[Peon]

Or put another way, what speed DDR4 would give the same level of performance as an average stick of DDR3-1600/1866/2133/etc. etc.?

 

[billbartuska]

Memory latency, or timings, like the number 10 in CL 10, are the length of time it takes the memory to complete a step in what it has to do. That "time" is measured in "clock ticks", ie CL 10 takes 10 clock ticks to complete before the memory can move on to it's next operation. The length of one clock tick is the speed at which the memory is running. 1800 MHz memory has a clock tick length of one 1,800,000,000th of a second (1,800,000,000 clock ticks per second), so the CL step takes 10 x 1/1,800,000,000 seconds.

A stick of memory always takes the same amount of time to complete it's CL step (or any other step) no matter what speed it is running.. If you run the above memory stick faster, say 2400 MHz, it still takes 10 x 1/1,800,000,000 seconds to complete the CL step, but each clock tick is now 1/2,400,000,000 of a second, so it now would take more clock ticks to complete the CL step. Namely, 24/18 times 10 (for CL step) or 13.3 clock ticks (10 times 24/18 clock ticks). But, alas, that has to be rounded to to CL 14 as memory can't use partial clock ticks..

So to compare different memory you compare the actual time it takes each one to complete what it has to do, Usually just calculating the length of time for the CL step for each will give you your answer.

 

[vs17e]

Quote:

Originally Posted by billbartuska

Memory latency, or timings, like the number 10 in CL 10, are the length of time it takes the memory to complete a step in what it has to do. That "time" is measured in "clock ticks", ie CL 10 takes 10 clock ticks to complete before the memory can move on to it's next operation. The length of one clock tick is the speed at which the memory is running. 1800 MHz memory has a clock tick length of one 1,800,000,000th of a second (1,800,000,000 clock ticks per second), so the CL step takes 10 x 1/1,800,000,000 seconds.

A stick of memory always takes the same amount of time to complete it's CL step (or any other step) no matter what speed it is running.. If you run the above memory stick faster, say 2400 MHz, it still takes 10 x 1/1,800,000,000 seconds to complete the CL step, but each clock tick is now 1/2,400,000,000 of a second, so it now would take more clock ticks to complete the CL step. Namely, 24/18 times 10 (for CL step) or 13.3 clock ticks (10 times 24/18 clock ticks). But, alas, that has to be rounded to to CL 14 as memory can't use partial clock ticks..

So to compare different memory you compare the actual time it takes each one to complete what it has to do, Usually just calculating the length of time for the CL step for each will give you your answer.

tl;dr version: latency / frequency = time, choose the one with the smallest number

 

[Slink3Slyde]

Well with respect I think there are differences between DDR4 and DDR 3 that go beyond using the formula thats always been used.

Anand went into it here



I dont have more then a passing understanding of the mechanics of how RAM works so I cant offer a formula to compare. But I can see that there are obviously other differences that make DDR4 C15 work faster then DDR3 1866 C9 even though applying the above formulas would suggest the DDR3 should be faster.

If you were using DDR4 3000 C15 as I would suggest that you use with Skylake these days the difference its safe to say would be even more. Theres also been reviews that have shown that faster DDR4 helps out a fair amount in gaming situatons.

 

[billbartuska]

Anand:

In fact despite the development of new memory interfaces, the true latency for DRAM under default specifications has stayed roughly the same since DDR. As we make faster memory modules, the CAS Latency rises to keep higher frequency memory stable, but overall the true latency stays the same.

And when he benchmarks there are other things (memory controller. etc.) that effect the scores and show the slight differences he reports.

 

[Slink3Slyde]

Quote:

Originally Posted by billbartuska

Anand:

In fact despite the development of new memory interfaces, the true latency for DRAM under default specifications has stayed roughly the same since DDR. As we make faster memory modules, the CAS Latency rises to keep higher frequency memory stable, but overall the true latency stays the same.

And when he benchmarks there are other things (memory controller. etc.) that effect the scores and show the slight differences he reports.

(DDR, DDR2, DDR3, DDR4) if it ultimately has no effect on performance.

(CL/(Freq-in-MHZ) * 1000) = speed in ns

9/933*1000= 9.95ns

15/1066*1000 = 14.07

If 1866 C9 should be much faster then 2133 C15 RAM if you apply the formula. But in actual fact its a bit slower effectively as shown in most of Anands tests and in the games tested by Hard OCP on the same processor, taking the memory controller out of the equation.

Something else is going on behind the timings listed?

I apologize if I missed something in your post before its been a long day so its quite possible

 

[BoredErica]

I'm surprised that 1866 C9 is a little slower than 2133 C15 in that Anandtech page. Based on their performance index, the difference there is 207 vs 142, which is huge. The guy said in previous articles to generally favor the one with the higher frequency and that the index is a rough estimation that fails more and more as the frequencies go up, but still.

 

[billbartuska]

Quote:

Originally Posted by Slink3Slyde

Quote:

Originally Posted by billbartuska

Anand:

In fact despite the development of new memory interfaces, the true latency for DRAM under default specifications has stayed roughly the same since DDR. As we make faster memory modules, the CAS Latency rises to keep higher frequency memory stable, but overall the true latency stays the same.

And when he benchmarks there are other things (memory controller. etc.) that effect the scores and show the slight differences he reports.

(DDR, DDR2, DDR3, DDR4) if it ultimately has no effect on performance.

