Sign in with
Sign up | Sign in
Your question
Solved

Memory latency vs GB/sec

Tags:
  • Latency
  • Memory
Last response: in Memory
Share
October 5, 2014 6:42:07 AM

It is better to have bigger memory latency
than higher memory score in GB/s?

More about : memory latency sec

October 5, 2014 6:52:25 AM

At a specific clock speed lower latency is better. Usually higher clock speeds will outweigh lower latency. So DDR 1600 CL9 vs DDR 2400 CL17, the 2400 memory would have more throughput because of the higher clock speed even with the higher latency. But DDR 1600 CL9 vs DDR 1600 CL13, the CL9 would be slightly better.
m
0
l
a b } Memory
October 5, 2014 6:58:59 AM

the general rule is the lower the cl and the higher the speed, the better! however depending hardware that may slightly change. for instance, running an APU you'd better have higher speed ram as it will also be used as video memory! for a system though running on dedicated gpu the benefits of runnig lower cl are better!
m
0
l
Related resources
a b } Memory
October 5, 2014 7:02:26 AM

There's a nice little formula to (roughly) determine RAM response time:

Latency * 1000 / Frequency = Response time in ns

Lower is better.
m
0
l
a b } Memory
October 5, 2014 7:38:27 AM

Lower latency is certainly better at a given frequency. That being said, you are best off with memory that has a good mix of frequency and low cas L.

For example DDR3 1600 cas 9, is slower than DDR3 2400 cas 10 or even 11
m
0
l
a c 145 } Memory
October 5, 2014 3:14:25 PM

Panagio Tsaou said:
It is better to have bigger memory latency
than higher memory score in GB/s?


Higher data rate always trumps lower latency. In terms of total round trip time for a random read or random write, a few extra DRAM IO bus cycles adds very little to the total latency. Microprocessors and memory controllers are designed with this in mind and will easily mask it.

There are many benchmarks showing the statistically noticeable benefits of higher data rate memory. It's very handy for data compression and encoding across the board, somewhat useful for Crossfire and SLI configurations, and not particularly useful beyond DDR3-1600 for most other applications.

There are no benchmarks showing statistically noticeable benefits of lower latency. They all show either no effect (within a margin of error), or use flawed methodology.
m
0
l
October 10, 2014 6:43:30 AM

So lets say that i get a mem score of 7.94GB/s and a Mem Lat of 66,8 ns but then i get 7.85GB/s and 59,3 ns then what is better? the first benchmark at 112MHz FSB CL12 and the second at 108FSB but with CL11...
Not sure about the fist one tho...
m
0
l
October 10, 2014 7:16:24 AM

Panagio Tsaou said:
So lets say that i get a mem score of 7.94GB/s and a Mem Lat of 66,8 ns but then i get 7.85GB/s and 59,3 ns then what is better? the first benchmark at 112MHz FSB CL12 and the second at 108FSB but with CL11...
Not sure about the fist one tho...


That is a 1.1% variation. Not significant in real world terms.
m
0
l
October 10, 2014 12:01:26 PM

kanewolf said:
Panagio Tsaou said:
So lets say that i get a mem score of 7.94GB/s and a Mem Lat of 66,8 ns but then i get 7.85GB/s and 59,3 ns then what is better? the first benchmark at 112MHz FSB CL12 and the second at 108FSB but with CL11...
Not sure about the fist one tho...


That is a 1.1% variation. Not significant in real world terms.


So that means that it is 1.1% overall lost performance?
Like 100fps going to 99 or like 1s faster at 100 seconds rendering?
m
0
l

Best solution

a c 145 } Memory
October 10, 2014 2:49:04 PM

Panagio Tsaou said:
kanewolf said:
Panagio Tsaou said:
So lets say that i get a mem score of 7.94GB/s and a Mem Lat of 66,8 ns but then i get 7.85GB/s and 59,3 ns then what is better? the first benchmark at 112MHz FSB CL12 and the second at 108FSB but with CL11...
Not sure about the fist one tho...


That is a 1.1% variation. Not significant in real world terms.


So that means that it is 1.1% overall lost performance?
Like 100fps going to 99 or like 1s faster at 100 seconds rendering?


Not at all. 7.94GB is what the synthetic test measured (this value is actually kinda low). Performance of real world applications is rarely memory bound.
Share
!