What’s the differency between memory real latency 12ns vs 11.25ns?

[NDRE28]

Hello! 

I’m curious about the perceived difference between these 2 memory kits: 

 

1. DDR4-2666 CL16-18-18 @1.2V (real latency: 12ns)

 

2. DDR4-3200 CL18-21-21 @1.2V (real latency: 11.25ns) 

 

The 2nd kit has a speed advantage of -0.75ns (6.25%) over the 1st kit. 

 

QUESTION: 

6.25% speed advantage translates to how much overall, in tests? 

(4%? 2%? 1%?)

[Mira Yurizaki]

Assuming my math is correct, in order there to be an accumulation of one second lost using the slower RAM, you're looking at 1,333,333,333 requests. Assuming a worst case scenario of requesting only 4 KiB pages, this is about 5TB of data.

 

Basically, you're looking at almost zilch with regards to speed advantage.

[NDRE28]

10 hours ago, Mira Yurizaki said:

Assuming my math is correct, in order there to be an accumulation of one second lost using the slower RAM, you're looking at 1,333,333,333 requests. Assuming a worst case scenario of requesting only 4 KiB pages, this is about 5TB of data.

 

Basically, you're looking at almost zilch with regards to speed advantage.

I’m afraid that your answer ain’t clear enough for me. 

Can you, please, be more specific?

[Mira Yurizaki]

1 minute ago, NDRE said:

I’m afraid that your answer ain’t clear enough for me. 

Can you, please, be more specific?

What about my answer isn't clear?

[GoldenLag]

16 minutes ago, NDRE said:

6.25% speed advantage translates to how much overall, in tests? 

it depends. 

[Firewrath9]

depends.

on the uArch and what workload you are using. for example, if you are using a memory bandwidth heavy application (i.e. idk AIDA64 disk and cache benchmark), it might make a 1-3% difference, might be more on ryzen 1st/2nd gen, less on intel. the bandwidth is major difference. The only time you should start to tighten timings is after you get to the highest possible frequency (with loosening timings, 4600mhz CL24 is better than 3200mhz CL16) or 3733mhz on ryzen 3rd gen.

subtimings don't really help much, even on memory and cache benchmarks.

[NDRE28]

10 hours ago, Mira Yurizaki said:

What about my answer isn't clear?

I was asking for percentages in benchmarks. 

(I’m on Intel platform).

[NDRE28]

10 hours ago, GoldenLag said:

it depends. 

I’m on Intel platform. 

I’m interested into knowing the speed differences, in percentages, in benchmarks. 

[Mira Yurizaki]

1 minute ago, NDRE said:

I was asking for percentages in benchmarks. 

(I’m on Intel platform).

It depends on what test they're doing and how long it's being done. If it's a bandwidth test, there's no advantage for using low latency RAM. If it's a "how fast can this RAM service requests", if it's only doing something like 1000 requests, the delay is 750ns on the higher latency RAM, assuming no other factors are involved. This is imperceptible as far as human timescales are concerned.

 

But if you want a stupid simple answer:

• Raw bandwidth test: there's no advantage
• Servicing requests tests: as high as 6.25% better for the lower latency RAM.

[NDRE28]

11 hours ago, Mira Yurizaki said:

It depends on what test they're doing and how long it's being done. If it's a bandwidth test, there's no advantage for using low latency RAM. If it's a "how fast can this RAM service requests", if it's only doing something like 1000 requests, the delay is 750ns on the higher latency RAM, assuming no other factors are involved. This is imperceptible as far as human timescales are concerned.

 

But if you want a stupid simple answer:

• Raw bandwidth test: there's no advantage
• Servicing requests tests: as high as 6.25% better for the lower latency RAM.

OK. I think I get it! 

With other words there’s no perceptible advantage to move to the higher MT/s kit...

[Mira Yurizaki]

1 minute ago, NDRE said:

OK. I think I get it! 

With other words there’s no perceptible advantage to move to the higher MT/s kit...

MT/s does matter depending on which platform you go with. AMD Ryzen systems require higher MT/s for optimal performance than Intel's.