Ryzen's Infinity Fabric Clock Speed is Linked to Memory Clock Speed... Might Explain Why Memory OCs Make a Noticeable Impact in Performance on Ryzen?

[Drak3]

4 hours ago, MageTank said:

That is actually a very smart question. It would confirm whether or not the IF is tied to raw bandwidth, or just the clock speed itself. I'll pass this question along to some friends with Ryzen CPU's. Thanks for the suggestion.

Keep me in that loop too

[DocSwag]

5 hours ago, MageTank said:

That is actually a very smart question. It would confirm whether or not the IF is tied to raw bandwidth, or just the clock speed itself. I'll pass this question along to some friends with Ryzen CPU's. Thanks for the suggestion.

And me!

[Energycore]

On 3/16/2017 at 8:52 PM, themctipers said:

*sighs*

You sigh, but I tell you this: if this process allows for $300 9-cores and $220 6-cores to be marketable, then the industry can just adapt to them. Start pushing for higher RAM clockspeeds.

 

Now on this topic. I would really love to see whether you can get better performance if you try to run your memory at some 4000MHz (with a 125Mhz BCLK and the 32x multiplier maybe), and just looser timings and maybe 1.5V (good news, Ryzen isn't a fucking diva about RAM voltage like Skylake was).

[MageTank]

7 hours ago, Energycore said:

You sigh, but I tell you this: if this process allows for $300 9-cores and $220 6-cores to be marketable, then the industry can just adapt to them. Start pushing for higher RAM clockspeeds.

 

Now on this topic. I would really love to see whether you can get better performance if you try to run your memory at some 4000MHz (with a 125Mhz BCLK and the 32x multiplier maybe), and just looser timings and maybe 1.5V (good news, Ryzen isn't a fucking diva about RAM voltage like Skylake was).

Skylake isn't a diva with ram voltage either. People just completely threw a hissy fit over things they don't understand. Intel has been saying VDIMM kills CPU's since forever. Haswell's "Max VDIMM" was 1.575v according to their whitesheets. Skylake was 1.4175v for DDR3 and 1.26 for DDR4 according to their whitesheets. Yet people ran 1.65v DDR3 on Haswell for years, and we are currently running 1.45v DDR4 on Skylake just fine. I see people benching at 2.1v and no IMC has died from this. Why, you might ask? As I've repeatedly stated on this forum: Vdimm does not kill IMC's. It doesn't even come into contact with the CPU itself, as it's supplied from the board, to the memory. The voltages you must watch out for in regards to IMC, is VCCIO/VCCSA. These voltages tend to get higher as memory frequency gets higher, to help keep things stable. Intel simply used a blanket excuse that just so happened to be wrong, and people ran with it. I just hate how they conveniently ignored the exact same statements for Haswell, and donned their tin-foil hats for Skylake.

 

If anyone would like to read the Intel whitesheets for Haswell and Skylake, let me know, i'll link them here.

[leadeater]

4 minutes ago, MageTank said:

If anyone would like to read the Intel whitesheets for Haswell and Skylake, let me know, i'll link them here.

Nah just qoute the entire white papers without using spoiler tags so no one can miss the information or can say it was not provided to them 

[MageTank]

1 minute ago, leadeater said:

Nah just qoute the entire white papers without using spoiler tags so no one can miss the information or can say it was not provided to them 

I actually have screenshots of the relative information, but I've been accused of "cherry picking" when providing just those parts. Sad part is, I have an ASUS Z170 board with 1.65v DDR3 (2133 C9, super tight) sitting right behind me that has been running non-stop as a NAS/backup PC for the niece since June of last year, and the 6600T is still chugging along. No memory degradation either, lol. If VDIMM killed Skylake, I would think non-stop 1.65v would have done something by now in the almost 1 year of it running. 

[Nup]

Digital foundry seemed to find that there wasn't huge performance improvement with faster ram. 2166-3200, if I remember. 

[DocSwag]

1 hour ago, Nup said:

Digital foundry seemed to find that there wasn't huge performance improvement with faster ram. 2166-3200, if I remember. 

