AMD Ryzen 7 1800X CPU Review

[-Fran-]

Actually, I also agree that it is not such a good idea to benchmark mmorpgs. My reason for that is that when you REALLY are cpu bound is when there are lots of players fighting lots of things with effects, and that is the hardest thing to produce consistently. In order to produce it consistently, you would require huge amounts of time and effort, that could be better spent benchmarking several extra games instead.

BUT there are some ways of doing it repeatably (or semi-repeatably) that I can think of:

1- Empty place, no people, some or many mobs, only 1 person fighting: Very repeatable, but useless as a benchmark since there is no taxing workload. Useless

2- Main commercial city filled with people on a busy day at the highest time: Repeatable enough, as long as it's done in the same day at the same approximate time. Somewhat taxing, but no effects and no fighting. The main disadvantage is that a result from a certain date is useless in any future, as game population varies too much on the medium term, so no comparability with future hardware. The benchmark results are only good for one review. Basically useless

3- Set up your own private server and configure many bots to fight repeatedly against certain mobs: Perfect repeatability and realistic taxing, but it would cost too much (money and time) to benchmark only one game. Even worse if you want several mmorpgs. But an advantage: Setting up just one server is good enough for several years, since mmorpgs don't change much over time, and only a few are popular at any given period of 3-5 years.


Last option would be the one that is feasible (although quite costly for only 1 game).


I don't know almost anything about crysis, so I wont comment on that.

Let's drop Crysis for the time being, lol. I don't want to go back to 2007 discussions XD

So, about MMORPGs... Yes, they are hard to test and most probably a PITA as well since to reduce the error margins in your measurements, you'd have to repeat the tests at least a dozen times.

Like you say, there are ways. Toms is a review power house, they have done a lot to discover issues and in these almost 20 years that I've been reading the site (since 2001), I've found a lot of interesting articles that actually go against "common understanding". The bet example I have is the 2D rendering investigation they did "just because" and prompted ATI to get their act together. Also the Pentium 3 errata.

I mean, I am raising the bar here, because I know Toms has capable people that can do it. I mean, they were already doing it with WoW. I don't see why they can't do it again. They could even have a dedicated review series just dedicated to investigate how CPU scaling works with MMORPGs and MP games as a separate review/article. Options are out there for them to get the information to us.

But yeah, it is a lot to ask and I do get that. We can contribute if they ask (referencing your next post) or they could contact the Devs behind some of the MMORPGs and see if they could accomodate a testing server for them or something? We could even go there and participate or something? A logistic nightmare for sure, but I'm 1000% sure it would be fun as hell.

Cheers!

EDIT: Typos.

 

[InvalidError]

Last option would be the one that is feasible (although quite costly for only 1 game).

The last option would be against the terms of use of most popular MMOs, which makes it a no-go due to the risk of legal action.

 

[Onus]

While I'd like to see more MMORPGs tested, I think the potential issues of repeatability are valid, since the question this test might answer has already been answered.
For example, the GW2 tests I ran on my Athlon X4-5350 showed little more than that the game was playable, which in context could be a useful piece of information; I also showed that Diablo II (2, not 3) was NOT playable on that little system. That wasn't a comparison vs. anything else though, but only helped find a Yes/No answer to fitness for purpose. Strictly on-topic for Ryzen, I think it's clear that even where Ryzen is out-performed, there is no task for which it is categorically unfit for purpose, even if it isn't the best choice. That would sharply limit the usefulness of MMORPG tests, where the goal is comparative results rather than a Yes/No answer.

 

[anbello262]

Last option would be the one that is feasible (although quite costly for only 1 game).

The last option would be against the terms of use of most popular MMOs, which makes it a no-go due to the risk of legal action.

From my understanding, private servers are allowed in several popular mmorpgs. And if they aren't in a specific one, I'm quite sure that there can be a workaround, like negotiating with the game managers based on the fact that this isn't for playing, it's just for reviewing/benchmarking. I would guess that most would accept under those terms, at least for this site.

 

[-Fran-]

While I'd like to see more MMORPGs tested, I think the potential issues of repeatability are valid, since the question this test might answer has already been answered.
For example, the GW2 tests I ran on my Athlon X4-5350 showed little more than that the game was playable, which in context could be a useful piece of information; I also showed that Diablo II (2, not 3) was NOT playable on that little system. That wasn't a comparison vs. anything else though, but only helped find a Yes/No answer to fitness for purpose. Strictly on-topic for Ryzen, I think it's clear that even where Ryzen is out-performed, there is no task for which it is categorically unfit for purpose, even if it isn't the best choice. That would sharply limit the usefulness of MMORPG tests, where the goal is comparative results rather than a Yes/No answer.

