New gaming rig, need part recommendations - Page 3

There are four slots on the board, we'll call them 1, 2, 3, and 4.
Slots 1 and 3 are a "pair" and they work together in a "push-pull" type setup. The external memory clock (which is at a given frequency, let's call it 100 MHz) has a HI and a LO signal. On the HI signal, it's sending data to slot 1 while at the same time it's reading data from slot 3. Then when the signal goes to LO it's reading data from slot 1 while it's sending data to slot 3. This is what I'm calling a "push-pull" as it's doing two jobs at once. This HI-LO transition happens 100,000 times every second. The idea that slot 1 and slot 3 are working as a push-pull pair is called "Dual-Channel" mode. This means that when you install RAM you want to do it into the colored slots that match one another. Typically, there will be four slots on a board, two of them blue and two of them grey or two red, two white, or something like that. By putting a DIMM (or stick of RAM) into each of the grey slots, and leaving the blue slots empty, for example, you're telling the motherboard that you want to work the RAM in a Dual-Channel mode. This makes the effective speed of your RAM on the external clock 200 MHz. (Remember, that clock is still only at 100 MHz, but since we're having it perform an operation on the HI and the LO phase simultaneously, we call it a 200 MHz external clock.)

Now, the RAM also has an internal clock, and that's where you get into RAM timings. DDR3 RAM, for example, runs at 8 times the speed of its external clock. So we take that 200 MHz we have as our external clock and we multiply that by 8. This is where you get the DDR3 SDRAM DDR3 1600 (PC3 12800) designation from. That 1600 represents the external 200 MHz clock we have multiplied by the 8x internal clock for a final internal RAM frequency of 1600 MHz. (1.6 GHz).

But there's more going on inside the RAM than that. Since we're writing and reading data to and from multiple physical chips on the DIMM unit and also doing this millions of times per second, we have to have a way to regulate it all inside the chip and that's where you get into things such as CL, which is Clock cycles between sending a column address to the memory and the beginning of the data in response, tRCD which is Clock cycles between row activate and reads/writes and tRP which is Clock cycles between row precharge and activate.

That's probably more than you want to know about RAM and it's beyond the scope of layman's terms, so a simplification that the industry uses is to call RAM "Cas Latency" 8 or 9 or 10 or whatever, and then publish the "Timings" of 9-9-9-24 for a CL 9 module, 8-8-8-24 for a CL8 module, and in some of the really high performance DDR3 RAM you can get 7-8-7-24 timings in CL7 modules.


The lower the numbers in those timings are, the better. And the higher the number after the DDR, the better. DDR3 is better than DDR2. DDR3 1600 is better than DDR3 1333. DDR3 2133 is better than 1600, and so on.

So what we want is a combination of high frequency (DDR3 1600 for example) and low Cas Latency or fast timings. CL8 at 8-8-8-24, for example.

This means that our memory is clocked very fast and this means that we can read and write data from and to it very quickly, taking advantage of that Dual-Channel "push-pull" mode we mentioned.

However, all of these clock frequencies have to be timed perfectly between chips. And the HI and LO values on those chips in our RAM come from the voltage that our motherboard is supplying to our RAM. The more chips we put on, the higher the voltage has to be. This is something that's handled automatically by the motherboard. The problem comes when we put too much voltage to a RAM stick, or DIMM. We either melt it and ruin it, or we introduce "False HI" values when we really want a "LO" value. The reason for this is also probably beyond the scope of layman's terms, but the simple explanation is that the voltage isn't perfectly square "HI and LO" 0.000V and 1.5000V for example. Our voltage, even though we want it to be ON/OFF or HI-LO on our commands is really slightly analogue, like a sine wave. So to prevent our RAM from making mistakes and mistaking a 0.2500 V signal as a "HI" what we have to do is slow it down a little bit. By slowing it down, we allow that 1.5V HI signal to bleed off to a stable 0.000V low signal so that it won't be mistaken for a HI. By slowing it down, we also give it time to develop a good 1.5000V HI signal on the high end, rather than 1.3565V or something. Basically, by timing it too quickly, it's taking a snap-shot of the __----____----____--- pattern at the wrong time, and not seeing it as a square wave. It's seeing more like a sawtooth _/---\_ and when it takes its reading on the / or the \ rather than the - hi or _ low, we have issues.