I've seen a bunch across the web and they tend to give around 9 or 10% performance jumps if I remember correctly  

[Misanthrope]

10 hours ago, Energycore said:

You sigh, but I tell you this: if this process allows for $300 9-cores and $220 6-cores to be marketable, then the industry can just adapt to them. Start pushing for higher RAM clockspeeds.

 

Now on this topic. I would really love to see whether you can get better performance if you try to run your memory at some 4000MHz (with a 125Mhz BCLK and the 32x multiplier maybe), and just looser timings and maybe 1.5V (good news, Ryzen isn't a fucking diva about RAM voltage like Skylake was).

Except it hasn't: Excavator had way more cores even if it lagged behind in IPC back then and the "industry" never pushed forward into more multithreaded apps outside of the workstation specialty ones which is a niche market vs consumer and gamer markets.

 

Could AMD survive as the cheap workstation manufacturer? Possibly but not with their current financials and spending on consumer segment advertising. We come back to the same points: it's been years since AMD enthusiasts have been telling us to just wait a bit more for software to catch up to this amazing hardware and it just never does. There's always "Oh but that DX12 game doesn't counts for x,y and z reasons, wait more" and we keep waiting and waiting and waiting and nothing gets done.

 

Sorry but the technology is obviously not enough if AMD doesn't has the business acumen to push it forward into mass adoption and a full paradigm shift from developers.

[Humbug]

4 hours ago, Misanthrope said:

Except it hasn't: Excavator had way more cores even if it lagged behind in IPC back then and the "industry" never pushed forward into more multithreaded apps outside of the workstation specialty ones which is a niche market vs consumer and gamer markets.

 

Could AMD survive as the cheap workstation manufacturer? Possibly but not with their current financials and spending on consumer segment advertising. We come back to the same points: it's been years since AMD enthusiasts have been telling us to just wait a bit more for software to catch up to this amazing hardware and it just never does. There's always "Oh but that DX12 game doesn't counts for x,y and z reasons, wait more" and we keep waiting and waiting and waiting and nothing gets done.

 

Sorry but the technology is obviously not enough if AMD doesn't has the business acumen to push it forward into mass adoption and a full paradigm shift from developers.

You make very good points.

 

I just want to say though that games today are a lot more multithreaded than they were 3-4 years ago. So things are moving although there is still more to do..

 

In the past maintstream i5s and i7s used to perform the same in games. It was conventional wisdom that nobody should buy more than an i5 for games. But now in modern games the i7s are clearly faster thanks to hyperthreading allowing 8 threads. When I bought my ivy bridge i7 3770k it was considered not worth the extra money vs the i5 3570k for gaming. But today it will not only beat an ivy bridge i5 it also destroys a haswell i5 and in a number of games even beats a kaby lake i5. This despite the lower ipc but just because of the 8 threads...

[Misanthrope]

11 minutes ago, Humbug said:

You make very good points.

 

I just want to say though that games today are a lot more multithreaded than they were 3-4 years ago. So things are moving although there is still more to do..

 

In the past maintstream i5s and i7s used to perform the same in games. It was conventional wisdom that nobody should buy more than an i5 for games. But now in modern games the i7s are clearly faster thanks to hyperthreading allowing 8 threads. When I bought my ivy bridge i7 3770k it was considered not worth the extra money vs the i5 3570k for gaming. But today it will not only beat an ivy bridge i5 it also destroys a haswell i5 and in a number of games even beats a kaby lake i5. This despite the lower ipc but just because of the 8 threads...

Not really a counter argument since I agree but just an extra thought: People always say AMD is future proofing and such but what this demonstrates is that it's intel that's moving the trend at a much more slower pace than AMD would when it comes to core and thread count. I leave it up to you if this is a reaction to intel dominating the market share or if on the other hand, intel it's just not spending anymore than they have to on multithreaded chips because they know the software development is not moving fast enough to really take advantage of much more than 4/8.

[Nup]

8 hours ago, DocSwag said:

I've seen a bunch across the web and they tend to give around 9 or 10% performance jumps if I remember correctly  

Not bad, but from a ~1ghz jump you'd think it'd show a good bit more... alas we'll see how it develops. 

[rjfaber91]

9 hours ago, DocSwag said:

I've seen a bunch across the web and they tend to give around 9 or 10% performance jumps if I remember correctly  

I'd love to be able to give you some data on that, since I do own a pair of 3200MHz sticks. Sadly, they refuse to run at anything higher than 2133MHz on my Aorus Gaming 5. Even 2400MHz, which is still well within AMD's specification, just results in a boot loop.

