Looking for memory for my i7-870 OC. I think I will OC to 4G. The MBD supports 1066/1333/1600(OC)/1800(OC)/2000(OC)/2100(OC).I am trying to decide on memory. Here are the two I have narrowed my selection to -
G.Skill Ripjaw - DDR3 1600 7-7-7-24 http://www.newegg.com/Product/Product.aspx?Item=N82E168...
Will the OCZ run at faster CL timings if they're set to slower than 2000 MHz?
Will the G.Skill's run at lower CL timings if they're set to faster than 1600 MHz?
The answer to both is: Maybe. Not trying to be flip, but you won't know what they're capable of until they're in your system and you run them through a gauntlet of tests.
If it was me, I would pick the one closest to what I expected to run most of the time, or pick based on price. So go for the cheaper G.Skills. If OCing the CPU past 4 GHz makes the RAM unstable, then back the RAM multiplier down to get it stable with your high CPU OC. There won't be that much difference in memory performance, outside of synthetic benchmarks, at that level anyway. 1600 CL 7 is already very fast.
EDIT: By the way, I just noticed the G.Skills you linked to are 2 x 1GB.
From what I can understand from the text and charts the memory bandwidth generally more important than the CL spec. For example the DDR3 2000 memory will support higher data transfer rates than the DDR3 1600. Also if I calculate the latency of the two sticks I was looking at (adjusting for the 2x2G) I get the following latencies
1600 CL7 = 8.75ns, 2000 CL9 = 9ns. So they are very close.
By not knowing if the 1600 CL7 memory will operate at 2000 I woudl have to purchase it and take a chance. If I end up running it at 1600 I have not selected the correct memory wrt to bandwidth.
I know that memory speed does not make that much difference in most applications but I plan on using this system for video editing. I have purchased a 6GB/s SATA drive and want to max out the memory.
My Motherboard is the MSI P55-G85 w/ I7-870 and I am planning on running w/o Turbo at 4.0Ghz (191Mhz BCLK w/ 21 Multiplier) Using a 10x multiplier on the memory takes me to 2000 on the memory clock.
Have I missed anything ? Go for the faster speed memory if the latency is not much different (assuming you can run that fast).
What I don't quite understand is the affect of the other numbers in the memory specs. there are 4 numbers on the 2000 speed memory 9-9-9-30 vs 7-8-7-24 on the 1600. How do these affect the scaling? Or are they insignificant when it comes to memory bandwidth?
Those 4 numbers are for various timings. The first is the CL or CAS Latency number which you noticed many references to in your research. The number are separate from bandwidth but not insignificant because the two impact each other in the sense you cannot have very low timings and very high frequency at the same time (without paying an arm and a leg for it). Thus there is always the decision, the same that you started with, as to whether or not to go for lower timings or higher bandwidths.
And so, I wouldn't agree that it is always true that "From what I can understand from the text and charts the memory bandwidth generally more important than the CL spec." You could have RAM with less bandwidth but much lower timings that performed faster as a result. You have to do the mat. And also consider what the primary use of the computer is going to be because some programs make effective use of high bandwidth while others do not.
Great references. I now understand how the memory bandwidth and latency play together to make up the overall performance of memory. It gets very complicated when you throw price back into the mixture as different vendors offer different prices for different performance memory.
I opted for OCZ Platinum memory DDR3-2000 CL9 for $91 after rebate from NewEgg. There were other choices for better performing memory but at higher prices. If you do the math I think I get the following performance calculations -
2000 Mhz CL9 = 9/1000Mhz or 9ns access time. (use 1/2 the frequency because dual access memory) ~ 1333Mhz CL6
Here is the formula I used ;
First calculate the actual latency of the memory at its given ratings. For my example the memory is rated at 2000 Mhz with CAS 9. So my Latency = 9ns [Latency = CL / SPEED * 2]
Then calculate the CAS at different speeds; CL = Latency * SPEED / 2
I finally received the memory from NewEgg. Unfortunately - however not surprising the internal tables in the DRAM only support up to 1600Mhz and with a CAS of 9, my calculations are for CAS 7 at 1600. Looks like I will have to manually set the settings but there are many more than just the first 5 numbers so not sure what to do with them.
Next step is to find a good benchmark tool that will help me baseline the configuration, do you have any suggestions?
If you really want to stress your system you could grab Intel Burn Test (running LinPack). It will put any OCed rig through the ringer. Run it before & after your changes and compare how the results (hopefully) improve.
My motherboard only allows for all memory timings to be auto or manual. When I set them to Auto I think they used the memory tables for the settings. The OCZ memory did not have the advertised timings in its tables so I was forced to set them manually and guess at many of the numbers.
