DDR2 Latency vs. Clock Speed

[imported_wicka]

Of course we all know by now that DDR2 latency is wildly higher than than DDR. However, my question is: where is the line drawn between buying lower latency or buying higher clock speed? For example, you could get CL3 DDR2 667, or CL5 DDR2 800, one of which is going to be faster. Both factors influence the speed, but at which point does an overwhelming gap in clock speed make the latency differences negligible?

 

[A554SS1N]

As far as I have seen in graphs, higher clockspeed on DDR2 always seems better, for example, 533Mhz at 3-3-3 would be slower than 667Mhz at 5-5-5, and 667MHz 3-3-3 would be a little slower than 800Mhz at 5-5-5 timings, or similar. Probably all due to the way that DDR2 works compared to DDR - seems less efficient per clockspeed to me.

 

[hans007]

uh no.

to the last one. this completely depends on the platform.

latency is measured in cycles. clock speed is an actual speed. a ddr2/533 cycle is slower than a ddr2-800 one.


533 cycle = 1.8ns or so
667 = 1.5ns
800 = 1.25 ns
400 = 2.5ns.


so basically cl3 ddr2-400 has the same latency as cl6 ddr2-800 would, and same as cl5-667 , cl4-533

on and intel platform, the bandwidth doesnt matter (well assuming no integrated video) if it exceeds the bus speed.


thus using dual channel, if you have an 800 bus intel chip, y ou only really need ddr2-400 to max that out, a conroe at 1066bus would need ddr2-533.

i would say the optimum is probably ddr2-667 cl4 which is pretty easy to get (almost all ddr2-667 cl5 will o/c to this at like 2.1 volts also). and ddr2-667 cl5 is pretty much mainstream memory now. you could probably do it with ddr2-533 cl4 + voltage also.

on amd it might make slightly more of a difference, though most people agree that the athlon is not memory constrained (i.e. memory bandwidth barely mattered between s754 and s939). so probably the same thing there.

 

[imported_wicka]

Originally posted by: hans007
uh no.

to the last one. this completely depends on the platform.

latency is measured in cycles. clock speed is an actual speed. a ddr2/533 cycle is slower than a ddr2-800 one.


533 cycle = 1.8ns or so
667 = 1.5ns
800 = 1.25 ns
400 = 2.5ns.


so basically cl3 ddr2-400 has the same latency as cl6 ddr2-800 would, and same as cl5-667 , cl4-533

on and intel platform, the bandwidth doesnt matter (well assuming no integrated video) if it exceeds the bus speed.


thus using dual channel, if you have an 800 bus intel chip, y ou only really need ddr2-400 to max that out, a conroe at 1066bus would need ddr2-533.

i would say the optimum is probably ddr2-667 cl4 which is pretty easy to get (almost all ddr2-667 cl5 will o/c to this at like 2.1 volts also). and ddr2-667 cl5 is pretty much mainstream memory now. you could probably do it with ddr2-533 cl4 + voltage also.

on amd it might make slightly more of a difference, though most people agree that the athlon is not memory constrained (i.e. memory bandwidth barely mattered between s754 and s939). so probably the same thing there.

So what about this: http://www.newegg.com/Product/Product.asp?Item=N82E16820145015

DDR2-675 CL4

 

[Suspicious-Teach8788]

I think now that AMD and Intel both use DDR2, you need to consider both architectures.

With AMD, we showed that bandwidth is not a huge issue and there is more than enough, but what about Intel?

I remember when OCForums had a post about DDR 500 2-2-2-5 vs DDR600. Basically it was like top of the line BH5 chips (VX or Redline) versus top of the line TCCD (like OCZ 4800 Platinum or Geil One TCCD) and the 600 won.

Timings matter to a certain point, and so you need to test. If you can do both timings and clock speeds, you should get that, but usually clock speeds are a lot more important simply because you're so limited in speeds when you consider timings.

 

[BitByBit]

Originally posted by: hans007
uh no.

to the last one. this completely depends on the platform.

latency is measured in cycles. clock speed is an actual speed. a ddr2/533 cycle is slower than a ddr2-800 one.


533 cycle = 1.8ns or so
667 = 1.5ns
800 = 1.25 ns
400 = 2.5ns.


so basically cl3 ddr2-400 has the same latency as cl6 ddr2-800 would, and same as cl5-667 , cl4-533

on and intel platform, the bandwidth doesnt matter (well assuming no integrated video) if it exceeds the bus speed.


thus using dual channel, if you have an 800 bus intel chip, y ou only really need ddr2-400 to max that out, a conroe at 1066bus would need ddr2-533.

i would say the optimum is probably ddr2-667 cl4 which is pretty easy to get (almost all ddr2-667 cl5 will o/c to this at like 2.1 volts also). and ddr2-667 cl5 is pretty much mainstream memory now. you could probably do it with ddr2-533 cl4 + voltage also.

on amd it might make slightly more of a difference, though most people agree that the athlon is not memory constrained (i.e. memory bandwidth barely mattered between s754 and s939). so probably the same thing there.

DDR533 certainly does not have a latency of 1.8ns!

This feels like the 50th time I've responded to a question / erroneous post regarding memory latency. But what the hell.

Memory latency in nanoseconds = CAS Latency (in cycles) / internal clock speed.
So a DDR533 module, with a CL of 3 cycles and an internal clock of 266MHz has a latency of:

3 / 266E+06 = 11.2ns

To calculate the latency in ns of a DDR2 module, follow exactly the same process. Although DDR2 runs at half the internal clock of equivalent DDR, its CAS latency takes account of that.
A DDR2 533 module with a CL of say, 5 cycles, has a latency of:

5 / 266E+06 = 18.8ns