Ryzen 5000 Series Reviews Discrepancy

[RacA]

As per GN vid posted earlier in this thread, seems there is even more mysteries about how Zen 3 scales with various ram configurations.
And it's the number of modules and/or ram chips on a single rank memory. Never seen that coming.
Steve suggested they might do a follow up investigation with Wendell at Level1Techs, I'm really looking towards this.

[RejZoR]

Just watched OC3D video where he overclocked Ryzen 5600X to 4.8GHz all core. Now that's stupid impressive for a Ryzen CPU. Add insane IPC on top of that and this generation is looking amazing. Now, sure he won the chip lottery a bit which can happen so you can't rely on it as a standard level, but it can happen apparently. I was equally lucky with my ancient 5820K that clocks like crazy at stupid low voltages. Hoping I'll be this lucky for Zen 3 when I go out and buy one.

[MageTank]

1 hour ago, RacA said:

As per GN vid posted earlier in this thread, seems there is even more mysteries about how Zen 3 scales with various ram configurations.
And it's the number of modules and/or ram chips on a single rank memory. Never seen that coming.
Steve suggested they might do a follow up investigation with Wendell at Level1Techs, I'm really looking towards this.

I don't think it's much of a mystery at all. We've always seen these kinds of interactions on memory controllers, even spanning back on older Intel platforms. I also don't believe it has anything to do with single rank DIMM's in particular as Steve didn't test dual rank DIMM's in his reviews. 

 

There are some important things to note involving this subject.

 

#1. When you populate 4 DIMM's, you activate 2DPC (DIMMS Per Channel) timings that are otherwise completely disabled or not in use. These timings, while technically more stressful on the memory controller, do have an impact on performance. If we are talking strictly bandwidth through channel interleaving, the addition of extra ranks (even on single rank DIMM's) will net a bandwidth improvement and potential reduction in latency simply by having these ranks open on a given channel. This is why Steve noticed an increase in performance despite not getting anywhere near his actual 32GB capacity in testing.

 

#2. Steve would have observed the exact same performance boost had he simply tested dual rank DIMM's in a 2x16GB configuration, and likely would have seen an even higher boost in performance. Now I want to point out real quick that he quoted Wendell from L1Tech as saying "He observes the best performance when using single rank 16GB DIMM's" however I simply do not believe this as it has been contradictory to my testing of every Ryzen 5000 series CPU I have on hand. It also goes against every bit of conventional wisdom we understand about memory ranks and timings in general. The reason we recommend dual rank DIMM's in a 2x16GB configuration is because it gives you the benefits of rank interleaving (the ability to read from one rank while simultaneously writing to another) while also maintaining a 1DPC configuration, saving some stress on the IMC. The use of 2 DIMM's for a 1DPC configuration also nets the highest potential for higher clock speeds and lower latency while maintaining the benefits of dual-rank (4 ranks total, much like Steve's testing configuration) AND the aforementioned benefits of rank interleaving.

 

TL:DR? We've been preaching this for years on this forum, reviewers are just now finally catching on.

[SpaceGhostC2C]

On 11/8/2020 at 10:27 AM, Nathanpete said:

Did you not consider the fact that 3600 MHz give AMD an unfair advantage over Intel. I'm not team blue, I just want an even playing field. 3200 MHz is a more realistic option for gamers but I believe it would still give AMD a small unfair advantage in the infinity fabric. Hence why I also suggested 3000 MHz. And a small unfair advantage is significant when we make our buying decisions based on performance differences of like 2%. 

No, it doesn't. There are only two ways to make a useful comparison:

1. Use the same RAM on both systems
2. Use the officially supported RAM speed of each system, even if they differ, as that's the only warranty they'll honor (manufacturers consider everything else "OC")

I can see no valid reason to qualify (1) by adding "and it has to be low speed RAM" or "it has to be this particular speed I have arbitrarily decided on". Hence, I can't fault any reviewer for using a specific kit over another. The argument of "what people will likely buy" isn't valid either: people watching reviews are going to (hopefully) buy based on reviews, and therefore the information reviews provide comes before their purchase decisions. Otherwise, we can review just one system: the one we assumed they were going to buy... 

"Normal" RAM speeds tend to go higher as the particular DDR revision matures, and whether consumers care about higher speeds depends on what kind of impact they expect it to have, and therefore you can't take "what people buy" (or "what i guess people buy") as a constant or reference point. I agree that testing at 5000MHz wouldn't be very useful, but 3600 is run of the mill memory, and even higher speeds are becoming commonplace in laptops.

