What is good RAM timing?

Aeraylia_

Honorable
Sep 14, 2014
180
0
10,680
The G.SKILL Ares Series 16GB (2 x 8GB) RAM has a timing of 11-13-13-31. I saw on a forum that a guy had RAM and the timing was "1600 9-9-9" is mine faster or slower? How does the timing work? Thanks :)
 
Solution


It's not always the case, but generally it is. It's more so, because of pricing... You can get 9-9-9 16GB (I don't know of any though) but i't'd be more expensive, and a waste of money frankly.
Smaller is better.

(CAS / Frequency (MHz)) × 1000 = X ns

(9 / 1600) x 1000 = 5.63 nanoseconds

(11 / 1600) x 1000 = 6.88 nanoseconds

While RAM speed / timings is oft poo poo'd as having no real difference, is supported only when ya don't look to deeply or compare relative price / performance ratios inappropriately. What this comes from is misinterpretations of RAM reviews whereby the benefit of faster RAM is judged against the relative costs. Example....

New RAM platform comes out and reviewer looks at two sets .... 1600 CAS 9 and 2400 CAS 10. he sees that the increase in average fps is only 0 - 5 % whereas the difference in cost of the two sets is such that one is twice the cost of the other . The conclusion is drawn that the improvement is not worth the increase in cost and this conclusion will be parrotted in forums for years. Is that valid now... and if not why so ?

1. When a new RAM platform comes out , yields on the hi spec stuff is very low and because it is rare, the high end stuff will draw a high premium. Years later this premium disappears.... 2133 CAS 9 right now is oft the same cost or no more then $3-4 more than 1600.

2. The RAM cost is red herring in that it's not just your RAM that goes faster, it's your whole system that sees the benefit so the cost of the build is what matters.

3. Average fps is not where we see the effect of faster RAM, the real impact is in minimum fps.

4. If your GFX card is the limiting speed factor, the impact of faster RAM might not show up till you go to multiple cards.

5. The impact may be significant in one game and 0 on another. Obviously in demanding applications like video editing, CAD, database manipulation and such this claim is no longer made much.

So let's look at this:

$155 Gskill Trident 16 GB 24500 CAS 10
$130 G.Skill Ripjaws X Series 16GB

Cost ratio is 20% and no one would argue that it's 19% faster. However, the 970 SLI system with 4690k came in at $1725 and adding $25 and that's a 1.4 % price increase. Will we ever see a 2% or more increase in what we do on our puters ? Most certainly yes.

http://www.tomshardware.com/reviews/32-gb-ddr3-ram,3790-10.html

Here we see an 11% difference in gaming performance in F1 ... in Metro we see virtually nothing. That represents the two extremes. More typically the differences range from 2 -5 % and while that might certainly be describes and not very significant, the 1.4% difference in system cost is even smaller. And before anyone mentions IGP, the test was done with a 4770k and 290x

Here's some test data where they looked at minimum rather than average frame rates

http://www.anandtech.com/show/2792/12

22.3 % (SLI) increase in minimum frame rates w/ C6 instead of C8 in Far Cry 2
18% (single card) / 5% (SLI) increase in minimum frame rates w/ C6 instead of C8 in Dawn of War
15% (single card) / 5% (SLI) increase in minimum frame rates w/ C6 instead of C8 in World in Conflict

Also see http://www.bit-tech.net/hardware/memory/2011/01/11/the-best-memory-for-sandy-bridge/1

In short..... when a new platform drops and prices for the high spec stuff is high, it's hard to make a recommendation to get faster RAM for gaming boxes. For tough applications, especially in a production environment, it's not hard at all. However, once a platform has matured as we are today, buying 2133 CAS 9 for $74 over 1600 CAS 9 for $71 is the proverbial "no brainer". As to whether to "go higher", 2400 is a reasonable price premium for a high end gaming box; higher than that it's must tougher to justify.



