DDR4 CAS Latency

[buildzoid]

That's a non-answer.



Non-point as in what? The OP is obviously wondering whether DDR4 will perform better or not in the end. There's no "crappiness" of high speed kits.

A 2933 kit WILL perform better than a 2400 kit regardless of it's higher timings. Period.

sorry but no due to how 2933+mhz ICs are designed they are hopeless when it comes to real world performance if this wasn't the case pro overclockers would use them for benchmarking not just ram frequency suicide runs. Dave even reviewed a 2933 kit and it got destroyed in everthing by a 2400 kit. Timings matter and frequency does to but at somepoint the increase in time spent finding and accessing data negates the speed at which the data travels. This is typically the case with ddr3 ram going over 2800mhz.
As for op's question just look at some haswell-e superPi scores compared to haswell superPi scores and you can see that ddr4 2666 14-10-10-28 is pretty much on par with ddr3 2133 8-10-8 and ddr3 2800 9-10-10.

 

[EarthDog]

Make sure you are looking at 32M runs to see the effect on ram. While 1M does respond to faster memory speeds, it is more or less handled all in cache. So 32M will give you a better idea.

 

[Mike Rollins]

Your math is accurate for the old SDRAM but for all DDR SDRAM you must multiply by 2000.
Therefore, in your examples:
2133MHz CL15 DDR4 memory, ( 15 / 2133 ) * 2000 = 14.06ns
1600MHz CL9 DDR3 memory, ( 9 / 1600 ) * 2000 = 11.25ns


Actually, 2400-CL9 is faster than 2933-CL12 (the only CL I can find)...
DDR3 2400-CL9 = 7.50ns (~$270US for a 16G Kit)
DDR3 2933-CL12 = 8.18ns (~$700US for a 16G Kit)
DDR4 3000-CL15 = 10.00ns (~$390US for a 16G Kit)

As for now, the only reason to go to DDR4 is for the low power draw (1.2V vs 1.65V).
Give it some time though and we'll see DDR4 prices drop and speeds soar.
Eventually, they'll get up to DDR4 4200-CL15 (7.14ns) using only 1.35V.

I know this post is dated, but I ran across it since I'm researching some new memory to purchase.

The example you used here to calculate the speed, which you said is more accurate and the formula used previously, towards the beginning of the post (using the divisor as the stated speed cut in half and then multiplying by 1000), which you said was inaccurate because it's for old SDRAM. Did you even stop to think more into this formula you said was correct while bashing the other (using the stated memory speed as the divisor and then multiplying it by 2000 instead of 1000)? If you stopped and thought about it for a second, when you increase the divisor by 2x and then increase the multiplier from 1000 to 2000, it would be the exact same formula. That would be like saying instead of the answer being (15/1500)*1000=10ns vs. (15/3000)*2000=10ns. Just my $0.02.

 

[Lagittaja]

I know this post is dated, but I ran across it since I'm researching some new memory to purchase.

The example you used here to calculate the speed, which you said is more accurate and the formula used previously, towards the beginning of the post (using the divisor as the stated speed cut in half and then multiplying by 1000), which you said was inaccurate because it's for old SDRAM. Did you even stop to think more into this formula you said was correct while bashing the other (using the stated memory speed as the divisor and then multiplying it by 2000 instead of 1000)? If you stopped and thought about it for a second, when you increase the divisor by 2x and then increase the multiplier from 1000 to 2000, it would be the exact same formula. That would be like saying instead of the answer being (15/1500)*1000=10ns vs. (15/3000)*2000=10ns. Just my $0.02.

I know this post is dated, but I ran across it since I'm researching DDR4...

Actually Laytonjnr's calculations were incorrect and JoshuaAJones's was correct..
Laytonjnr calculated using
(CL / Data rate) * 1000
Which is correct for SDR SDRAM but incorrect for DDR SDRAM.

What we're calculating here (for DDR) can be achieved using for example:
a)
(CL / (Data rate / 2)) * 1000
b)
(CL / IO Bus Clock) * 1000´
c)
(CL / Data rate) * 2000
or whatever variation you can think of.

JoshuaAJones calculated using c which gives the same results as a and b.