It is a fair point, but as usual the devil is in the details. In every single review, there's no "yes/no" type of answer in my opinion, but a behavior to analyze. You have the hard numbers, but also the impressions from the reviewer which are as important (IMO) as the numbers presented.

Also, you can even see it in SP games tested and their benchmarks used: If you don't use the internal benchmark then you might be skipping specific parts of the game engine that could put more stress in different components. That goes for every single piece of software I would say. Drivers could even optimize for specific benchmarks and "cheat"? Well, tinfoil hat off, I do like when reviewers get clever and try to mix things a bit, just in case.

From my understanding, private servers are allowed in several popular mmorpgs. And if they aren't in a specific one, I'm quite sure that there can be a workaround, like negotiating with the game managers based on the fact that this isn't for playing, it's just for reviewing/benchmarking. I would guess that most would accept under those terms, at least for this site.

Plus, at a first glance, it sounds like a "win win" 😛

Cheers!

 

[InvalidError]

I would guess that most would accept under those terms, at least for this site.

Sending review samples costs hardware manufacturers only a few hundred dollars. Asking Blizzard and other companies to setup a server entirely for a single site's testing purposes would cost them thousands of dollars if at all possible and since most companies wouldn't hand over server code to third parties, such tests would have to be done on a hosted server, which means that test results may also be affected by network latency.

I would be surprised if MMO operators would be willing to go that far even for THG. Maybe the small obscure ones that almost nobody knows about who can host their whole game on a single server. In the case of WoW and other major MMOs, a server or realm is a few racks worth of equipment.

 

[anbello262]

Sending review samples costs hardware manufacturers only a few hundred dollars. Asking Blizzard and other companies to setup a server entirely for a single site's testing purposes would cost them thousands of dollars if at all possible and since most companies wouldn't hand over server code to third parties, such tests would have to be done on a hosted server, which means that test results may also be affected by network latency.

I would be surprised if MMO operators would be willing to go that far even for THG. Maybe the small obscure ones that almost nobody knows about who can host their whole game on a single server. In the case of WoW and other major MMOs, a server or realm is a few racks worth of equipment.

Actually, there are cracked servers of many popular mmorpgs, and those are usually run by people on their basements or by using some cheapish hosting service. In those cases the code is already "out there", so I doubt the companies would have any problems with giving it to a review site with a ND agreement.
But my main point here is the hardware where it is being run.
The huge arrays of servers are only necessary so that thousands of players can play at the same time with no lag, and to be able to update is quickly (and patch-fix big issues in no time).
A private server hosting only 50-100 players can be run on some workstation hardware on a gigabit lan. Even consumer i7s are powerful enough for most.

On the cases where the server files are still private and companies wish them to remain so, it wouldn't be hard to set up a special "instance" in-game, where you could only access with a code (given to the editors). Just like an instanced dungeon (so common in many games). A different instance for each registered reviewer site, even, with no added cost.
Just inform a game mod in advance that you wish to benchmark (maybe some low traffic weekdays/hours would be preferable), and they open the instance for you and your bots.
It wouldn't be hard to make sure that those (registered) bots only work in that specific 'dungeon' instance, and that they get no exp in there (to keep consistent levels), and that all loot disappears if they leave the place. Or they can even be NPCs mimicking player classes, behaviour and skills (just a server side bot).

It might cost a few thousand to set it up the first time, but the maintenance costs would probably be low, and the added publicity could be worth a lot more.


The more I think about it, the more feasible it seems, although the costs are still high, given the different options available. The highest costs would probably be the negotiation and agreement of terms previous to the actual implementation, and making sure it is done in a secure way.

 

[InvalidError]

Actually, there are cracked servers of many popular mmorpgs

And how many of those 'cracked servers' aren't considered a violation of the ToS by the original MMO's creators?
How many of those 'cracked servers' are an accurate duplication of the original servers' mob and boss mechanics?
How many of those 'cracked servers' accurately duplicate the game play experience on normal live servers?

I have tried playing WoW on many "private servers" and most mobs have little more than follow and auto-attack scripts attached to them. There isn't much point in playing such a massively neutered version of PvE content.