 

On the other hand, for mid- and low-end CPUs, option (2) is more compelling, especially on the Intel side where memory speeds are locked for several SKUs, and also locked for chipsets that are more likely to be paired with such CPUs. Still, it's not "unfair" to use this for higher end CPUs as well, but viewers need to consider the relevance of each test. Best case: test multiple configs, provide full information. (Intel tends to be overly conservative with their "supported" RAM speeds, which means it's also overly restrictive in its locked platforms. Hence, even though (2) it's "fair", there would be even more complaints about "favoring AMD"...)

[SpaceGhostC2C]

2 hours ago, MageTank said:

. Now I want to point out real quick that he quoted Wendell from L1Tech as saying "He observes the best performance when using single rank 16GB DIMM's" however I simply do not believe this as it has been contradictory to my testing of every Ryzen 5000 series CPU I have on hand. It also goes against every bit of conventional wisdom we understand about memory ranks and timings in general.

 

I may add: I've heard Wendell latest streaming with Ian Cutress and I'd swear one of their comments on Ryzen 5000 was something along the lines of "go dual rank with Ryzen 3". I may be remembering it incorrectly, but it is aligned with what you are saying, and would point to GN misunderstanding or mis-saying what was being said on the other channel. But I haven't watched Level1Techs' review, so maybe Wendell said it in backwards there? 

[terminalinfinity]

Looks like my suspicion has been confirmed.  Greg Salazar tested the AMD-provided BIOS + 2 public variants and found up to a 300 point swing in cinebench + his processor stopped boosting properly with the public BIOS.

 

Furthermore, I pretty much confirmed earlier today the issue at least extends to Asrock and possibly ASUS boards.  See the details in my thread about it.

So it maybe more how the newest BIOS interacts w/specific board designs w/ MSI having the most boards in whatever design the new BIOS doesnt play nice with

[porina]

5 hours ago, MageTank said:

Now I want to point out real quick that he quoted Wendell from L1Tech as saying "He observes the best performance when using single rank 16GB DIMM's" however I simply do not believe this as it has been contradictory to my testing of every Ryzen 5000 series CPU I have on hand. It also goes against every bit of conventional wisdom we understand about memory ranks and timings in general.

That was the biggest headscratcher of the video. Have to wonder if that was misquoted, or otherwise misunderstood somewhere along the line. If we assume it is right, then could there be some new mechanism to understand? It should be easy enough to test, if people get their hands on 16GB SR modules as well as a Zen 3 CPU. Maybe not the easiest right now.

[MageTank]

23 minutes ago, porina said:

That was the biggest headscratcher of the video. Have to wonder if that was misquoted, or otherwise misunderstood somewhere along the line. If we assume it is right, then could there be some new mechanism to understand? It should be easy enough to test, if people get their hands on 16GB SR modules as well as a Zen 3 CPU. Maybe not the easiest right now.

I have 2 of each Zen 5000 series CPU currently out right now in my test lab along with almost every possible memory kit you could imagine. I've driven myself a little crazy the past few weeks with testing and have not observed what Steve is claiming Wendell observed. What I observed is a direct contradiction.

 

2x16GB is consistently out-performing the other configurations, followed by 2x8GB single rank (clock speed advantage, rank disadvantage), followed by 4x8GB (rank advantage when interleaving channels, clock disadvantage due to 2DPC timing configurations, especially when using daisy chain board topology), with 4x16GB coming in dead last due to the massive clock speed disadvantage imposed by the IMC stress brought about by both dual rank timings and 2DPC timings being enabled.

 

It is important to remember that these reviewers do not manually overclock their DIMM's. They are far more likely to load XMP and get to testing, which will often allow for variances in their results. This is especially true when you factor in the many different types of XMP options available these days, as very few are considered equal. ASUS and ASRock both have XMP optimization modes that allows the board to load XMP data for clock speeds and primary timings, but allows the board to complete the training process for secondary and tertiary timings. This training is going to differ based on several different variables, and will dramatically impact both stability and overall performance depending on the memory/cpu/board combination. 