[MageTank]

18 minutes ago, Nup said:

Not bad, but from a ~1ghz jump you'd think it'd show a good bit more... alas we'll see how it develops. 

Sadly, we don't exactly know what is happening with the IMC in order to obtain that higher frequency. If tertiary timings (especially RTL/IO-L) is suffering to obtain higher speeds, then a 1000mhz increase in frequency won't bring much performance at all. It would be like going from 3200 to 3400. In fact, I can demonstrate this with my memory kit, showing lower frequency kits with tighter tertiary timings absolutely dominating higher frequency kits in both raw bandwidth and latency if you would be interested in seeing such information.

 

Good news is, I have been tasked to build a video editing rig with a Ryzen 1700. This means I will personally have my hands on the CPU to test it's IMC. Whether or not anyone has any faith in my ability to overclock memory, at the very least, I'll be able to point out what I see and answer any questions/test specific pieces that other reviewers might be overlooking. 

[Nup]

On 3/19/2017 at 10:49 PM, MageTank said:

Sadly, we don't exactly know what is happening with the IMC in order to obtain that higher frequency. If tertiary timings (especially RTL/IO-L) is suffering to obtain higher speeds, then a 1000mhz increase in frequency won't bring much performance at all. It would be like going from 3200 to 3400. In fact, I can demonstrate this with my memory kit, showing lower frequency kits with tighter tertiary timings absolutely dominating higher frequency kits in both raw bandwidth and latency if you would be interested in seeing such information.

 

Good news is, I have been tasked to build a video editing rig with a Ryzen 1700. This means I will personally have my hands on the CPU to test it's IMC. Whether or not anyone has any faith in my ability to overclock memory, at the very least, I'll be able to point out what I see and answer any questions/test specific pieces that other reviewers might be overlooking. 

It would be good to see!

Goes a little over the level i understand at the moment, but gotta start somewhere (makes me dream of what techquickie could have achieved ). 

Timings can't be improved by the user, right? 

 

keep us posted with what you find, hit me with a tag. 

[MageTank]

13 minutes ago, Nup said:

It would be good to see!

Goes a little over the level i understand at the moment, but gotta start somewhere (makes me dream of what techquickie could have achieved ). 

Timings can't be improved by the user, right? 

 

keep us posted with what you find, hit me with a tag. 

All timings can be improved by the user on Intel platforms, and previous AMD platforms, but on Ryzen, only the Primary timings (TCL, TRCRDD, TRCWRD, TRPT and TRAS). In order to adjust secondary/tertiary timings, you must lower the DRAM ratio (memory strap) and increase reference clock. An example of this being tCWL. At 2133, it's at a value of 10, and at 2400, it's at a value of 11. 

 

What we do know when it comes to Samsung B-Die, is that ram performs at it's best when reference clock is 145mhz, and DRAM ratio is 2400 (or 135mhz and 2666 DRAM ratio for 3600mhz). This equates to a 3480/3600mhz memory clock speed, and you can tighten the primary timings as much as you can. So far, the best I have seen is 3480mhz C11-10-10-10-22 and 3600 C11-11-11-11-22. Obviously, this is not a 24/7 stable profile, but it's the best I've seen for Ryzen. 

[DocSwag]

58 minutes ago, ONOTech said:

 

I got all R7 CPUs so I can help out. Just let me know how/what you want me to test. Granted, I won't be back at my dorm until Thursday. Currently in Florida enjoying that 80 F weather and the cool ocean breeze 

Considering I've found someone with a ryzen CPU, have you checked if you can disable your CPU to 4+0 cores (on CCXs)? I know that you can change it to 2+0 or 1+1 (thanks to AdoredTV), so I'm wondering if you can also change it to 4+0 or 2+2, as that would be quite an interesting test IMO. His video showed there is some difference (around 5-10%) between 1+1 and 2+0 so I'm wondering if the difference is larger with 4 cores.

