Memory Timing and Frequency Comparison
02.22.07 - 08:54pm
Rarely does the RAM in a computer receive all that much attention, normally the processors and video cards get all the hype with their 3D and numbercrunching benchmarks. Traditional desktop processors normally have a small amount of highspeed onboard memory or cache but this cache isn’t nearly enough memory to store all the variables necessary for a functional computer. Computer hard drives on the other hand have enough memory to store millions of images and files yet the access speeds are measured in milliseconds rather than nanoseconds like onboard cache. RAM fills in the gap between these two storage mediums with moderate capacities and decent access times. More often than not the operating frequency of the modules overshadow the memory timings so I will be doing a short investigation into the importance of memory timings.
SDRAM which composes most of the ram modules found in modern desktops operates by constantly refreshing the data stored and accessing and modifying data in set periods of times that correspond to the RAM timings programmed into the motherboards BIOS. Most often the CAS, RAS to CAS, RAS Precharge, and Precharge Delay are the timings that memory modules are rated for while the various sub-timings are rarely mentioned. An example would be my Crucial 10th Anniversary sticks which I believe were rated for DDR2-667 at 3-3-3-9 timings or something similar to that. Each number corresponds to a set number of clock cycles, so the CAS is 3 cycles, RAS is 3 cycles, RAS precharge is 3 cycles, and the Precharge Delay is 12 cycles. More often than not your budget sticks will operate at the same frequency as your performance sticks but they will feature considerably longer memory latencies.
It is very common for your OEM computers to feature the slowest memory modules possible in an attempt to save as much money as possible. In most instances this is not a problem as most desktop computers aren’t run at 100% load with a priority placed upon performance. But memory performance can make a difference when maximum performance is required such as 3D modeling, gaming, and simulations. I will be using my favorite memory-limited program to test for memory performance, SuperPi. SuperPi is a great application that responds very well to increases in effective memory bandwidth as well as memory latencies. The fastest times are usually recorded with a combination of tight memory timings and high memory frequencies which make perfect sense considering how memory latencies will actually drop as clockspeeds advance but then the timings are increased to compensate for this.
Test
For the tests I used my Gigabyte DS3 with the C2 north bridge, 2×1GB kit of Crucial 10th Anniversary sticks (high-bin D9GMH), my trusty E6400, and the Scythe Ninja Rev B. To try and eliminate processor speed gains I ran the processor at the stock 2.13GHz and used the memory dividers in the BIOS to adjust the memory speeds. Through the entire test the processor was set to 1.35 volts to insure stability and the ram was set to 2.20 volts while the FSB, MCH, and PCIe voltages were left on default. The memory was tested with the 1:1, 4:5, 4:6, and 1:2 dividers to deliver DDR2-533, 667, 800, and 1066 frequencies. Memory timings were set either as 3-3-3-9, 4-4-4-12, 5-5-5-15, or 6-6-6-18 with 4-3-3-10 being substituted once for 3-3-3-9. The time to compute SuperPi 1M was recorded after three iterations were performed and the average was taken. SANDRA was used to measure the available bandwidth along with the bandwidth efficiency.
Results
Most of the results scaled just how I expected but the available bandwidth had a huge gap at DDR2-800. I do know that at a 400FSB the 965 chipset is known to jump from the 1066 to the 1333 northbridge strap which effects the FSB latencies but I didn’t except to see such a huge drop. From these basic results I can safely say that memory frequency is more effective than memory timings but memory timings also can play a crucial role. A great example of the differences between low timings and high frequency would be the DDR2-533 3-3-3-9 compared to the DDR2-667 5-5-5-15, the DDR2-667 was still faster while providing roughly 300 more mb/s of available bandwidth. At the end of the day if it came down to having to decide between frequency and timings, frequency would be the surefire winner. The improvements shown by SuperPi make it obvious that performance will scale greatly with frequency while only moderately with timings.
So what does this mean to you? This can be taken a lot of different ways, for those individuals that are looking to purchase new memory modules then this could help make the difference between some DDR2-533 and DDR2-667 or 800 sticks. One thing that needs to be taken into account though will be what the individual memory sticks will be capable of, for example some DDR2-800 rated for 5-5-5-15 might be capable of DDR2-800 4-4-4-12 or DDR2-1066 5-5-5-15 making them potentially a better buy than some DDR2-1066 5-5-5-15 sticks that cost twice as much. Research is crucial in this aspect if you are attempting to stretch your dollar as far as you can. When I purchase memory I look for sticks that will scale very well with decent volts, I’d rather have DDR2-1066 5-5-5-15 at 2.1 volts rather than DDR2-800 3-3-3-9 at 2.3 volts even though the DDR2-800 3-3-3-9 sticks could be faster. Memory voltage is another thing to keep in mind, generally 2.2 volts is considered the max for a safe 24/7 setting with current DDR2. I personally operate my memory between 2.1 and 2.2 volts with a fan cooling them. I hope my research was useful, the differences between high frequency loose and low frequency tight sticks was rather surprising as I thought they’d be a bit more even in performance.