 

I am not faulting or mocking the reviewers for this practice, mostly because it's impossible to manually overclock your DIMM's for each and every possible CPU and still have it be a fair review due to the difference in IMC quality even on the same exact families/models of chips. It would also take far too much time and would not be consistent with how the average user would use their system. My point is that what is being observed isn't something unique to Zen 3 and it has been noted before in previous Zen reviews. AMD's Robert Hallock has done some extensive testing as well on this very subject and came to pretty similar conclusions: https://community.amd.com/community/gaming/blog/2017/07/14/memory-oc-showdown-frequency-vs-memory-timings

 

One thing to note from that review is that at the time, they believed tuned single rank DIMM's to be superior to tuned dual rank DIMM's (and I agreed back then in my personal testing, dual rank DIMM's were simply too hard to OC on that older AGESA firmware) however this has changed dramatically since then. We now have a better understanding of which memory to pair with Ryzen along with newer firmware and better boards with superior memory topology. Vendors are finally taking memory OCing seriously again, and it's allowing us to take advantage of overclocking these dual rank DIMM's despite the stress they bring about.

[GDRRiley]

18 minutes ago, MageTank said:

It is important to remember that these reviewers do not manually overclock their DIMM's. They are far more likely to load XMP and get to testing, which will often allow for variances in their results.

most do not but people like GN Steve do. He sets the sub timings on ram too

[porina]

14 minutes ago, MageTank said:

I have 2 of each Zen 5000 series CPU currently out right now in my test lab along with almost every possible memory kit you could imagine. I've driven myself a little crazy the past few weeks with testing and have not observed what Steve is claiming Wendell observed. What I observed is a direct contradiction.

 

2x16GB is consistently out-performing the other configurations, followed by 2x8GB single rank (clock speed advantage, rank disadvantage), followed by 4x8GB (rank advantage when interleaving channels, clock disadvantage due to 2DPC timing configurations, especially when using daisy chain board topology), with 4x16GB coming in dead last due to the massive clock speed disadvantage imposed by the IMC stress brought about by both dual rank timings and 2DPC timings being enabled.

Ok, I suppose there's two dimensions to be tested here. One is if we have the same speeds and timings (as far as is practical), does each ram configuration give different performance? This is the "dual rank" advantage. The other is, if you overclock the memory to its best, where each configuration may affect the speed timings on top of those factors, where you say 2x2R > 2xSR > 4xSR > 4x2R.

 

Out of interest, do you see much of a difference between 2x2R vs 4xSR at equal clocks/timings? They seem pretty much identical to me.

 

 

14 minutes ago, MageTank said:

I am not faulting or mocking the reviewers for this practice, mostly because it's impossible to manually overclock your DIMM's for each and every possible CPU and still have it be a fair review due to the difference in IMC quality even on the same exact families/models of chips. It would also take far too much time and would not be consistent with how the average user would use their system.

I think this is an argument point at the moment. Personally I'd like to see at least two configurations. 1, at the officially supported speed, 2, at a typical high end user speed, such as 3600. Enthusiasts are more likely to get higher speed kits, but I'd suspect there's still many systems out there at more moderate speeds. Similarly, most users will turn on XMP and hope it works, which isn't 100% of the time. The numbers of people manually tuning ram are pretty small in proportion. I've pretty much stopped doing it as I'm taking a break from competitive benchmarking. I just want things to work reliably, never mind that last % of performance.

 

14 minutes ago, MageTank said:

My point is that what is being observed isn't something unique to Zen 3 and it has been noted before in previous Zen reviews. AMD's Robert Hallock has done some extensive testing as well on this very subject and came to pretty similar conclusions:

I never did it as seriously, but on both Intel and up to Zen 2 systems, I did see quite a benefit to DR over SR at the same speed in compute uses, but not so much in gaming. I wonder if, with Zen 3s higher than ever potential, it is starting to show more sensitivity than in the past?

[MageTank]

19 minutes ago, GDRRiley said:

most do not but people like GN Steve do. He sets the sub timings on ram too

If he is going through and setting every single secondary and tertiary timing, I applaud his effort, but that is absolute madness. Unless he is doing so from a median perspective (an overclock configuration that all of his CPU's can achieve), this in and of itself would introduce more variables for his testing if memory controller quality differs per chip. This is extra troublesome on modern boards that let you POST with obvious unstable RAM overclocks but end up correcting the instability through additional training/correction procedures that negatively impact performance. If you have any information on his testing methodology, I'd love to give it a read.