 

[-Fran-]

Actually, there are cracked servers of many popular mmorpgs

And how many of those 'cracked servers' aren't considered a violation of the ToS by the original MMO's creators?
How many of those 'cracked servers' are an accurate duplication of the original servers' mob and boss mechanics?
How many of those 'cracked servers' accurately duplicate the game play experience on normal live servers?

I have tried playing WoW on many "private servers" and most mobs have little more than follow and auto-attack scripts attached to them. There isn't much point in playing such a massively neutered version of PvE content.

I side with InvalidError 100% here. If you want to do something, do it right (and lawfully, lol).

Also, thanks for the follow ups on the purported problems it has in gaming.

Will you guys try and get a bit deeper? 😀

Cheers!

 

[anbello262]

Actually, there are cracked servers of many popular mmorpgs

And how many of those 'cracked servers' aren't considered a violation of the ToS by the original MMO's creators?
How many of those 'cracked servers' are an accurate duplication of the original servers' mob and boss mechanics?
How many of those 'cracked servers' accurately duplicate the game play experience on normal live servers?

I have tried playing WoW on many "private servers" and most mobs have little more than follow and auto-attack scripts attached to them. There isn't much point in playing such a massively neutered version of PvE content.

I'm sorry, I might have not expressed my point correctly in my previous post.
I wasn't trying to talk about the cracked servers themselves, my main point was that those servers are usually run on workstation level hardware and/or use cheap-ish hosting, and they work well enough for a lot more people than would be necessary to conduct a benchmark.

I am not saying "use those cracked servers". I was just giving my opinion (with some evidence) about your claim that "running a bechmark server would cost way too much and require several powerful servers". No, it wont. You only require that amount of server power if you want to have tens of thousands of people at the same time with no lag. For any benchmark scenario, 50 to 100 (maybe 200 max?) would be more than enough, and that can be run with simple hardware, or take up very little server power in the official servers.



As a little extra, I don't think it's important "how many" of those can give you a playable experience. With it just being 1 is enough to prove my point about the requirements, and about the game files already being out there. And even more important, even if it isn't enjoyable, it will probably be basically the same in resource taxing, for a benchmark.

 

[InvalidError]

I am not saying "use those cracked servers". I was just giving my opinion (with some evidence) about your claim that "running a bechmark server would cost way too much and require several powerful servers". No, it wont. You only require that amount of server power if you want to have tens of thousands of people at the same time with no lag.

Most of those private servers are heavily dumbed down compared to the real ones, they are just quick and dirty hacks to get the game clients working with them, not full-blown re-implementations. Blizzard and other MMO operators won't custom-build a production-grade server environment optimized to accommodate a 50 people benchmark server when the server-side architecture is geared for scaling across multiple racks to 10 000+ simultaneous players. That sort of infrastructure does not necessarily scale well down to a single system handling everything due to the overheads introduced to make that scaling possible.

 

[-Fran-]

I am not saying "use those cracked servers". I was just giving my opinion (with some evidence) about your claim that "running a bechmark server would cost way too much and require several powerful servers". No, it wont. You only require that amount of server power if you want to have tens of thousands of people at the same time with no lag.

Most of those private servers are heavily dumbed down compared to the real ones, they are just quick and dirty hacks to get the game clients working with them, not full-blown re-implementations. Blizzard and other MMO operators won't custom-build a production-grade server environment optimized to accommodate a 50 people benchmark server when the server-side architecture is geared for scaling across multiple racks to 10 000+ simultaneous players. That sort of infrastructure does not necessarily scale well down to a single system handling everything due to the overheads introduced to make that scaling possible.

Counterpoint: Blizzard hosts special events in local hardware in competitions.

I can remember they do allocate specific servers for the High Ranking Arena matches in official tournaments; or at least, I remember they did. Same with Valve and CS:GO.

And talking about Valve, TF2 with their robots vs humans update can be an interesting test. You can have a customized server+client running to put stress on the system. I have run local servers for my friends while playing and it's quite interesting to play with them. The more people you have in it, the more taxing it becomes.

Just food for thought 😀

Cheers!

 

[InvalidError]

I can remember they do allocate specific servers for the High Ranking Arena matches in official tournaments; or at least, I remember they did. Same with Valve and CS:GO.

None of those are anywhere near the scale of MMORPGs and none of those have a persistent world either.

 

[-Fran-]

I can remember they do allocate specific servers for the High Ranking Arena matches in official tournaments; or at least, I remember they did. Same with Valve and CS:GO.

None of those are anywhere near the scale of MMORPGs and none of those have a persistent world either.