(CL/(Freq-in-MHZ) * 1000) = speed in ns

9/933*1000= 9.95ns

15/1066*1000 = 14.07

If 1866 C9 should be much faster then 2133 C15 RAM if you apply the formula. But in actual fact its a bit slower effectively as shown in most of Anands tests and in the games tested by Hard OCP on the same processor, taking the memory controller out of the equation.

Something else is going on behind the timings listed?

I apologize if I missed something in your post before its been a long day so its quite possible

What you're forgetting is the Data Transfer Rate. Data goes to and from the 1066 memory faster, so even though the data is "managed" slower internally by the 1066 memory it gets to and from the motherboard faster.

 

[Slink3Slyde]

So as pertains to the OP, 2133mhz C15 DDR4 is slightly faster then 1866 C9 DDR3 overall because of the newer faster memory controller/ z170 chipset improvements, despite the RAM itself actually working slower?

I'm guessing then, the highest sort of speeds you would get from DDR3L @ 1.5 volts might be 2400 C10 at the very top end, if you wanted to run 24/7 closer to Skylake spec on DDR3L at 1.35-1.4 volts even less. So maybe something like DDR4 2666 C15 or 2800 C15/16 would probably roughly match that overall at another educated guess on a Skylake chip. Going by the numbers that Anand and H got.

 

[billbartuska]

Quote:

Originally Posted by Slink3Slyde

So as pertains to the OP, 2133mhz C15 DDR4 is slightly faster then 1866 C9 DDR3 overall because of the newer faster memory controller/ z170 chipset improvements, despite the RAM itself actually working slower?

I'm guessing then, the highest sort of speeds you would get from DDR3L @ 1.5 volts might be 2400 C10 at the very top end, if you wanted to run 24/7 closer to Skylake spec on DDR3L at 1.35-1.4 volts even less. So maybe something like DDR4 2666 C15 or 2800 C15/16 would probably roughly match that overall at another educated guess on a Skylake chip. Going by the numbers that Anand and H got.

Overclockability of any memory depends on the ICs and PCB (brain) in the sticks. You can but DDR 1600 that will overclock 5% and you can buy DDR 1600 that will overclock 50%

 

[BoredErica]

Quote:

Originally Posted by billbartuska

Overclockability of any memory depends on the ICs and PCB (brain) in the sticks. You can but DDR 1600 that will overclock 5% and you can buy DDR 1600 that will overclock 50%

Based on that, slight deviation from the topic...

I've seen Gskill Trident and Ripjaws V memory both for Skylake platform, both have the same frequency and timings, and the Trident one costs more. Are they different in any way or is it only a branding thing?

 

[Slink3Slyde]

I know, but most times you will need to add voltage I believe., I do have some experience overclocking RAM but I lack detailed technical knowledge to explain the why's

I was taking a stab based on the fact most current ddr 3 kits at 2133 and over need 1.65 volts at stock, I believe there are some expensive DDR3L kits that will do 2133 C10 at 1.35 volts, based on that 2666 24/7 might be possible I suppose. But l think at that point it's probably cheaper to buy DDR4 3200+ which would be faster for Skylake anyway.

It depends on the reason for ops question really.

Edit: I think that Trident generally have better ICS then Ripjaws, and Ares is more budget then both I think.

 

[damric]

I don't know but here's my latest DDR3 test with Skylake....if anyone has anything ot compare it to...

 

[BoredErica]

Quote:

Originally Posted by Slink3Slyde

I know, but most times you will need to add voltage I believe., I do have some experience overclocking RAM but I lack detailed technical knowledge to explain the why's

I was taking a stab based on the fact most current ddr 3 kits at 2133 and over need 1.65 volts at stock, I believe there are some expensive DDR3L kits that will do 2133 C10 at 1.35 volts, based on that 2666 24/7 might be possible I suppose. But l think at that point it's probably cheaper to buy DDR4 3200+ which would be faster for Skylake anyway.

It depends on the reason for ops question really.

Edit: I think that Trident generally have better ICS then Ripjaws, and Ares is more budget then both I think.

ICS as in ICs or 'integrated circuits'?

Ram's a toughphie because it's hard to tell if an application wants more frequency or prefers a good mix of low latency as well... especially if that application is not used often or benchmarked often.

16gb 4000 C19 for like $210 now, prices have really dropped. Going to 4266 C19 costs $300 for only 8gb though. I calculated the performance index of 4000 C19 and 4266 C19 and 3200 C13, but at this point I think the performance index would be way off...

 

[Slink3Slyde]

Quote:

Originally Posted by Darkwizzie

ICS as in ICs or 'integrated circuits'?

Ram's a toughphie because it's hard to tell if an application wants more frequency or prefers a good mix of low latency as well... especially if that application is not used often or benchmarked often.

16gb 4000 C19 for like $210 now, prices have really dropped. Going to 4266 C19 costs $300 for only 8gb though. I calculated the performance index of 4000 C19 and 4266 C19 and 3200 C13, but at this point I think the performance index would be way off...

ICs the actual RAM chips on the PCB, I think Trident is Gskills premium line. The trident kits probably have tighter lower timings as well.

I'm of the opinion currently that you can use the index to give a rough guide as long as it's the same generation of RAM

 

[billbartuska]

Memory binning

http://www.maximumpc.com/exceleram-explains-ram-binning-process/

http://www.gamersnexus.net/guides/1140-silicon-die-bin-out-process-explained

 

[orlfman]

i've had the chance to use both corsair dominator platinum 3000mhz cas 15 kit and a gskill trident z 3000mhz cas 14 kit.

the trident z is essentially gskill's version of corsair's dominator platinum series. highest binned IC's go into their trident line similar to how corsairs highest binned IC's go into their dominator platinum line.