We know Ryzen scales with speed and latency, speeding up the IF is crucial. But how much does Coffee Lake really benefit when you start going over ~3600 into the "holy crap that's expensive" RAM territory? I mean given the price premium it seems like it would only make sense after you've already bought the highest end CFL chip you can lay your hands on.
IMO is for the benchmark champions, not real world use, at least I cannot think of one, when the latency mucks up the bandwidth available it becomes not really worth it for much of anything else.
There might be a very small handful of games, pro apps, workstation stuff etc that really cares about raw bandwidth over snappy latency, but most times it is the other way around, which is why it seems to be DDR2 at 800, DDR3 at 1600-1866, DDR4 at 2600-3200 (with latencies to match the speed such as CL5 for the 800, CL 7-9 for the 1600 and CL 14-18..subtimings are important as well)
I've got G-Skill memories running at DDR4-3466 with G.Skill Trident Z 2x8GB CL16 and Asus Crosshair VI Hero and Ryzen 1700. It refused to work at that speeds, but after a BIOS update around the launch of R2 series it does.
Whether or not you'll see performance benefits past X frequency is very different from whether or not it continues to run faster. I've seen memory scaling tests that show Ryzen still scaling the same all the way up to 3466. Other configs may continue to demonstrate good scaling beyond that point... again depends on the software and where the bottleneck lies.
DDR4-5066 CL21 is 4.145 nanoseconds DDR4-4000 CL18 is 4.5 nanoseconds
My RAM is DDR4-3200 CL14 which is 4.375 ns. I always look for lowest latencies, higher speeds are usually over-hyped. It's all about how fast RAM can return data, but the speed in MHz is a measure over one second not per request and to get higher speeds they add wait cycles. 4.5 nanoseconds is how long that RAM takes to provide the requested data, the lower the better. Mine is actually faster per access even though it has a lower MHz speed.
Take the rated speed and divide it by the CL to get your effective speed, that's what you should be comparing. Mine is 228MHz. I think that is 228 MB per second? 1.8GBps after accounting for the 64bit data bus. More due to "double data rate"? QDR now? So 7.2GBps? Let's not forget dual channel memory that pushes it up to 14.4GBps. It's been too long, that's just guessing now.
When you talk about memory speeds, you also need to look at the actual impact on performance. When I see high latency memory, the question becomes what the real world performance impact will be. 3200CL14 vs 3466CL15 or 3466CL16, to 3600CL16 to 4000CL20. You would hope that 4000MHz memory would be fast enough to have better performance at the higher latency compared to 3466CL15, but seeing a true benchmark comparison, and on Ryzen where things are far from simple when it comes to high speed memory, it is a good question.
I just hope future Zen releases get much better memory controllers. I keep wondering if my Samsung B-die can run faster on a better controller, and if so that might save me some money when/if Zen 3 or 5 comes to AM4. Or what ever the last major Zen version update is for AM4. I'm assuming Zen 3 is the last AM4 chip, but you never truly know until the time comes. If Zen 3 can run 4000 or higher without issues maybe just maybe so can my 3200 b-die, if I'm lucky.
If you have memory rated for DDR4-3200, then running it at 3200 won't be a problem. If you buy DDR4-4000 rated memory and you can't run it at that speed due to the memory controller, THAT is a problem. Getting even Hynix M-die to run at 3200 isn't a big problem with relatively recent BIOS versions. 3466 and 3600 will be a bit tougher, but Samsung B-die shouldn't have too many problems with Ryzen.
Expecting memory rated for 3200 to run at 3466 if the platform gets a magic update is going to be more a function of the memory itself, rather than the memory controller. In general, you may need to use normal CPU overclocking techniques, such as using more voltage, or using looser timings(go to 16 or even 18 CAS latencies for example).
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
18 Comments
Back to Article
peevee - Monday, June 11, 2018 - link
"CL21 26-26-54"The game is old.