 
  • Like
Reactions: ReveurGAM


It's slower, but if you're getting 16GB instead of 8GB, or just a bit cheaper 16GB then I wouldn't worry too much. It has no noticeable impact in real world tests.
 


So because I have 16GB instead of 8GB (which he had) I will have longer timing because it's bigger?
 


It's not always the case, but generally it is. It's more so, because of pricing... You can get 9-9-9 16GB (I don't know of any though) but i't'd be more expensive, and a waste of money frankly.
 
Solution


Yeah I see your point. It would be pointless to pay more money for a few more nanoseconds haha thanks! :)
 


No problem 😀. Any more questions do ask them, that's what we're here for!
 


No, you can get 1600 in just about any size down to CAS 6 and up to 11. The 6 will generally be the most expensive and the 11 will be the least expensive.

You have longer timings because as RAM comes off the production line not every module is equal. Just like some CPUs OC well and some don't, some RAM will pass 1600 CAS 9 specs and some won't..... what fails at 1600 CAS 9 will be tested at CAS 10 and if it fails will be tested as CAS 11.... if it passes, it will be sold as DDR3-1600 CAS 11 but will garner a lower price than 1600 CAS 9.

See the formula in my post above and you can get a good understanding of CAS here;

http://en.wikipedia.org/wiki/CAS_latency

At 1600 CAS 9 is by far sold more than any other timings. Looking in newegg, the GSkill CAS 9 has generated 418 reviews to just 36 for the CAS 11 which is indicative of the CAS 9 selling at 11;1 over the CAS 11


And it is by no means pointless to ante up the minuscule price difference (test results linked above show that) ... prices fluctuate depending upon how much the market has of each you could have had 2133 CAS 9 for between $0 and $10 more.
 


If I could add just a couple things here, the 1600 in DDR3 1600 is not the frequency, it's the Data Rate in Mega Transfers / second. All DDR3 1600 sticks run at the same clock settings, the I/O memory bus clock (Command Rate) is 800Mhz, the internal clock (the stick itself) runs at 200Mhz, which gives you a cycle time of 5ns.

So the first and eighth word time for CL11 would be 13.75ns and 18.13ns respectively, for CL9 your looking at 11.25ns and 15.63ns. Your millage may vary depending on the word burst allowance being used in a transfer, so these are just basic timing numbers.

CL 8/9 is pretty good timing for 1600 sticks, a bit faster than 11.

If you want to read up a bit on how DRAM works as far as data/timing rates here are a few links I use for reference....enjoy :)
https://en.wikipedia.org/wiki/Memory_timings
https://en.wikipedia.org/wiki/CAS_latency#cite_note-eighthword-6
https://en.wikipedia.org/wiki/DDR3_SDRAM
https://en.wikipedia.org/wiki/SDRAM_latency
https://en.wikipedia.org/wiki/Transfer_%28computing%29

And a good read on what DDR4 is all about with respect to DDR3 and future DRAM configurations...
 
  • Like
Reactions: ReveurGAM
Yes, that is well described in the link I posted (Edit ... oops didn't paste in) but doesn't make anything else less valid.

http://en.wikipedia.org/wiki/CAS_latency

http://en.wikipedia.org/wiki/DDR3_SDRAM

As with earlier memory generations, faster DDR3 memory became available after the release of the initial versions. DDR3-2000 memory with 9-9-9-28 latency (9 ns) was available in time to coincide with the Intel Core i7 release.[10] CAS latency of 9 at 1000 MHz (DDR3-2000) is 9 ns, while CAS latency of 7 at 667 MHz (DDR3-1333) is 10.5 ns.

(CAS / Frequency (MHz)) × 1000 = X ns

Example:

(7 / 667) × 1000 = 10.49475 ns



 
Just to add a bit more to this DRAM discussion, here are a couple comments from Pinhedd on another thread we talked about it. Memory is sort of complicated to nail down (for me anyway) just how fast it is going to run for you, there are many variables. Just to illustrate this Pinhedd took the time to point this out here.............
""
You're both wrong.