(15/(2133/2)*1000 = 14,06ns
(15/1066,5)*1000 = 14,06ns
(15/2133)*2000 = 14,06ns

(15/2133)*1000 = 7,03ns

 

[sneekypeet]

In case you like pictures of how it all correlates... http://www.crucial.com/usa/en/memory-performance-speed-latency

 

[Aquinus]

Just for clarity, there are cases where things big sequential reads offer improved performance because some latencies are omitted because of the kind of change in address that occurred. For example, it takes fewer cycles to go from column to column as opposed to row to row. None of this really matters in the end though because most workloads can buffer memory read/write requests in cache. The biggest win here is writeback cache. Where active memory isn't written to immediately but (in terms of modern Intel CPUs,) will get written to L3. The CPU will acknolege a successful memory operation once it has been written to L3 but, in reality it's being queued up to be written to main memory. Meanwhile, between now and then, that data is accessible by all of the cores and attached cores (via QPI on MP systems,) without touching system memory. This caching concept starts to degrade when writing a lot of data to system memory, faster than system memory can be written to but, so long as the IMC has overhead is within that capabilities of the last level of cache accessible to the cores, memory bandwidth and latency means very little and it is in fact the speed of the SRAM cache in the CPU than can have a greater impact on performance.

All in all, memory has a point of diminishing returns, just as cache does because in the end, you can only feed a core so much data and you can only shove so many cores on a die and software can only be made to be so parallel. It's the nature of computers; the bottleneck problem.

 

[Mike Rollins]

Wow! This thread continues to revive itself. Well, since I posted here almost a year ago, I sold my Corsair Vengeance LPX 2800 CL16 DDR4 memory and bought Kingston HyperX 2400 CL12 DDR4 memory, and I must say, it worked out better than the Corsair. But, I think there may have been some incompatibility issues with the Corsair LPX memory and the Asus X99 Deluxe MB, which I still have but it's on the shelf while I work on a new build.

With the X99 chipset, anything over 2666Mhz you had to increase the BCLK to 125 from 100, actually 127.3 BCLK was the sweet spot. Running a higher BCLK presented other problems, so I downgraded my speed and got a faster latency. Since my rig is torn apart and has been since Sept, I can't comment on how my HyperX memory runs on my new Rampage V Extreme. Who knows, maybe in another year, someone else will post on this thread and say, "What are you talking about? RAM? We're all using 3D Xpoint memory."

 

[dr_lucas]

What you are referring to there is known as RAM 'access time', or more casually as RAM 'speed' or 'performance', which is basically the time it takes in nanoseconds for the RAM to locate a single piece of information and make it available to the processor (a very rough definition). Latency is a reference to CAS timings, which refers to the delay between the memory controller telling the RAM to 'find' the information, and the information being made available.

The general way to work out RAM access time is: ( CL / Frequency ) * 1000.

So, for 2133MHz CL15 DDR4 memory, ( 15 / 2133 ) * 1000 = 7.03ns

And for 1600MHz CL9 DDR3 memory, ( 9 / 1600 ) * 1000 = 5.63ns

So actually, DDR3 memory is faster in terms of access time between these two examples (as AFAIK 2133MHz is meant to be the standard DDR4 RAM frequency). As many people have already said, DDR4 RAM probably will eventually have better access time than DDR3 in due time.

Layton

Crucial has an interesting article that doesn't really match this calculation:
http://www.crucial.com/usa/en/memory-performance-speed-latency

I am a bit confused...can anyone explain which one is correct?

 

[Lagittaja]

Crucial has an interesting article that doesn't really match this calculation:
http://www.crucial.com/usa/en/memory-performance-speed-latency

I am a bit confused...can anyone explain which one is correct?

Could you even bother to try and read the whole thread first?

And just in case you're lazy, you'll find your answer a few posts above of yours

I know this post is dated, but I ran across it since I'm researching DDR4...

Actually Laytonjnr's calculations were incorrect and JoshuaAJones's was correct..
Laytonjnr calculated using
(CL / Data rate) * 1000
Which is correct for SDR SDRAM but incorrect for DDR SDRAM.

What we're calculating here (for DDR) can be achieved using for example:
a)
(CL / (Data rate / 2)) * 1000
b)
(CL / IO Bus Clock) * 1000´
c)
(CL / Data rate) * 2000
or whatever variation you can think of.

JoshuaAJones calculated using c which gives the same results as a and b.

(15/(2133/2)*1000 = 14,06ns
(15/1066,5)*1000 = 14,06ns
(15/2133)*2000 = 14,06ns

(15/2133)*1000 = 7,03ns