3900x vs 9900k Base Supported RAM Speeds?

[Nozyspy]

Hi everyone,

 

I am going to be upgrading my computer quite soon but there are quite a lot of new developments since the last build i did (almost exactly 5 years ago), not least of which is AMD's resurgence.

 

One thing i am curious about is the base memory speeds supported for the 3900x vs the 9900k. The 3900x has 3200mhz listed whilst the 9900k has 2666mhz listed as the base memory speed supported. Now of course you can use XMP to boost up your memory, but as i understand it it can be a bit jenky and cause system instability (i would prefer the best performance reasonably possible but with trouble free motoring). Am i right in assuming no profiles are required to use the 3200mhz memory on the Ryzen 3900x, since it is supported by the chip natively? 

 

What performance difference is there between the 3900x using 3200mhz RAM and the i9 9900k using 2666mhz RAM, given that the AMD has a slightly slower base clock but the Intel has slower base memory support?

 

Thanks for any info you can offer or point me to to clarify this!

 

Nozy

[Ravendarat]

I was thinking about this the other day actually. 3600 mhz seems to be the agreed upon sweet spot for 3rd gen ryzen but what speed is actually RECOMMENDED for intel chips like the 9900? Im sure someone will chime in here shortly

[Mateyyy]

Those are kind of the "guaranteed" speeds that you can achieve with the respective CPUs, but I haven't heard of anyone with Zen 2 that has had problems with 3600MHz memory, which is indeed the sweet spot for these chips. For Intel, generally after 3000-3200MHz you don't really notice much of an improvement anymore in performance.

Also, anything over 2133MHz is technically an overclock, as that's the base speed for DDR4 (at 1.2V).

I would go for the 3900X with a dual-channel 3600MHz CL16 memory kit personally.

[Sakkura]

You will need to use memory profiles in either case.

 

As for performance, the lower official memory speed supported by Intel is not a factor at all. In reality the Intel memory supporter is still better than AMD's, though the difference is becoming fairly negligible in practice. At the same time, quirks of the infinity fabric design of Ryzen means you need fast memory to get the full performance. With Intel, memory speed is less important (though there's still no reason to get slow 2666 memory, 3000 costs the same).

[Jedimaster99]

3 minutes ago, Mateyyy said:

I would go for the 3900X with a dual-channel 3600MHz CL16 memory kit personally.

This system makes a lot more sense than the 9900k. You have all around better performance for a better price with Ryzen, and (2x8GB/4x8GB) of 3600mhz CL16 (CL17/CL18 doesn't kill ya) is the best priced RAM for the performance. 

[Jurrunio]

If you ask here we would just recommend pushing it higher lol, if a higher speed kit doesnt work properly.at rated speed you could always manually lower the frequency a bit and be done with it

 

As for memory support, Ryzen's memory spec is pretty much nailing for the worst sample of CPUs so 3200MHz is the lowest guarantee from AMD. Meanwhile 9th gen's memory spec is just because Intel got lazy... Even on cheap.boards Intel CPUs.often crack 4000MHz memory without much complaints.

 

And yes, good boards are crucial

 

That said, memory support is far less important than say, performance.and platform upgrade options. 9900k's platform is dead and it's slower in multicore performance by around 30% if I recall correctly.

[porina]

10 minutes ago, Mateyyy said:

Also, anything over 2133MHz is technically an overclock, as that's the base speed for DDR4 (at 1.2V).

Not exactly... JEDEC defines the official DDR4 speeds and timings, and they do go up to 3200 speeds. However, the lowest CL defined at 3200 is 20! Most enthusiast kits run around CL16 at that speed, or even lower if you throw more money at it. These are not standard profiles and go via XMP, and thus are arguably overclocks. I've long wondered why enthusiast ram manufacturers don't program in a SPD profile at JEDEC 3200 timings (or whatever is fastest the ram supports). At least you get the speed, if not the timings if you forget to turn on XMP.