The point wasn't scale of it, but feasibility to do so. What you want to put in the servers is up to the lender and as long as you can run a benchmark in them, you should not give a damn about the server itself and focus on the client side IMO.

Think of it as the same scale as "closed beta testing" for MMORPGs; I have participated in a lot of them and the difference client side is small, unless the servers get hammered (as it's usually the case).

Cheers!

 

[palladin9479]

I am not saying "use those cracked servers". I was just giving my opinion (with some evidence) about your claim that "running a bechmark server would cost way too much and require several powerful servers". No, it wont. You only require that amount of server power if you want to have tens of thousands of people at the same time with no lag.

Most of those private servers are heavily dumbed down compared to the real ones, they are just quick and dirty hacks to get the game clients working with them, not full-blown re-implementations. Blizzard and other MMO operators won't custom-build a production-grade server environment optimized to accommodate a 50 people benchmark server when the server-side architecture is geared for scaling across multiple racks to 10 000+ simultaneous players. That sort of infrastructure does not necessarily scale well down to a single system handling everything due to the overheads introduced to make that scaling possible.

It would be possible if, and this is a huge ass if, the developer designed their back end infrastructure to be portable and modular. In Blizzard's case it's obvious they did, but other MMO's might not be able to do that. MMO's are just giant databases with front end correlators and lobby servers. Sometimes those databases can be scaled down and run on a single system, other times they are designed to require this massive infrastructure just to properly start up.

 

[anbello262]

I understand that my first option is highly debatable, and the discussion will keep going with no definitive conclusion, and therefore will remain as nothing more than an idea.

But what about the second option?
Instanced spaces to run the benchmark inside the same server. No different than an instanced dungeon, so common in so many mmorpgs. No extra hosting required, only programming the bots (or npcs) and adding a sufficient number of mobs/npcs to make it realistically taxing.
The "benchmark dungeon" would only be accessible by certain registered reviewers with verified identities. One different instance for each registered review group/site.

 

[daneracer]

Unless I missed it, the benchmarks for Project Cars does not mention how many cars on track. This can make a big difference in the results. Also no mention whether AA was used. I presume it was not. I will get Ryzen for my office but keep intel for sims.

 

[goldstone77]

Uh yeah, well all that aside, everyone's favourite Scotsman at AdoredTV pointed out something that completely blew my mind about low-resolution gaming benchmarks. Although they do seem to make logical sense to me and to pretty much everyone else, he points out a lasting trend that will really have you scratching your head. I'm a long-time and well-decorated Tom's expert. I built my first PC 28 years ago (286-16) when I was 12 years old (yes, I'm 40 so 😛~~~), I've seen the CPU wars between not only Intel and AMD but also Cyrix and VIA. I had an ATi EGA Wonder, I've seen video cards not made not only by ATi and nVidia but also Oak, Orchid, 3dfx, Matrox, S3 and CirrusLogic. I saw the 84-key keyboard in my IBM PC model 5150 replaced by the 101-key keyboard from the IBM PC/AT. I saw bus standards like 8-bit, 16-bit ISA, 32-bit EISA and VESA, 64-bit PCI, AGP & PCI-Express. Hell, I've seen MFM, RLL, IDE and SATA but I've NEVER seen ANYTHING like he's going to describe here and to be honest, the fact that I've seen so much but never noticed what he talks about before embarrasses me. You really need to see this because THIS is some of the greatest investigative tech journalism I've seen since Charlie Demerjian exposed nVidia:
https://www.youtube.com/watch?v=ylvdSnEbL50

Keep in mind that he's comparing the 8350, which was released October 2012, to an i5 2500k that was released 1 year and 9 months earlier in January of 2011. If you look at the other benchmarks in this video, you'll see the Ivy Bridge chips (released April 2012, 6 months prior to the 8350) maintain their leads and in some cases increase it. The gains are only realized against a chip nearly 2 years older. The closest contemporary to the 2500k (since it wouldn't be fair to compare with Phenom) would be the 8150, released October 2011. That's still 9 months after Sandy, and it does not fare nearly as well despite the higher core count.

This a very skewed way of thinking because those chips are four cores while this is eight. Extra cores aren't free from an engineering point of view, they draw power, generate heat and take up die space which all could be trimmed in favor of single thread "gaming" performance. Its why we don't see "gaming" benchmarks on Intel server CPU's every though their quite expensive. Would you argue that Intel server CPU's are quite poor and those users should instead drop i5's and i7's into their infrastructure servers? This CPU is designed for very wide workloads present in higher end applications just like it's predecessor. The big benefit is that it doesn't make monstrous sacrifices to achieve that wider workload, which makes a cut down version quite appealing.

