambus RDRAM memory seldom provides a significant performance advantage over SDRAM and may, in fact, be slower in some applications than even PC133 SDR-SDRAM. We suggest avoiding RDRAM memory entirely. In the past, we recommended RDRAM for Pentium 4 systems for which memory performance was a very high priority and the additional cost of RDRAM was not a deciding factor. With the advent of dual-channel DDR-SDRAM systems, that advice is now obsolete, because dual-channel PC2700 or PC3200 memory outperforms RDRAM in every respect. Intel will soon discontinue all of their RDRAM motherboards, making the point moot.
As of July 2003, the memory landscape for PCs appears to be predictable for the next couple of years. PC133 SDR-SDRAM and PC1600/PC2100 DDR-SDRAM are useful only for upgrading older systems. PC2700 DDR-I SDRAM is the current standard, although the Intel 875P- and 865-series chipsets have made PC3200 DDR-SDRAM a mainstream technology. Inexpensive systems use single-channel PC2700 DDR-SDRAM, and mainstream or higher systems use dual-channel PC2700 or PC3200 DDR-SDRAM. This state of affairs is likely to remain unchanged for the next year or more, so PC2700 or PC3200 DDR-I SDRAM remains a "safe" purchase.
As we move into 2004, PC2700 and PC3200 DDR-I SDRAM will begin yielding its position as the mainstream memory technology to dual-channel PC3200 DDR-II SDRAM, at first in high-end systems and later in the year in midrange systems. By late 2004, only entry-level PCs will use PC2700 or PC3200 DDR-I SDRAM. Beginning in 2005, even inexpensive systems will use PC3200 DDR-II SDRAM, and high-end systems will use faster variants of DDR-II SDRAM.
Rambus RDRAM never became a mainstream memory technology despite Intel's efforts to push it. Those efforts were particularly futile with Pentium III-class processors, which do not require the additional bandwidth available with RDRAM. Those early efforts to promote RDRAM failed miserably because people noticed that despite the hype, RDRAM provided little or no performance benefit relative to PC133 SDRAM with sixth-generation processors.
The advent of the bandwidth-hungry Pentium 4 processor should have made the advantages of RDRAM compelling, but the issue of relative memory performance has been overtaken by events. Intel's contract with Rambus has expired, its enthusiasm for RDRAM has faded, and it has now developed dual-channel DDR-SDRAM chipsets that provide more throughput than RDRAM with better latency and at a lower price. In our testing, RDRAM-based Pentium 4 systems provide better memory performance than those that use single-channel DDR-SDRAM, but worse performance than those that use dual-channel DDR-SDRAM. That leaves RDRAM as an expensive technology with no remaining market niche, and we expect it to fade quickly.