Interesting. It took me a while to figure out completely what you were saying, but I think I get it… but how would these results make a difference for a Mac? Is that just a completely different world? I ask because I´m thinking of getting a Mac and upgrading/scaling up in the future could be an issue. Sorry if this stumps you… cheers!
Well current Macs are based around Intel processors and motherboards so there would be a slight performance gain between DDR2-533 and say DDR2-800. The only problem with Macs are that they only pump 1.8 volts through the memory modules, a good portion of DDR2-800 memory modules won’t even boot at 2.0 volts, let alone 1.8 volts. Since Macs are a “sealed system” when deciding on a memory upgrade you’ll have very few options to choose from. Right now your best bet will be the cheaper DDR2-667 sticks being sold for desktops and laptops as most will boot easily on lower voltages. Just do a little research for Mac review websites and I’m sure you’ll be able to find memory that will be compatible. In regards to the performance difference I don’t know if the daily user will notice the differences between DDR2-533 and DDR2-1066. When using a computer for daily activities like surfing the web and other non-processor intensive applications you won’t notice a difference at all. In a select few situations where you are maxing out the machine for a lengthy period of time such as video encoding you may notice a ~5% difference in performance between 533 and 1066 which usually won’t justify the costs for the high-end memory. You will only notice these differences when you are performing a benchmark, I used SuperPi because I happen to actively compete against others with it and it is very sensitive to changes in bandwidth. So all in all if you have to make a choice between DDR2-667 4-4-4-12 and DDR2-533 3-3-3-9 and the prices are similar, just go with the DDR2-667. It won’t be worth your money to try and get the tightest timings at the highest speeds as you only see those benefits in artificial benchmarks. Above all things though will be capacity, you will see a huge improvement comparing 1 gigabyte to 2 gigabytes of memory and less of an improvement with 4 gigabytes. Just check how many memory slots your Mac will support, I’d suggest going with 2 1GB sticks and if it supports 4 slots then upgrading down the road to 4 1GB sticks. Don’t waste your money on 512MB sticks, you’ll just have to replace them later. I hope that helps a bit, I’m a bit verbose but I could talk for days, well hours, about this subject.
Great write up Gomez.
Hey Chris Morrell, I read your research on memory latencies. However, I’m not sure what you mean by scaling with the memory voltage. Please explain. Thanks
Great Work Man!! I’m glad that there are people like u arround.. I really appreciate ur research. :)
No problem, glad I could help ;)
Hey Chris,
You sound like you like to be a great helper out there, memory timings and performance right now are my greatest weakness, Im building a new gamer. specs gigabyte ms790fx ds5 amd athlon 6400+ 3.2 gig twin asus 3870×2 set up in quad crossfire 2×16x pci-e, coolmax 1200 watt power supply, processsor cooling swiftech apx ultima full kit but just processor on the cooling (at least for now, Antech 900 case. Memory timeing has always been my broken leg so to speak and i dont totally comprehend its affect against performance my plan on memory is geil evo one as it was origionally recomended to my by dfis head tech who really hit it on the head but this was ddr and athlon xp but geil has really proved itself to me my plans are a quad matched set of evo one 8 gig total (4×2gig) 800. newegg got in a brand new order of the 44412 and they sold out overnight, im disabled so i only get one check a month and didnt make it they still have the 55515 at 1.8 volts the 44414 is 1.9 to 2.0, so now im really at a loss should i save the 4bucks buy the little cheaper and push it, or should i get the latter which lists higher voltages, I have had great luck with geil including runnning really cool. any help explanation, or even some websites to help me understand better would be so sincerelly appreciated, I am very smart and can pretty much turn most anything i read into real world use, I guess thats all I can think of at the moment please help.
Sincerely,
Paul
Great article. Mentionned alot of info I was looking to clear up especially about the DDR2 1066 @ 5-5-5-15.
Well spoken, well researched, Thanks!
i want to upgrade my memory to 4 gig so i was doing a bit research and i was left with making a choice between kingstone’s hyperx 2×2gig kit of 667mhz with 4-4-4-12 and 800mhz 5-5-5-18 and i choose to go with second one as the price of both the ram were same ….
but i was wondering dose the timing would have made a difference in this case in performance compared to frequency ….
both the ram operate at 1.8 v and the one the i bought can be overclocked to 4-4-4-12 but do u recon to overclock a ram
Very good article, but some explanations of the less important timings would have been nice, you can find brief explanations at the memset website at tweakers.fr.
I have a Phenom II 9150e with DDR2-800 at what I set them to 6:5:4:17 with tRC 21, and are obviously a nice improvement over 6:6:6:18 tRC 26. I just bought a new set of 2×2GB ram and a new video card and my rig should really fly when I get them, I bought Patriot DDR2-1600 with 4:4:4:12 and can probably be changed to 4:4:3:12 at least, as well as a Gecube HD 3850 factory overclocked to 733MHz. I’m so happy with my choices I could squeal like a little school girl!