Von Neumann is dead.
rahvin - Monday, June 11, 2018 - link
With latencies that high (21-26-26-54) you have to wonder if it's even faster at all.edzieba - Monday, June 11, 2018 - link
Remember CAS latencies are not absolute, they're measured in clock cycles. e.g. CL21 at 5066MHz is equivalent to CL10 at 2400MHz.StevoLincolnite - Monday, June 11, 2018 - link
Common DDR5, where arth thou?Alexvrb - Monday, June 11, 2018 - link
We know Ryzen scales with speed and latency, speeding up the IF is crucial. But how much does Coffee Lake really benefit when you start going over ~3600 into the "holy crap that's expensive" RAM territory? I mean given the price premium it seems like it would only make sense after you've already bought the highest end CFL chip you can lay your hands on.Dragonstongue - Monday, June 11, 2018 - link
IMO is for the benchmark champions, not real world use, at least I cannot think of one, when the latency mucks up the bandwidth available it becomes not really worth it for much of anything else.There might be a very small handful of games, pro apps, workstation stuff etc that really cares about raw bandwidth over snappy latency, but most times it is the other way around, which is why it seems to be DDR2 at 800, DDR3 at 1600-1866, DDR4 at 2600-3200 (with latencies to match the speed such as CL5 for the 800, CL 7-9 for the 1600 and CL 14-18..subtimings are important as well)
Kalelovil - Monday, June 11, 2018 - link
IIRC Ryzen doesn't run IF any faster above DDR4-3200.alex-w - Tuesday, June 12, 2018 - link
I've got G-Skill memories running at DDR4-3466 with G.Skill Trident Z 2x8GB CL16 and Asus Crosshair VI Hero and Ryzen 1700. It refused to work at that speeds, but after a BIOS update around the launch of R2 series it does.Alexvrb - Tuesday, June 12, 2018 - link
Whether or not you'll see performance benefits past X frequency is very different from whether or not it continues to run faster. I've seen memory scaling tests that show Ryzen still scaling the same all the way up to 3466. Other configs may continue to demonstrate good scaling beyond that point... again depends on the software and where the bottleneck lies.PixyMisa - Tuesday, June 12, 2018 - link
Still two years away, most likely.zodiacfml - Monday, June 11, 2018 - link
We forgot to tell it is overclocked. Oh, the water chiller too.'nar - Tuesday, June 12, 2018 - link
Word games and marketing hype, mostly.DDR4-5066 CL21 is 4.145 nanoseconds
DDR4-4000 CL18 is 4.5 nanoseconds
My RAM is DDR4-3200 CL14 which is 4.375 ns. I always look for lowest latencies, higher speeds are usually over-hyped. It's all about how fast RAM can return data, but the speed in MHz is a measure over one second not per request and to get higher speeds they add wait cycles. 4.5 nanoseconds is how long that RAM takes to provide the requested data, the lower the better. Mine is actually faster per access even though it has a lower MHz speed.
'nar - Tuesday, June 12, 2018 - link
Take the rated speed and divide it by the CL to get your effective speed, that's what you should be comparing. Mine is 228MHz. I think that is 228 MB per second? 1.8GBps after accounting for the 64bit data bus. More due to "double data rate"? QDR now? So 7.2GBps? Let's not forget dual channel memory that pushes it up to 14.4GBps. It's been too long, that's just guessing now.Targon - Tuesday, June 12, 2018 - link
When you talk about memory speeds, you also need to look at the actual impact on performance. When I see high latency memory, the question becomes what the real world performance impact will be. 3200CL14 vs 3466CL15 or 3466CL16, to 3600CL16 to 4000CL20. You would hope that 4000MHz memory would be fast enough to have better performance at the higher latency compared to 3466CL15, but seeing a true benchmark comparison, and on Ryzen where things are far from simple when it comes to high speed memory, it is a good question.SkOrPn - Wednesday, June 13, 2018 - link
I just hope future Zen releases get much better memory controllers. I keep wondering if my Samsung B-die can run faster on a better controller, and if so that might save me some money when/if Zen 3 or 5 comes to AM4. Or what ever the last major Zen version update is for AM4. I'm assuming Zen 3 is the last AM4 chip, but you never truly know until the time comes. If Zen 3 can run 4000 or higher without issues maybe just maybe so can my 3200 b-die, if I'm lucky.Targon - Wednesday, June 13, 2018 - link
If you have memory rated for DDR4-3200, then running it at 3200 won't be a problem. If you buy DDR4-4000 rated memory and you can't run it at that speed due to the memory controller, THAT is a problem. Getting even Hynix M-die to run at 3200 isn't a big problem with relatively recent BIOS versions. 3466 and 3600 will be a bit tougher, but Samsung B-die shouldn't have too many problems with Ryzen.Expecting memory rated for 3200 to run at 3466 if the platform gets a magic update is going to be more a function of the memory itself, rather than the memory controller. In general, you may need to use normal CPU overclocking techniques, such as using more voltage, or using looser timings(go to 16 or even 18 CAS latencies for example).
Oxford Guy - Thursday, June 14, 2018 - link
"at a high voltage"Whatever that means.