I ended up returning the OCZ memory as defective. I could not ever get it to run at 2000 CL9. MY MOBO needed additional timings because the memory did not publish its timings at 2000. This should have been in the memory XMP table, however the XMP table only had 1333, not 1600 or 2000 and therefore was not able to provide my MOBO with the proper timings. When I reached out to OCZ support they offered no information and told me to ask their community. That was not something I was interested in - I don't think is it unreasonable to expect the manufacturer to have timings for the memory it is selling at the advertised speed, but they did not. It is very low cost and at lower speeds would probably work but I would not call it 'High Performance". It was not at the performance I was looking for especially given the 2T command rate.
I went with Mushkin 2000 CL7 memory. I am extremely satisfied with that memory. It was almost twice the cost of the OCZ memory but since I am expecting to use the system for video editing I needed the high memory bandwidth and I did not want to waste a lot of time fighting with memory timings. Their product support was excellent, valuable, and my emails were returned within 1 day. OCZ took several days for a response and their response was not helpful.
The next challenge I encountered was to get a stable OC configuration on the CPU. The MOBO only lets me change the multiplier for memory off the base clock. 4x, 5x, 6x so I needed to find a stable base clock / memory multiplier that runs the CPU 'cool' enough and is stable. I have the system running at base clock of 160.5Mhz, cpu multiplier of 24, memory multiplier of 6. This gives me a CPU running at 3.85GHz and memory at 1905Mhz CL 7. Running SiSandra the memory bandwidth was 23.55 Gb/s and latency was 57ns. I am not completely happy with the memory at 1905Mhz so I am trying a base clock of 200.5Mhz with a CPU multiplier of 19x and memory multiplier 5. This will provide a CPU running at 3.81Ghz and memory at 2005Mhz CL7. My initial measurements on SiSandra reported memory bandwidth of 25.37Gb/s and latency of 51ns. I was in the middle of trying to find the lowest voltages to keep it stable when I decided to replace my CPU Heat Sink. I returned my Thermalright MUX 120 and have ordered an Thermalright Venomous X. I have also been on vacation for two weeks - thus my extended delay i this post.
I have been able to push the memory to 2075Mhz but the system configuration was not stable under Prime95. SiSandra did report bandwidth of 26.08Gb/s and 50ns. The CPU Temps were in the 80's so that is why I am changing out my heat sink to try and get that lower. At this configuration the CPU was running at 3.94Ghz and the temp was 80C. It failed under Prime95 and IMO too hot so that is why I am changing out the heat sink. We will see when I get the new one.
In summary I think my recommendation for others who are trying to max out memory here are some points to look for.
1) Unless you have a lot of time and don't mind experimenting go with a good brand of Memory. The OCZ memory IMO should have had XMP table settings for its rated speed. It did not and the manufacturer support could not provide it so you were left experimenting. I don't see how they can provide lifetime warranty support by having people just experiment. I selected Mushkin memory
2) Push for the lowest CL you can afford at the highest speed. Most MOBO are very happy around 1600 and that seems to be a nice overclock for average users. I think it is hard to push up to 2000 and find a stable configuration. If you keep the CPU voltages at the 'auto' settings then your CPU will get very hot and may be too hot for a long life running at that temp. Rule of thumb, if you are providing 1.45 v to the CPU and you read in CPU-Z 1.23v underload IMO that 1.22v is going into excess heat.
3) Memory speed and latency are both important, I started this thread trying to understand which was more important and I gave up and just purchased memory that provided good numbers for both.
4) Purchase memory at the speed / latency you want. Trying to adjust or push too far may not work and will take a ton of experimenting. If you don't have the exact timings at the speed you want for COMMAND RATE, CL, tRCD, tRP, tRAS, tWR, tRFC, tRRD, tFAW, tRTP then you are just guessing for those numbers and will have a difficult time getting to a stable config. If you have a MOBO that will do 'Auto' for those numbers (assuming it is not just going to the tables in the memory) you may be able to get a stable config.
5) XMP / Memory Profiles in memory. Each memory DIMM has built into it a table of memory timings. This was so MOBO's could find a common timing table between different memory DIMMs. The MOBO will use the set of timings that match across all memory DIMMs installed. Anyway I only tried two manufacturers - OCZ and Mushkin. The OCZ table did not support the advertised speed, the Mushkin did. I found that my MOBO would use that table when it was trying to make the 'auto' settings.
6) Unless you are trying to use an application that is strongly dependent on memory speeds you may not benefit from the extra cost of high performance memory. From the reviews I have read Video editing seems to be the only application space that benefits. My guess is that any memory bound app would benefit.