<!–#set var="article_header" value="Perfect Timing:
DDR Performance Analysis” –>
Perfect Timing: DDR Performance Analysis
It’s a never-ending story: the media strongly recommends the use of fast memory, while in the meantime many computers sold in the big retail stores are factory-equipped with cheap and slow memory, thus hampering the performance of supposedly “modern” computer systems. Are we being sold nothing more than junk? Or is the performance gap simply much smaller than we thought?
There are some core components that are responsible for the overall performance of a computer; these are, namely, the processor, the hard drive, the graphics card and, last but not least, the chipset, including the memory controller. But there is no sense in buying premium hardware if even one component remains below average. One popular strategy among system integrators is to use cheap memory. Modules that only run at CL3 (SDRAM) or CL2.5 (DDR) are usually slightly cheaper than DIMMs for CL2.
There’s no question that CL2 has a performance advantage over slower modes, but what does the difference look like in the real world? This article takes a look at four different memory settings, ranging from the absolute minimum to the most optimal using an up-to-date Pentium 4 system.
The Magic Word is “CAS-Latency”
Users that are less experienced may have heard of the catchwords CL2 or CAS2 already. Unfortunately, only a few dealers are able to explain the meaning of these equivocal words. Attempts to explain sometimes end with other equally vague terminology, such as “Cache Latency” or other funny expressions.
Latency is a type of waiting time – which makes sense in relation to memory chips. The “C” or “CAS” stands for “Column Address Strobe.” To explain this term, we should start with the memory chip itself:
It’s best to imagine a memory chip as some kind of storage matrix or table. In order to address each position, the matrix is organized in rows and columns. Thus, CAS is the column address strobe. As you can imagine, there is also a row address strobe (RAS).
From top to bottom: Infineon 512 MB DDR266 registered DIMM with ECC (72 instead of 64 Bit) for servers and workstations; Corsair 256 MB PC2700 DDR-DIMM (up to 166 MHz) for overclocking; Infineon 256 MB DDR-DIMM CL2 and Micron 128 MB DDR-DIMM CL2 – both are quite good modules.
What Does Memory Access Look Like?
In practice, this is what happens: in order to read or write to a memory point, the memory controller first transmits the row number of the desired address, and the RAS signal is activated. However, it takes a few clock cycles (the RAS-to-CAS delay time) until the columns can be accessed. Likewise, after the CAS signal is activated, further clock cycles must take place. With standard SDRAM based on the PC133 standard, this is always two or three; with DDR RAM, four or five clock cycles. In the DDR process, the “real” CAS latency times are two or two and a half clock cycles.
The RAS-to-CAS time depends on the technology, and, since it’s mostly between five to seven clock cycles, this is the fundamental factor for delays. You could also say that CL2 memory is by far superior to slower models with CL2.5 (DDR) or CL3 (SDRAM). Theoretically, this is true, but other factors put the advantages of fast data transfer in another light – for one thing, modern processors have a high hit rate with cache memory, which means that they seldom have to read directly from the memory. Secondly, the rows are also switched most of the time, so that RAS-to-CAS time is also needed, thus delaying the read process. Finally, the burst reads will occasionally happen — in this case, many of the memory cells that are directly adjacent to one another are read, but here the CAS latency occurs only once, and is therefore insignificant.
Buying Memory: Always Go For Speed?
Experience dictates that the advantages of fast memory are worth the slightly higher price that you have to pay to get it. However, many retailers still offer high-end PCs that make do using the slower memory with which they are equipped. This makes it entirely possible for a new Pentium 4 PC with a hefty 2 GHz to end up lagging behind the neighbor’s 1.8 GHz computer in the benchmarks. But don’t forget that the chipset also plays an important role here. Aside from Gigaherz and Gigabytes, the inner values (especially in the memory interface of the chipset) contribute considerably to higher performance.
When buying memory, if the price for CL2 memory is only a little bit higher, then we recommend that you go for it. However, you should keep in mind that the slowest unit in the group will be the performance bottleneck. For instance, if you use three DIMMs with CL2 and one with CL2.5, then all four will run with only CL2.5
The most important thing is the price/ brand comparison: make sure that you always distinguish the brand name modules (Infineon, Micron, Samsung, Hyundai, NEC, etc.) from the no-name modules. In terms of price, this usually translates to about a 30% difference!
Looking Ahead: CL1.5
From a technical point of view, there are two ways to increase performance for DDR SDRAM. The first method is to increase the clock speed to 200 MHz (400 MHz DDR), a method which is currently championed by VIA. This will come to fruition this year with the release of the KT400 chipset for Athlon systems. The second method is to further reduce CAS latency. This is technically possible with DDR SDRAM, and CL1.5 would be the next level of development. Still, it’s questionable whether this method will really become established or not, since it requires considerable effort. Making the step to four data transfers per clock cycle alone would also require much effort, similar to the system bus of the Pentium 4. Here, the bandwidth would increase significantly, thereby the question of fast timings would be put aside for a while.
Test Setup
Test System | |
CPU | Intel Pentium 4, 2GHz |
Motherboard | Abit BD7i845 DDR Chipset |
RAM | 256 MB PC2100 DDR-SDRAM,CL2 (Corsair Micro) |
FireWire-Controller | Texas Instruments |
Graphics Card | nVIDIA GeForce 3,64 MB DDR |
Network | 3COM 905TX PCI 100 MBit |
OS | Windows 2000 Professional |
Benchmarks | |
Office Applications Benchmark | BAPCo SYSmark 2002 |
OpenGL Performance | SPECviewperf 6.1.2Quake III Arena |
Memory Performance | SiSoft Sandra 2002 Pro |
Game Performance | 3DMark 2001 SE |
Drivers & Settings | |
Graphics Driver | nVIDIA Reference Driver 28.32 WHQL |
Chipset Driver | Intel INF Update 3.20.1008& Application Accelerator 2.1 |
DirectX Version | 8.1 |
Resolution | 1024×768, 16 Bit, 85 Hz |
Although Windows XP is managing to become established as the standard operating system, the dynamic memory administration is a roadblock when it comes to getting exact benchmark results. In evaluation of memory performance, the subtlest of nuances are also important, therefore we turned to Windows 2000 Professional.
Explanation: the charts show you results that are labeled “Slowest” and “Fastest.” The following table shows you the settings we chose:
Fastest | Slowest | |
Act to Precharge | 5 | 7 |
RAS-to-CAS | 2 | 3 |
RAS Precharge | 2 | 3 |
The differences in the implementation of these three parameters are so small that the benchmarks are not able to give us results that can be interpreted, due to the measurement tolerance. Therefore, we ran all three values with the fastest and the slowest possible settings, each with CL2 and CL2.5.
Maximum Memory Transfer Rate: SiSoft Sandra 2002 Pro
Office and Internet Content Creation: SYSmark 2002
OpenGL Performance: SPECviewperf 6.1.2
Game Performance: Quake III Arena
3D Performance: 3DMark 2001 SE
Conclusion
Looking at the benchmark results, you can clearly see the potential of RAM. In terms of percentage, the differences don’t play a very important role, but when you consider the usual performance differences in motherboard comparisons or the small performance increases within an individual processor series, CL2 memory is a must.
Above all, fast memory is a great purchase when you consider the slight difference in price. And, if you are planning to overclock your computer in the future, your efforts are sure to be more successful with brand-name modules than with 08/15 DIMMs.
If you are gearing up for the future, look out for modules with 166 MHz memory clock (DDR333 or PC2700), which can also be run in CL2.0 mode. As of yet, these modules are still too rare and too expensive, but they should become increasingly available in the months to come.