Question Memory Latency result in MemTest86 a good measure of RAM performance?

rasmasyean

Distinguished
Mar 15, 2008
264
1
18,795
I ran some MemTest86ing on 2 computers.

MSI MAG Z590 TOMAHAWK WIFI LGA 1200 ATX Intel Motherboard - Newegg.com (i9-11900K)
ASUS PRIME Z490-A Intel Z490 ATX Intel Motherboard - Newegg.com (i9-10850K)

I found it surprising that the "Memory Latency" is higher in the newer MoBo/CPU. (like 23 vs 20). Thats seems like a lot.
Is this a thing where it depends on how well the model of RAM interacts with a specific motherboard/controller?
Does it depend on kit lottery and/or combination of sticks in different orders?

What is the significance anyway? Can you interpret his as being slower by 15%?

Thanks!
 
Do you have XMP enabled on both boards? Is the memory installed in the A2 and B2 slots on both boards?

If these are those same Neo Forza kits from your other thread then they should be the same EXACT CL latency on both boards unless they are not in fact the same model of memory kit or unless there is something wrong with one of the kits which might be why they didn't want to play nice together.

Take screenshots of the SPD and Memory tabs while running CPU-Z on each board, for all four DIMMs. So, install one kit in one board, enable XMP, run CPU-Z and take screenshots. Then do the same thing with the other kit, in the SAME board, after removing the first kit.

Then, do the same thing with both kits in the other board. Post the screenshots here.
 

rasmasyean

Distinguished
Mar 15, 2008
264
1
18,795
The "computer" or "motherboard" should have ZERO to do with the "first word" or "true" latency.

This is how you determine true latency. I don't know what Passmark is relating as "memory latency" but for the same kit it can't change just because you put it in a different board.

1 / Frequency (not DDR) x Latency = True Latency (nanoseconds).

Stock 3200 @ 14 is faster than Stock 3600 @ 16:

1 / 1.600GHz x 14 = 8.75nS
1 / 1.800GHz x 16 = 8.89nS

Stable Overclock with 3733 @ 16 is faster:

3733 @ 16 is 1 / 1.867GHz x 16 = 8.57nS

By calculating True Latency, when experimenting with memory overclocking, it becomes much easier to determine with greater consistency, where you might expect to find that elusive fuzzy-grey edge of stability.

Accept for applying excessive or prohibitive voltages above 1.35, which can potentially damage your processor's IMC, even with the best chips, any Frequency / Timing combinations that result in True Latencies below about 8.5'ish to 8.4'ish may be unstable, or unbootable.
 
Right. So has nothing to do with ACTUAL latency based on the memory itself. The actual True Latency is entirely dependent on the memory speed and timings and is independent of any outside factors.

That's assuming same frequency and timings though. And technically, true latency isn't 100% representative of actual memory performance because how a specific board auto configures some of the secondary and tertiary timings can obviously have a direct impact on memory performance which then can have an impact on overall performance BUT the only way you are usually going to notice most of those variables is through synthetic bench tests. In real world applications, you'd never know the difference unless you've made significant changes to the frequency or main timings. And even then, it's unlikely you'd notice anything majorly different unless it's a very memory dependent operation.
 

rasmasyean

Distinguished
Mar 15, 2008
264
1
18,795
OK, so let me get this straight. The term "Latency" is specifically deemed to be CYCLE Latency (CL).
And the OC'ed time numbers (ns) you're converting them to can be too short such that it does something like:
Make the 1's and 0's rise and fall before the rest of the transistors can catch up.

The "Memory Latency" in Passmark should be named something like "Average Fetch Delay Time"" or something to that effect to distinguish itself from RAM Latency in industry parlance.