This CPU should be compared to Intel's 6c/12t offerings on a cost per cost basis. I fully expect AMD to release cut down versions with four cores which will have the same or better (clock speed) gaming performance but at a much reduced cost, making them competitive with the Intel i5's and i7's.

Actually, Infinity Fabric appears to benefit from RAM speed and lower latency even more. So, single thread performance might not be such a great indicator anymore with the new architecture of Ryzen. Here is a thread I started on the AMD forum when I noticed the significance. Also, we are seeing great results with just preliminary observations with gaming engines showing sizable gains with Ryzen only after a relatively small time. Considering the much better single thread scores of Intel vs Ryzen there are definately signs of diminishing returns vs. optimization.
CORSAIR Dominator Platinum
DDR4 2400 (PC4 19200)
Timing 10-12-12-28
Cas Latency 10
Voltage 1.35V
https://community.amd.com/message/2779568#comment-2779568

 

[InvalidError]

Actually, Infinity Fabric appears to benefit from RAM speed and lower latency even more.

If memory timings had any impact on infinity fabric, there is something horribly wrong with infinity fabric since an external interface should have no impact on internal circuitry beyond the effects of their common clock. If external timings have an effect on chip performance, it would be due to how tighter memory timings affect the memory latency and CPU cores' wait cycles.

On a possibly related note, new BIOS/microcode updates making the rounds recently appear to add 10-20% performance in some titles that had significantly sub-par performance. The bulk re-benches with the R5-series,.updated BIOS/firmware and updated games less than two weeks from now should be interesting.

 

[goldstone77]

Actually, Infinity Fabric appears to benefit from RAM speed and lower latency even more.

If memory timings had any impact on infinity fabric, there is something horribly wrong with infinity fabric since an external interface should have no impact on internal circuitry beyond the effects of their common clock. If external timings have an effect on chip performance, it would be due to how tighter memory timings affect the memory latency and CPU cores' wait cycles.

On a possibly related note, new BIOS/microcode updates making the rounds recently appear to add 10-20% performance in some titles that had significantly sub-par performance. The bulk re-benches with the R5-series,.updated BIOS/firmware and updated games less than two weeks from now should be interesting.

I'm sorry I pasted the wrong link.
https://community.amd.com/thread/214043

Here is the video that first caught me onto the idea that what I did say may hold some truth. And later was confirmed by one person in the thread above. Do you have access to Ryzen can you perform a test? https://www.youtube.com/watch?v=RZS2XHcQdqA

Also, http://wccftech.com/amds-infinity-fabric-detailed/
How could it not have an impact?

 

[palladin9479]

Actually, Infinity Fabric appears to benefit from RAM speed and lower latency even more.

If memory timings had any impact on infinity fabric, there is something horribly wrong with infinity fabric since an external interface should have no impact on internal circuitry beyond the effects of their common clock. If external timings have an effect on chip performance, it would be due to how tighter memory timings affect the memory latency and CPU cores' wait cycles.

On a possibly related note, new BIOS/microcode updates making the rounds recently appear to add 10-20% performance in some titles that had significantly sub-par performance. The bulk re-benches with the R5-series,.updated BIOS/firmware and updated games less than two weeks from now should be interesting.

It's not an external interface, the fabric is where the system memory bus is located, think old school FSB. It's why Ryzen is so picky about higher speed memory and why you can get a large performance increase from using higher clocked memory. Ryzen's design is basically two CPU's sharing a single memory bus via their system interconnect.

A more technical description is that AMD has separated the Compute (CPU / GPU) components from the data components (Memory / IO) and created a single language for them all to speak. In theory you can have as many CPU's and GPU's interconnected together with any number of memory or IO channels and they would all speak to each other on the same networked interconnect. This interconnect can be between components on die or even extended across multiple sockets or interfaces such that you can expand into supercomputer power without needing custom glue.

The upside to this is you can easily redesign any system IO implementation in days instead of months. If a system design needs more memory channels, PCIe lanes or sockets / cores then it's a snap to extend it. The downside is the growing pains associated with such an endeavor.

 

[InvalidError]

It's not an external interface, the fabric is where the system memory bus is located, think old school FSB.