[79wjd]

4 minutes ago, DocSwag said:

Considering I've found someone with a ryzen CPU, have you checked if you can disable your CPU to 4+0 cores (on CCXs)? I know that you can change it to 2+0 or 1+1 (thanks to AdoredTV), so I'm wondering if you can also change it to 4+0 or 2+2, as that would be quite an interesting test IMO. His video showed there is some difference (around 5-10%) between 1+1 and 2+0 so I'm wondering if the difference is larger with 4 cores.

http://www.hardware.fr/articles/956-24/retour-sous-systeme-memoire-suite.html

[DocSwag]

1 minute ago, djdwosk97 said:

http://www.hardware.fr/articles/956-24/retour-sous-systeme-memoire-suite.html

Hmmm so it does, but it seems like in the games Ryzen isn't doing as well in in makes a bigger difference...

 

Could explain some of the reason why Ryzen falls a little behind in gaming performance.

[MageTank]

21 minutes ago, DocSwag said:

Hmmm so it does, but it seems like in the games Ryzen isn't doing as well in in makes a bigger difference...

 

Could explain some of the reason why Ryzen falls a little behind in gaming performance.

I am still going to blame the IMC, simply because I know the impact that tight ram has on frametime variance. Going from 3200 XMP to 3600 C14-14-14-28-2 (while my framerate itself didn't change) yielded a reduction of about 2ms frametime on average at 120fps in overwatch (7ms before, 5ms now).

 

For me, minimum framerates and frametime mean the most. I could care less about how many frames Ryzen can actually drive, I care more about whether or not it's a stuttering mess when attempting to do so. A bad IMC can make for a bad gaming experience in modern titles. That being said, I don't know enough about these CCX's to simply write them off, so I won't pretend that they are or are not the issue. I am only going to go with what I know, lol. 

[Ezio Auditore]

On 3/19/2017 at 4:22 PM, MageTank said:

If anyone would like to read the Intel whitesheets for Haswell and Skylake, let me know, i'll link them here.

Yes

[MageTank]

Just now, Ezio Auditore said:

Yes

http://www.intel.com/content/dam/www/public/us/en/documents/datasheets/4th-gen-core-family-desktop-vol-1-datasheet.pdf (Section 7.8, table 49 for IMC voltages)

http://www.intel.com/content/dam/www/public/us/en/documents/datasheets/desktop-6th-gen-core-family-datasheet-vol-1.pdf (Section 7.2.1.3, table 7-4 for IMC voltages)

 

"Max Typ +5%" = max typical voltage (for DDR3L, this is 1.35v) + 5% (1.35 + 5% =  1.4175v). For DDR4, it is 1.2 + 5% = 1.26v.

[DocSwag]

19 hours ago, ONOTech said:

I've only tested 2 or 3 games on my 1700 using 4+0 and 2+2. So far, only BF1 shows a noticeable difference (about 15 ish percent). The rest, GTA V & FC Primal, aren't noticeable.

Just had a thought, was the testing done on DX12 or DX11? Could you retest with the other API? Sorry for bothering you again but I just have this little hunch that DX12 might not be doing too well on Ryzen, just a small nagging feeling though  

[Nup]

On 3/20/2017 at 11:38 PM, MageTank said:

All timings can be improved by the user on Intel platforms, and previous AMD platforms, but on Ryzen, only the Primary timings (TCL, TRCRDD, TRCWRD, TRPT and TRAS). In order to adjust secondary/tertiary timings, you must lower the DRAM ratio (memory strap) and increase reference clock. An example of this being tCWL. At 2133, it's at a value of 10, and at 2400, it's at a value of 11. 

 

What we do know when it comes to Samsung B-Die, is that ram performs at it's best when reference clock is 145mhz, and DRAM ratio is 2400 (or 135mhz and 2666 DRAM ratio for 3600mhz). This equates to a 3480/3600mhz memory clock speed, and you can tighten the primary timings as much as you can. So far, the best I have seen is 3480mhz C11-10-10-10-22 and 3600 C11-11-11-11-22. Obviously, this is not a 24/7 stable profile, but it's the best I've seen for Ryzen. 

Super info, appreciate the time taken. I think this summer i will spend some time tinkering with my ram. 

 

Keep us posted on that 1700x build you have coming. 

[PocketNerd]

So, long story short, pair Ryzen with the fastest RAM you can get. Gotcha.