DRAM timings are measured in clock cycles, not bus transfers.

DDR3-2133 is not 2133Mhz SDRAM, it is 1066Mhz SDRAM. Both the command clock (which synchronizes the commands) and strobes (which synchronize the data transfers) are periodic and oscillate at nominally 1066Mhz. Data is transferred onto the bus with the rising and falling edge of the strobe for two transfers per cycle, or double-data-rate.

A 1066Mhz clock has a clock period of 0.9381 nanoseconds, or 938.1 picoseconds. A DDR3-2133 module with a nine cycle column latency (CAS9) will have to wait nine cycles between the issuance of a read command and the first word of the resulting burst being stable on the IO bus (accompanied by a strobe); the data on the IO bus will change ever half cycle until all eight words of the burst have been transferred.

In this case

Tcl = 9 * 0.9381 = 8.443 nanoseconds

Tburst = 4 * 0.9381 = 4.221 nanoseconds.

The first word of the burst is stable after 8.443 nano seconds, and the last word is stable after 11.72 nanoseconds. At 12.664 seconds the first word of the next read command (if one is issued) will be sent onto the IO pins.

In the grand scheme of things, CAS latency is almost completely inconsequential. A good memory controller can keep the IO bus busy nearly 100% by pipelining commands and interleaving banks.
""
and
""
A good memory controller can keep the IO bus busy nearly 100% of the time. There's always going to be a few cycles in which no useful operation can be performed (such as when selecting a new rank when the command rate is 2T or greater) and others in which the contents of the transaction queue just simply don't line up with the open pages or the timings don't permit the operation to complete within the time window.

Given the tough to analyze and non-deterministic nature of DRAM it's usually sufficient to analyze performance on the basis of theoretical peak transfer rate. 2133 megatransfers per second * 64 bits per transfer / 8 bits per byte = 17.064 gigabytes per second per memory channel. This works out to about 5.88 milliseconds per gigabyte on average at a minimum. If they were all sequential (they wouldn't be, they would span multiple rows) there would be a Tcl delay between the command to read the first address and the first word, and a Tcl + Tburst delay between the command to read the last address and the last word. There would also be a Tccd delay between read operations.
The reason for this is that it's possible to pipeline column operations. It is possible to issue a column read operation to a bank that has another column read operation in progress provided that a timing constraint is met. This timing constraint is the column command delay or Tccd. Tccd is lower bounded by Tburst, meaning that Tccd cannot be lower than 4 cycles (or 4Tck), but on certain memory architectures it may be greater than 4 cycles. This allows one burst to begin immediately after another burst finishes.

Now, 17.064 gigabytes per second for DDR3-2133 is a theoretical maximum and would be perfectly valid if DRAM/SDRAM had the same behavioural mechanics as SRAM/SSRAM. However, while SRAM has deterministic latency, SDRAM does not. The throughput of SDRAM is heavily dependant on the temporal and spatial locality of the data that's being accessed.
""

I agree with JackNaylorPE here, spending a bit more for CL9 over CL11 is money well spent. If your going to OC your CPU I would even consider 1866 or 2133, something in 1.6v...
 
Yes, I have shared many discussions with Pinhedd and I have walked away learning something each time. But for the most part .... I tend to pay more attention to tested performance results than the "how things work". I use the formula for weighing what's in the same ballpark are likely to bring the best return...for example:

2133 CAS 9 gives 4.22 ns
2400 CAS 10 gives 4.17 ns

That implies they are close but is 1.1% worth $30 ? Looking at the THG game results in the link above showed better numbers with a 3.4% advantage of the same modules at 2400 versus 2133 in gaming. Kinda like overclocking ... do ya want the highest P95 stable overclock ? .... or do ya want the highest overclock that is stable in something you are actually doing ?