 

I don't know how AMD define memory overclocks, if 3200 lower latency ram would be counted as an OC, as opposed to JEDEC timings.

 

7 minutes ago, Sakkura said:

quirks of the infinity fabric design of Ryzen means you need fast memory to get the full performance.

This is the main reason people aim for faster ram, moreso than the memory bandwidth in itself. IMO with the ever increasing core counts, there isn't anywhere near enough for more demanding uses, but for casual or lightweight uses (gaming, Cinebench) it doesn't matter. 3600 is the fastest speed safely support without going into async mode and the potential performance penalty from that. I believe there has been tests that running async, people with slower ram can still run high IF speeds and gain performance that way, but didn't look in detail which scenarios that may apply to.

 

7 minutes ago, Sakkura said:

With Intel, memory speed is less important (though there's still no reason to get slow 2666 memory, 3000 costs the same).

Similar problems to AMD here, more cores, more ram bandwidth can be useful. For casual uses I'd aim at 3000 minimum since you don't really save going any slower. It isn't as useful to go faster, but it still helps. 

 

Oh, on both Intel and AMD, you can run into various ram compatibility problems, especially past 3200 or so. It varies with mobo, bios, and how the modules are programmed.

[Nozyspy]

26 minutes ago, Mateyyy said:

Those are kind of the "guaranteed" speeds that you can achieve with the respective CPUs, but I haven't heard of anyone with Zen 2 that has had problems with 3600MHz memory, which is indeed the sweet spot for these chips. For Intel, generally after 3000-3200MHz you don't really notice much of an improvement anymore in performance.

Also, anything over 2133MHz is technically an overclock, as that's the base speed for DDR4 (at 1.2V).

I would go for the 3900X with a dual-channel 3600MHz CL16 memory kit personally.

Hmmm, but presumably on a 3900x then, it is only overclocking the RAM to bring it up to the natively supported speed of 3200mhz. For the Intel 9900k to use the 3200mhz memory it would also have to apply a slight overclock to the CPU too right?

[Stockholmes]

2 minutes ago, Nozyspy said:

Hmmm, but presumably on a 3900x then, it is only overclocking the RAM to bring it up to the natively supported speed of 3200mhz. For the Intel 9900k to use the 3200mhz memory it would also have to apply a slight overclock to the CPU too right?

It makes your CPU faster by definition, so I guess you could call it 'overclocking'. But the actual clockspeeds on the CPU aren't affected.

[Nozyspy]

4 minutes ago, Stockholmes said:

It makes your CPU faster by definition, so I guess you could call it 'overclocking'. But the actual clockspeeds on the CPU aren't affected.

But to get the same memory speeds on the Intel it does adjust the clock speeds of the CPU right? My concern is whether this adds in another instability variable?

[Stockholmes]

Don't worry, there is no instability variable. Your CPU won't overclock unless you overclock it yourself.

 

edit: I think you're thinking about the last generation of Ryzen here, where RAM clockspeeds mattered more concerning the CPU. With the 3900x there should not be such an issue. Even then, the RAM doesn't clock the CPU any higher. It just works better with a certain clock (ie. 3600MhZ if I recall correctly).

[Nozyspy]

2 hours ago, Mateyyy said:

Those are kind of the "guaranteed" speeds that you can achieve with the respective CPUs, but I haven't heard of anyone with Zen 2 that has had problems with 3600MHz memory, which is indeed the sweet spot for these chips. For Intel, generally after 3000-3200MHz you don't really notice much of an improvement anymore in performance.

Also, anything over 2133MHz is technically an overclock, as that's the base speed for DDR4 (at 1.2V).

I would go for the 3900X with a dual-channel 3600MHz CL16 memory kit personally.

I have never bought AMD before, but i do know that XMP is an Intel thing. How exactly would you get the memory to run at its advertised speed on an AMD board? I have heard that the XMP profiles still work anyway, but does AMD have its own version?