The memory controller is likely only one stop on the fabric and unless AMD was incredibly dumb, the fabric should be some form of transactional bus with the memory controller having a command buffer to decouple memory latency from the internal bus/fabric/whatever. Using the memory bus for cache snoops and synchronization between CCX would also nuke memory performance - that's what killed shared FSB CPU architectures' performance and forced CPU designers to go with point-to-point links between CPUs in multi-socket configurations ~10 years ago.

 

[goldstone77]

Actually, Infinity Fabric appears to benefit from RAM speed and lower latency even more.

If memory timings had any impact on infinity fabric, there is something horribly wrong with infinity fabric since an external interface should have no impact on internal circuitry beyond the effects of their common clock. If external timings have an effect on chip performance, it would be due to how tighter memory timings affect the memory latency and CPU cores' wait cycles.

On a possibly related note, new BIOS/microcode updates making the rounds recently appear to add 10-20% performance in some titles that had significantly sub-par performance. The bulk re-benches with the R5-series,.updated BIOS/firmware and updated games less than two weeks from now should be interesting.

It's not an external interface, the fabric is where the system memory bus is located, think old school FSB. It's why Ryzen is so picky about higher speed memory and why you can get a large performance increase from using higher clocked memory. Ryzen's design is basically two CPU's sharing a single memory bus via their system interconnect.

A more technical description is that AMD has separated the Compute (CPU / GPU) components from the data components (Memory / IO) and created a single language for them all to speak. In theory you can have as many CPU's and GPU's interconnected together with any number of memory or IO channels and they would all speak to each other on the same networked interconnect. This interconnect can be between components on die or even extended across multiple sockets or interfaces such that you can expand into supercomputer power without needing custom glue.

The upside to this is you can easily redesign any system IO implementation in days instead of months. If a system design needs more memory channels, PCIe lanes or sockets / cores then it's a snap to extend it. The downside is the growing pains associated with such an endeavor.

You appear to agree with my assessment. The potential of low latency RAM vs. High speed incompatibility look like they could have a significant effect on Ryzen! Tests need to be done! Now we just need someone who has Ryzen to try that 2400MHz CAS 10 RAM!

 

[palladin9479]

It's not an external interface, the fabric is where the system memory bus is located, think old school FSB.

The memory controller is likely only one stop on the fabric and unless AMD was incredibly dumb, the fabric should be some form of transactional bus with the memory controller having a command buffer to decouple memory latency from the internal bus/fabric/whatever. Using the memory bus for cache snoops and synchronization between CCX would also nuke memory performance - that's what killed shared FSB CPU architectures' performance and forced CPU designers to go with point-to-point links between CPUs in multi-socket configurations ~10 years ago.

In order to get to main memory it needs to go through the fabric and for some reason AMD seems to run them both on the same clock generator. For data to go from one CCX to another it must traverse the fabric, for that same data to go out to a PCIe device it against must traverse this fabric.

https://www.techpowerup.com/231585/amd-ryzen-infinity-fabric-ticks-at-memory-speed

This is why installing higher speed DDR4 memory gives such a large boost in performance, it's directly raising the clock speed of the fabric interconnect between the two CCX units. The two 8MB segments of L3 are kept separate but coherency is synchronized across the fabric.

 

[goldstone77]

It's not an external interface, the fabric is where the system memory bus is located, think old school FSB.

The memory controller is likely only one stop on the fabric and unless AMD was incredibly dumb, the fabric should be some form of transactional bus with the memory controller having a command buffer to decouple memory latency from the internal bus/fabric/whatever. Using the memory bus for cache snoops and synchronization between CCX would also nuke memory performance - that's what killed shared FSB CPU architectures' performance and forced CPU designers to go with point-to-point links between CPUs in multi-socket configurations ~10 years ago.

In order to get to main memory it needs to go through the fabric and for some reason AMD seems to run them both on the same clock generator. For data to go from one CCX to another it must traverse the fabric, for that same data to go out to a PCIe device it against must traverse this fabric.

https://www.techpowerup.com/231585/amd-ryzen-infinity-fabric-ticks-at-memory-speed

This is why installing higher speed DDR4 memory gives such a large boost in performance, it's directly raising the clock speed of the fabric interconnect between the two CCX units. The two 8MB segments of L3 are kept separate but coherency is synchronized across the fabric.

Then you agree the lower latency RAM is going to have a significant increase in Infinity Fabric performance. Everyone has been testing with 2800-3200MHz and say 16 or 15 CAS. How about 2400 at 10 CAS that would be much faster than the 3200 at 16 or 15 CAS.