How real is CL / MHz * 1000 ?

[Lee Saxon]

I've read that in trying to estimate the interaction of latency and clock speed for memory this formula can be used: CL / MHz * 1000 = best responsiveness in milliseconds

But how far / how seriously can you really take that? I don't have trouble believing that DDR4-3733 with CL17 could be faster than DDR4-3800 with CL19, for example, but it's hard to accept that DDR4-2400 CL10 and DDR4-3600 CL15 perform the same.

If you filter DDR4 by CAS Latency at Newegg, take the fastest clock speed available for each, and rank by the formula you get an order about which I'm fairly skeptical:

CL19 DDR4-4600 - 4.13
CL10 DDR4-2400 - 4.17
CL15 DDR4-3600 - 4.17
CL14 DDR4-3200 - 4.38
CL16 DDR4-3600 - 4.44
CL18 DDR4-4000 - 4.50
CL17 DDR4-3733 - 4.55
CL13 DDR4-2666 - 4.88
CL12 DDR4-2400 - 5.00

And hell, that'd put my current CL8 DDR3-1600 from 2011 still at a pretty competitive position.

 

[Campy]

Just because they have the same latency doesn't mean they have the same performance. latency and bandwidth are two different metrics of performance. Over the years bandwidth has increased massively while true latencies have stayed roughly the same.

 

[eton975]

Will be in nanoseconds, but essentially, yes if you mean CL/Mhz (so ie. 17/3733*1000).

 

[dullard]

it's hard to accept that DDR4-2400 CL10 and DDR4-3600 CL15 perform the same.

CL10 DDR4-2400 - 4.17
CL15 DDR4-3600 - 4.17

And hell, that'd put my current CL8 DDR3-1600 from 2011 still at a pretty competitive position.

Like the others have said, the formula is basically correct*, but you are not capturing the entire picture. Yes, both CL10 DDR4-2400 and CL15 DDR4-3600 have the same latency until the first data point is ready to be read from the memory module. But, if you read sequential data from the memory, then after that latency delay CL15 DDR4-3600 can deliver 50% more data in a given amount of time as the CL10 DDR4-2400.

Your CL8 DDR3-1600 might be somewhat competitive with the memory on that list if you only need one piece of data from any given memory area (random access memory). But, if you need sequential data (which is quite common), then the CL19 DDR4-4600 would get you the first piece of data 17% faster but also be able to deliver nearly 3x as much data in a given amount of time after that latency delay. Your memory is not even remotely close to competitive for sequential memory access.


* DDR memory gives data at both the leading edge and the trailing edge of the clock signal (the first D stands for Double data rate). But latency is only at one of those edges. So the real formula for DDR is CL / MHz * 2000 in nanoseconds. So all of your calculations are off by a factor of 2.

 

[dullard]

For example, suppose you want to manipulate a 10 MB image that is stored sequentially in memory. Lets compare your CL8 DDR3-1600 to the best memory on that list, the CL19 DDR4-4600.

The CL8 DDR3-1600 will give your processor the ability to read the first pixel information at 8/1600*2000 = 10 ns. The last pixel information will arrive at that 10 ns + 10 MB / (1600 * 8 MB/s) * (10^9 ns / 1 s) = 781260 ns = 0.78 ms. The motherboard + CPU will take its own time to read, transfer, and process the information, but the memory is ready with the data at 0.78 ms.

The CL19 DDR3-4600 will give your processor the ability to read the first pixel information at 19/4600*2000 = 8.26 ns. The last pixel information will arrive at that 8.26 ns + 10 MB / (4600 * 8 MB/s) * (10^9 ns / 1 s) = 271747 ns = 0.27 ms. The motherboard + CPU will take its own time to read, transfer, and process the information, but the memory is ready with the data at 0.27 ms.

 

[kjboughton]

https://www.anandtech.com/show/3851...now-about-sdram-memory-but-were-afraid-to-ask