[Mateyyy]

2 minutes ago, Nozyspy said:

I have never bought AMD before, but i do know that XMP is an Intel thing. How exactly would you get the memory to run at its advertised speed on an AMD board? I have heard that the XMP profiles still work anyway, but does AMD have its own version?

XMP profiles should work in the same way, sometimes it's just got a different name, like for example on Asus boards it's called DOCP.

[Nozyspy]

1 hour ago, Stockholmes said:

Don't worry, there is no instability variable. Your CPU won't overclock unless you overclock it yourself.

 

edit: I think you're thinking about the last generation of Ryzen here, where RAM clockspeeds mattered more concerning the CPU. With the 3900x there should not be such an issue. Even then, the RAM doesn't clock the CPU any higher. It just works better with a certain clock (ie. 3600MhZ if I recall correctly).

I'm getting a bit confused.  I had always thought that for RAM that was advertised to run above the standard DDR4 speed, it required an overclock to be applied to the CPU (either manually or via XMP) to hit that speed, UNLESS the CPU supported higher than standard speed natively. In which case an overclock would only be applied to the CPU if the RAM speed was above its native support.

You'll have to forgive my ignorance, i only used XMP profiles once years ago, but ended up turning them off because it seemed to cause the occasional BSOD, i am a bit out of touch still with what has changed in the last few years.

 

 

16 minutes ago, Mateyyy said:

XMP profiles should work in the same way, sometimes it's just got a different name, like for example on Asus boards it's called DOCP.

 

Is XMP something that the majority of people use? Linus seems to have a particular hatred of XMP. 

[xg32]

the minimum recommend specs for intel is 2666, and zen 2 is 3200, zen 2's sweet spot is 3600 while intel can go as fast as the motherboard tolerates.

 

if your mobo has 4 ram slots the optimal setup is 4x8gb with low CL, and make sure you optimize the timings.

 

For intel, the difference between 2600 and 3200 is about 5%, 3200 to 4000 is about 3%.

For zen 2 is probably a 10% diff from 2600 to 3200 and 3%~ from 3200 to 3600 under heavy loads.

 

For minmaxing it's probably 4 sticks of ddr 3600 with a x570 taichi, or ddr4000 with a z390 gene/dark

[Mateyyy]

13 hours ago, Nozyspy said:

Is XMP something that the majority of people use? Linus seems to have a particular hatred of XMP. 

Yeah I mean it's just that they're not always guaranteed to work, but I've never had any problems myself with XMP personally.

[Nozyspy]

Just to make sure i am understanding all this correctly:

 

1. The JEDEC standard for DDR4 is 2133mhz.

 

2. The supported operating speeds of 2666mhz and 3200mhz for the i9 9900k and Ryzen 9 3900x respectively is the minimum recommended speed that your RAM should have.

 

3. Getting RAM to that minimum recommended speed will require an XMP profile anyway (since the RAM by default will be at 2133mhz?)

 

4. The 9900k and 3900x will support those minimum recommended speeds without applying any overclock to the processor?

 

5. If you want to run at above the minimum recommended speeds (say 3200mhz for the 9900k or 3600mhz for the 3900x) the XMP profile that comes with that speed of RAM will also apply a small overclock to the processor to compensate for the increased RAM speed?

[PixelPol]

On 4/11/2020 at 11:16 PM, xg32 said:

if your mobo has 4 ram slots the optimal setup is 4x8gb with low CL, and make sure you optimize the timings.

Wait a minute... Excuse my supreme ignorance ...artist, not so versed in hardware, only in what I see my tasks do better (ie, lots of RAM for heavy load in photoshop, fast single core in that one, many cores in Blender, good 8GB VRAM card in Davinci, etc). But is 4x8 still dual channel ? I have read everywhere that for Ryzen is gotta be dual channel. I am going (planning a rig) for a 3900X, 3200mhz CL16 memory, 2060 S 8Gb or 1660 S (the latter for replacing and selling in October... The former to keep it. Depending on final budget and/or getting or not certain set of gigs for Davinci). And dunno why I thought it HAD to be 2x16 , not 4x8, to be dual channel. That quad channel would make it perform worse or not supported well.