 

14 minutes ago, porina said:

Ok, I suppose there's two dimensions to be tested here. One is if we have the same speeds and timings (as far as is practical), does each ram configuration give different performance? This is the "dual rank" advantage. The other is, if you overclock the memory to its best, where each configuration may affect the speed timings on top of those factors, where you say 2x2R > 2xSR > 4xSR > 4x2R.

If you test with all things equal (clock speeds, primary timings, etc) then dual rank will automatically win out due to a few factors. Rank interleaving is a pretty serious boon in and of itself, but simply being able to avoid fetching page data from a different DIMM entirely saves you a massive latency penalty which we know is important given Ryzen's latency issues in general. The real question is, at what point does the advantages of rank interleaving and 2DPC channel interleaving through the additional rank bandwidth give way to the advantages of being able to simply OC your ram higher due to the lack of stress brought about by those timings being enabled in the first place?

 

That's the question I've been wasting a lot of time on and thankfully with the design of Ryzen, it's fairly simple to answer. Since Ryzen doesn't scale well beyond 3800mhz (even on Ryzen 5000 series, even with the latest AGESA firmware version, 4000mhz is extremely difficult to achieve and nigh impossible on dual rank DIMM's at this time) due to the behavior of the infinity fabric when unstrapping the FCLK. So since we have that ceiling in-place regardless of the limitations imposed by DPC timings or multi-rank timings, we know that if our board and memory controller can push 3600-3800 with dual rank, it will most likely out-perform anything clocked higher than that simply by design of the IMC/infinity fabric.

 

28 minutes ago, porina said:

Out of interest, do you see much of a difference between 2x2R vs 4xSR at equal clocks/timings? They seem pretty much identical to me.

At equal clocks/timings? Yes, but only in synthetics/compression benchmarks. While the actual number of memory ranks (and their respective bandwidths) are identical, you cannot emulate the ability to read from one rank while writing to another from within the same DIMM. With all things completely equal (with the exception of the multi-rank timings, which I cannot keep equal as they simply do not work on 4xSR configuration), I observed a consistent difference of roughly 8-10% favoring 2x16GB in synthetic benchmarks (biggest being Linpack) and 7-zip compression/decompression workloads. I'll be completely honest, I don't waste a lot of time on gaming benchmarks mostly because my line of work doesn't require that I do so, but I am happy to extend my testing if any of you have some suggestions.

 

34 minutes ago, porina said:

I never did it as seriously, but on both Intel and up to Zen 2 systems, I did see quite a benefit to DR over SR at the same speed in compute uses, but not so much in gaming. I wonder if, with Zen 3s higher than ever potential, it is starting to show more sensitivity than in the past?

I had considered this as well, and it's still not entirely outside of the realm of possibilities, but I know for absolute certain that the IMC itself has not changed on Zen 3. It feels exactly the same as Zen 2 down to the most basic behaviors it has with various IC's in the DIMM's I use. The real question is, did AMD's change with the removal of CCX's and adjustment to their cache hierarchy change how their CPU's interact with memory? I personally don't know, and based on my testing, it doesn't really seem to be the case, but I also don't feel like we have the final AGESA firmware either. There are still a lot of quirky interactions that need worked out and I feel like there is still some performance left on the table that we can take advantage of.

 

I don't do a lot of XMP testing mostly because I don't believe in XMP, but I'll see if I can put something together to try to mimic these review systems and see what these XMP profiles are actually doing when loaded. How they are actually being trained, and whether they differ per board being used. For all we know, this could be a board vs board variance issue, or a bug in the AGESA firmware that isn't properly adjusting DPC timings when running in a 1DPC configuration (leaving tertiary timings too loose). If GamersNexus doesn't beat me to it, would be worth exploring this new firmware on Zen 2 to see if it changes the way memory training behaves on those older CPU's.

[GDRRiley]

4 minutes ago, MageTank said:

If he is going through and setting every single secondary and tertiary timing, I applaud his effort, but that is absolute madness. Unless he is doing so from a median perspective (an overclock configuration that all of his CPU's can achieve), this in and of itself would introduce more variables for his testing if memory controller quality differs per chip. This is extra troublesome on modern boards that let you POST with obvious unstable RAM overclocks but end up correcting the instability through additional training/correction procedures that negatively impact performance. If you have any information on his testing methodology, I'd love to give it a read.

he is 3200 C14 in a 4x8gb config

Spoiler

We use four 8GB sticks of G.Skill TridentZ RGB 3200MHz memory unless otherwise noted, with all primary and with select secondary and tertiary timings manually controlled for test consistency. We’ve found this to be very important for test-to-test consistency, as allowing the motherboard to auto-select timings (which it will do) could ruin the data.