This is why, when you fully populate the DIMM slots (or fill up the motherboard with as many sticks of RAM as it will hold) that you slow things down. The motherboard has to supply power to all four modules rather than just two. Furthermore, the bigger each module is, (4GB module vs 2GB module) the more power it takes too. So the motherboard sees your 4 modules at 4 GB each and it knows it needs lots of power to feed them all. It supplies that, but to keep your RAM from having errors (giving you Blue Screen of Death) it has to slow down several things. First, it slows down your internal multipliers. It does use the dual-channel, so you get that 200 MHz external, but it can't push your 1600 MHz ram at its full x8 internal multiplier for 1600, so it slows it down to a final value of 1333 MHz. Furthermore, it can't run that at the 8-8-8-24 timings either, because then your internal timings won't match your internal speeds, so it slows it down to something like 11-11-11-32, for example.

Also, with all four DIMMs populated, we have only pairs 1 and 3 working as a push-pull pair. Pairs 2 and 4 are also working as a push-pull pair, but the two pairs are independent from one another. In other words, you have 2 DIMMS in dual-channel mode, and 2 more DIMMs in dual-channel mode. You do not get quad channel mode. So this is an added bottleneck at slowing things down.

The end result is that you now have 16 GB of RAM, which is a large amount, but it's actual timings are slow, at 1.333 GHz and 11-11-11-32. This means that the speed of the data you're sending to and from the RAM is slowed down accordingly, resulting in longer loading times, slower performance, and so on.

If you cut it down to 8 GB of RAM, which is still quite a bit, you can run at the full 1600 MHz and 8-8-8-24 timings. This means that although you have less RAM on your system, the speed of the data being sent to and from it is fast.

We haven't even overclocked our RAM yet, and we're getting a 25% performance increase in our speeds, just by going with 8 GB of RAM.


So why do they sell 16 GB packages? You make more money on 4 x 4 GB than you do on 2 x 4 GB for one, and if people want it, why not give it to them? That's one reason. Another, and more relevant reason is that some people are editing massive video files and they need 16 GB of RAM otherwise they'd have to use the pagefile on their HDD and that's slow. For those people, it's the size of the RAM that matters more than the speed. That's why they're willing to run at 16 GB, 32 GB, or even 64 GB of RAM on some of the high end graphic workstations and such.

Gamers only need 8 GB of RAM to handle every game on the market today at maximum settings and with a pagefile disabled. Extreme overclockers, who want the absolute most out of their systems will often go with 2 x 2GB in order to get wicked fast timings in the CL5 range. Although this makes those insanely high RAM timings look awesome on benchmarks, they're going to suffer in real-world throughput on their RAM and they'll have to use a page-file to help load games that want more than the 4 GB of memory they have available. The sweet spot is actually at about 6 GB of RAM, but the problem is, they don't make things in 3's in the PC world. Things come in 2's, so your options are 1, 2, 4, or 8. I don't know of any 6 GB or 3 GB modules. So, your choice is 2 x 1 GB of wickedly high performance RAM for insane CPU benchmarks and no real utility for anything else, 2 x 2 GB of RAM at very high timings for an overclocked rig that will still need to use a page-file to load textures or 8 GB of relatively high performance RAM with no page file to run your games blazingly fast but not enhance your e-peen on your benchmarks as well. Or, on the opposite extreme, you get 16GB, 32GB, 64GB, or even 128GB of RAM for your graphics editing workstation that's using quadro cards and lots of CUDAs to pump out insanely high resolution 3D renderings without thrashing your hard drive.