 

Another idea to stick to budget is 3700X, 2060 super (ideally Radeon VII - a beast in benchmarks in my stuff- if getting some crazy ebay offer..700+ eu is too much for me), but still 32 GB, as my workflows do see the best bang for the buck in more memory. A pity that 64 GB (ZB, AE and PS heavy print stuff eat lots of ram) would make me lower too much other components to stay on budget.  Also, I 'd love to leave room for another 2x16, for a total 64gb (but that globally wouldn't be dual channel?). And there's there the thing of Autumn with a ton of new stuff released... I know am gonna get mad then, but current machine is not cutting it, and gotta work faster...

[Sakkura]

On 4/15/2020 at 4:42 AM, Nozyspy said:

Just to make sure i am understanding all this correctly:

 

1. The JEDEC standard for DDR4 is 2133mhz.

 

2. The supported operating speeds of 2666mhz and 3200mhz for the i9 9900k and Ryzen 9 3900x respectively is the minimum recommended speed that your RAM should have.

 

3. Getting RAM to that minimum recommended speed will require an XMP profile anyway (since the RAM by default will be at 2133mhz?)

 

4. The 9900k and 3900x will support those minimum recommended speeds without applying any overclock to the processor?

 

5. If you want to run at above the minimum recommended speeds (say 3200mhz for the 9900k or 3600mhz for the 3900x) the XMP profile that comes with that speed of RAM will also apply a small overclock to the processor to compensate for the increased RAM speed?

1. There are several different JEDEC standard speeds for DDR4 memory. Most DDR4 memory defaults to the JEDEC 2133 standard, but some defaults to JEDEC 2400.

 

2. Those are not minimums, they are the highest memory speed the CPU manufacturer promises will definitely work. In practice, Intel is under-promising a lot.

 

3. Going above the speed the memory defaults to (2133 or 2400) will require changing memory settings in one way or another. XMP (or equivalent DOCP) is an easy way to do that, but you can also manually input clocks, timings and voltages.

 

4. Yes. If you go above those maximums, you're technically overclocking the memory controller in the CPU. But the CPU cores are not overclocked.

 

5. If you go above the maximum, the memory controller is technically overclocked, but not the CPU cores.

 

8 hours ago, PixelPol said:

Wait a minute... Excuse my supreme ignorance ...artist, not so versed in hardware, only in what I see my tasks do better (ie, lots of RAM for heavy load in photoshop, fast single core in that one, many cores in Blender, good 8GB VRAM card in Davinci, etc). But is 4x8 still dual channel ? I have read everywhere that for Ryzen is gotta be dual channel. I am going (planning a rig) for a 3900X, 3200mhz CL16 memory, 2060 S 8Gb or 1660 S (the latter for replacing and selling in October... The former to keep it. Depending on final budget and/or getting or not certain set of gigs for Davinci). And dunno why I thought it HAD to be 2x16 , not 4x8, to be dual channel. That quad channel would make it perform worse or not supported well.

 

Another idea to stick to budget is 3700X, 2060 super (ideally Radeon VII - a beast in benchmarks in my stuff- if getting some crazy ebay offer..700+ eu is too much for me), but still 32 GB, as my workflows do see the best bang for the buck in more memory. A pity that 64 GB (ZB, AE and PS heavy print stuff eat lots of ram) would make me lower too much other components to stay on budget.  Also, I 'd love to leave room for another 2x16, for a total 64gb (but that globally wouldn't be dual channel?). And there's there the thing of Autumn with a ton of new stuff released... I know am gonna get mad then, but current machine is not cutting it, and gotta work faster...

4 sticks of memory can enable quad channel... but only if the CPU supports that. Otherwise it will just run dual channel, with 2 sticks in each channel.

 

For stability reasons, 2 sticks rather than 4 may be slightly preferable with Ryzen. Not as big of a deal with the latest generation though, memory issues have gotten less common with each generation.