Ryzen: if can't run full (3200MHz) speed, is DDR4-3000 CL15 faster than DDR4-3200CL16 b/c of the CL?

[wchang99]

With Ryzen, if for whatever reason you can't run your RAM at max (3200MHz) speed, would it actually be better to buy DDR4-3000 CL15 than DDR4-3200 CL16, because both would be running at 2933MHz or less, but the DDR4-3000 CL15 would be running at CL15 while the DDR4-3200 CL16 would be running at CL16?

Or should you be able to set the DDR4-3200 CL16, if it's at <=2933MHz, to run at CL15, and have it run just as easily and stably as if it were at its normal 3200MHz CL16? DDR4-3200 CL16 costs slightly more than DDR4-3000 CL15, which leads me to think it's slightly higher quality memory and that its abilities might be inclusive of what the DDR4-3000 CL15 can do (i.e. the DDR4-3200 CL16 could run equally well at <=2933MHz CL15 as at 3200MHz CL16, since the DDR4-3000 CL15 can do so).. but I have no idea if it actually works that way.

Does anyone know? (Apologies for the noob question... I did try, and Google skills failed me.) Thanks.

 

[JimmiG]

Ryzen only supports even CAS latencies, so only C14, C16, C18 etc. You would have to run the 3000 C15 memory at 2933 C16 (or C14, if you manage to get it stable). 3200 C16 is probably a slightly better bin than 3000 C15.

New BIOS/AGESA updates are improving compatibility, but it still pays off to get the best memory you can afford. It's more likely to actually work at the advertised speed, and if it doesn't, it will still probably work at a higher speed than a lower bin would have.

Higher frequencies and loose timings usually perform better with Ryzen than low frequencies and tight timings. E.g. 3200 C18 would perform better than 2933 C14 in most cases. The advertised spec is just what the RAM was tested at, so there's nothing stopping you from trying to run 3000 C15 RAM at 3200 C16 or C18, for example (as long as it's stable).

 

[elpokor]

TLDR: think in nanoseconds for real-world latency comparation.

3000Mhz = 3000*1000² clock cicles per second at double data rate (DDR4), so clock cicle=1/(1500*1000²) seconds= 0.6666 nanoseconds. Then simply multiply by CAS latency to get the actual "total completion time", which is the total time it takes to perform a memory access and get ready for the next one. Your examples:

3000Mhz CL15: 1/(1500*1000²)*15 = 10 nanoseconds total completion time
3200Mhz CL16: 1/(1600*1000²)*16 = 10 nanoseconds total completion time

so those settings are effectively equal in terms of latency. Guess what would happen with 3600Mhz CL18 or 4000Mhz CL20...?

 

[JimmiG]

Yep, but keep in mind that actual bandwidth does increase with clock frequency, even if latency is the same. So 3200 C16 is still "better" than 3000 C15, and requires a better bin.

Ryzen also has this "peculiarity" where the inter-CCX communication latency and bandwidth are linked to memory clock (not memory latency), which is why higher frequencies and higher CAS latency perform better than lower frequencies and tighter timings, even if theoretical DRAM latency would be the same or worse.

 

[Elixer]

For what it is worth, friend is trying the new AGESA code, and he is at 3466MHz now on a ASUS board. Before, he couldn't get higher than 2933MHz.

I do agree though that you should buy the best DIMMs you can afford.

 

[JimmiG]

For what it is worth, friend is trying the new AGESA code, and he is at 3466MHz now on a ASUS board. Before, he couldn't get higher than 2933MHz.

I do agree though that you should buy the best DIMMs you can afford.

Looking forward to trying the new AGESA on my X370 Prime and seeing if the improvements apply to the cheaper board, too.
Definitively going to try running my LPX 3600's at their rated speed, or if that doesn't work, at least aim for 3466 MHz, and I'll also give my old LPX 3000's (3000 @ C16 Hynix) another try (previously had cold boot problems with them).

Still, it always pays off to go with quality RAM, because it plays such an important role in the stability of your system (bad RAM can cause anything from boot problems to data corruption to crashes and freezes).

 

[VirtualLarry]

Still, it always pays off to go with quality RAM, because it plays such an important role in the stability of your system (bad RAM can cause anything from boot problems to data corruption to crashes and freezes).

I suppose that's true at the higher-end. When sticking with pedestrian DDR4 speeds like 2133 or 2400, I've generally bought the cheapest branded DIMMs available, and had no issues. (Avexir, Geil, Team, and last but not least, GSkill)

 

[wchang99]

Ryzen only supports even CAS latencies, so only C14, C16, C18 etc. You would have to run the 3000 C15 memory at 2933 C16 (or C14, if you manage to get it stable). 3200 C16 is probably a slightly better bin than 3000 C15.

New BIOS/AGESA updates are improving compatibility, but it still pays off to get the best memory you can afford. It's more likely to actually work at the advertised speed, and if it doesn't, it will still probably work at a higher speed than a lower bin would have.

Higher frequencies and loose timings usually perform better with Ryzen than low frequencies and tight timings. E.g. 3200 C18 would perform better than 2933 C14 in most cases. The advertised spec is just what the RAM was tested at, so there's nothing stopping you from trying to run 3000 C15 RAM at 3200 C16 or C18, for example (as long as it's stable).

Thanks, JimmiG. I researched Ryzen only supporting even CAS latencies on your pointing me in that direction (and it was slightly distressing, I thought by now I knew all of Ryzen's quirks that impact building a system, but apparently I don't). Thanks for pointing that out.

I read of the new BIOS/AGESA updates (that came out after you posted) -- 4000MHz max is such a difference. Hopefully all compatibility issues will get easier accordingly.

I understand (and am on board with) your advice of getting the best memory you can afford (I will get 3200 CL16 over 3000 CL15).

Thanks for your detailed response -- really appreciated.

TLDR: think in nanoseconds for real-world latency comparation.

3000Mhz = 3000*1000² clock cicles per second at double data rate (DDR4), so clock cicle=1/(1500*1000²) seconds= 0.6666 nanoseconds. Then simply multiply by CAS latency to get the actual "total completion time", which is the total time it takes to perform a memory access and get ready for the next one. Your examples:

3000Mhz CL15: 1/(1500*1000²)*15 = 10 nanoseconds total completion time
3200Mhz CL16: 1/(1600*1000²)*16 = 10 nanoseconds total completion time

so those settings are effectively equal in terms of latency. Guess what would happen with 3600Mhz CL18 or 4000Mhz CL20...?

Thanks elpokor. I only recently came across this formula for real-world latency (nanoseconds)(I'm still low on the learning curve. )

Thanks for your responses, Elixer and VirtualLarry -- definitely helpful. I can't get enough good information on this.

Ryzen also has this "peculiarity" where the inter-CCX communication latency and bandwidth are linked to memory clock (not memory latency), which is why higher frequencies and higher CAS latency perform better than lower frequencies and tighter timings, even if theoretical DRAM latency would be the same or worse.

I came across a few sources on this -- you guys might've seen it, but I'll post this here:
http://www.legitreviews.com/ddr4-memory-scaling-amd-am4-platform-best-memory-kit-amd-ryzen-cpus_192259