 

[MageTank]

9 minutes ago, GDRRiley said:

he is 3200 C14 in a 4x8gb config

  Hide contents

We use four 8GB sticks of G.Skill TridentZ RGB 3200MHz memory unless otherwise noted, with all primary and with select secondary and tertiary timings manually controlled for test consistency. We’ve found this to be very important for test-to-test consistency, as allowing the motherboard to auto-select timings (which it will do) could ruin the data.

 

Sounds like some of the older standard B-Die, all single rank. I am still curious as to what he means by "select secondary and tertiary timings manually controlled" as depending on the selection, those are very important. For example: If he controlled SG/DG (Same Group / Different Group) timings but not DD (Different DIMM) timings (DR or Different Rank wouldn't matter for his configuration given the use of single rank DIMM's), he would be dealing with an issue where the board is training his DPC timings which would definitely skew the results of his testing. I'd imagine Steve is smart enough to know this already, but if he would be willing to dive a little deeper in his methodology, that would be appreciated. Might be worth one of those "Ask GN" segments and see if he responds.

 

Thanks for finding that info for me.

[leadeater]

5 hours ago, MageTank said:

It is important to remember that these reviewers do not manually overclock their DIMM's. They are far more likely to load XMP and get to testing, which will often allow for variances in their results.

That may be the difference with the GN results, he controls the primary and secondary timings (he notes a few of them in the video). What he may be setting might react differently etc.

 

4 hours ago, MageTank said:

If he is going through and setting every single secondary and tertiary timing, I applaud his effort, but that is absolute madness. Unless he is doing so from a median perspective (an overclock configuration that all of his CPU's can achieve), this in and of itself would introduce more variables for his testing if memory controller quality differs per chip. This is extra troublesome on modern boards that let you POST with obvious unstable RAM overclocks but end up correcting the instability through additional training/correction procedures that negatively impact performance. If you have any information on his testing methodology, I'd love to give it a read.

It's only a few that he has found matter the most or need to be controlled to a known setting that gives the best consistency. I don't think he's using any aggressive settings or ones that are at all likely to fail, just controlling the values of them so they are known across all tests and platforms. Which is actually the best thing to be doing but does require extra setup time, not a lot however.

[ion_mswd]

Anyone explain what is 19 c in amd ryzen 5000 series. The brand new AMD Ryzen 5000 collection processors are nonetheless utilizing AMD’s 7nm course of however supply a 19 p.c enhance in directions per cycle, together with a whole redesign of the chip structure and the next max increase velocity. AMD Ryzen 5000 series (The new chipsets are leaping straight to Ryzen 5000 collection branding to keep away from any confusion of the brand new Zen 3 chips with the Zen 2-based Ryzen 4000 desktop chips that AMD launched over the summertime. Check this link for post  .

[LAwLz]

3 minutes ago, ion_mswd said:

Anyone explain what is 19 c in amd ryzen 5000 series. The brand new AMD Ryzen 5000 collection processors are nonetheless utilizing AMD’s 7nm course of however supply a 19 p.c enhance in directions per cycle, together with a whole redesign of the chip structure and the next max increase velocity. AMD Ryzen 5000 series (The new chipsets are leaping straight to Ryzen 5000 collection branding to keep away from any confusion of the brand new Zen 3 chips with the Zen 2-based Ryzen 4000 desktop chips that AMD launched over the summertime. Check this link for post  .

That entire article is horribly written. It reads like someone ran it through Google translate.

What I think they are referring to is the 19% IPC uplift.

 

IPC stands for "instructions per clock". In everyday speak it refers to the amount of calculations a core can do in a given number of clock cycles.

Let's say we have a CPU called CPU-X. It runs at 1Ghz and gets a benchmark score of 100.

Then we have a CPU called CPU-Y. It also runs at 1GHz but it gets a benchmark score of 120.

 

Both processors run at 1GHz, but CPU-Y has a better architecture and it is able to do 20% more work at the same frequency as CPU-X. That means CPU-Y has 20% higher IPC than CPU-X.