Ryzen with 3600MHz RAM Benchmark

[CSI_PC]

Yep, that was a good review. I think you should pick one of these up just because events like these don't happen that often anymore. Going from Bulldozer to RyZen is such an incredible jump or more like miracle for AMD. Also OCing is more like the old fashion way where you have to grind away to get somewhere but is rather fun.

I possibly will but not until they release the updated version and I see that reviewed, so I am giving it 6 months and will be the minor update Ryzen I expect them to launch or Skylake HEDT.
Reason I am considering Skylake HEDT is the number of PCIe lanes as I will be using a fair bit of NVMe and also looking to the future next year with something similar to Optane - so weighing choices that may include the AMD 'HEDT'.

Micron I think is launching their own product in 6-9 months on enterprise so will be interesting to see how that fits but Optane has nice hooks into the system as a cache solution that comfortably fits a whole loaded game and other functions.

Cheers

 

[OrangeKhrush]

Yep, that was a good review. I think you should pick one of these up just because events like these don't happen that often anymore. Going from Bulldozer to RyZen is such an incredible jump or more like miracle for AMD. Also OCing is more like the old fashion way where you have to grind away to get somewhere but is rather fun.

I like your systems

 

[OrangeKhrush]

I guess cache latency is not a problem anymore or never was and the bandwidth is crazy which if you want to run a system like the 1800X you should be running 3400mhz as status quo.

I am hoping to se re-benches done with DDR4 3400/3600 at some point over the next month or so once board partners are happy with the ucode.

 

[CSI_PC]

I guess cache latency is not a problem anymore or never was and the bandwidth is crazy which if you want to run a system like the 1800X you should be running 3400mhz as status quo.

I am hoping to se re-benches done with DDR4 3400/3600 at some point over the next month or so once board partners are happy with the ucode.

Not quite, look at how hardware.fr showed different behaviour depending upon the dataset size and whether it was 1st core or cross complex.
What size data set is used for the L3 cache in that Aida benchmark?
How does it use affinity and test one CCX and then across both?

It is not a simple test unfortunately to test the Ryzen architecture as there are a couple of different mechanisms involved that will have different behaviour and performance results, such as thread dependency on other CCX compared to cache data dependency on other CCX, but as I mentioned there is also more such as a complete cache coherency test.
That said I think the bigger challenge for games is probably thread dependency, but it can also come down to data set size (unusual behaviour seems to start above 4MB both local and more broader with data on other CCX) with the cache coherency to some extent.
Cheers

 

[OrangeKhrush]

Not quite, look at how hardware.fr showed different behaviour depending upon the dataset size and whether it was 1st core or cross complex.
What size data set is used for the L3 cache in that Aida benchmark?
How does it use affinity and test one CCX and then across both?

It is not a simple test unfortunately to test the Ryzen architecture as there are a couple of different mechanisms involved that will have different behaviour and performance results, such as thread dependency on other CCX compared to cache data dependency on other CCX, but as I mentioned there is also more such as a complete cache coherency test.
That said I think the bigger challenge for games is probably thread dependency, but it can also come down to data set size (unusual behaviour seems to start above 4MB both local and more broader with data on other CCX) with the cache coherency to some extent.
Cheers

They to will have to retest to see what improvement is made with all the updates to date to see if some issues were corrected.

 

[CSI_PC]

They to will have to retest to see what improvement is made with all the updates to date to see if some issues were corrected.

And what updates do you think is going to improve this?
The motherboard update is to improve DDR4 behaviour/voltages/maybe frequency turbo performance-behaviour/timings/etc but not core CPU related functions, motherboard manufacturers would not be responsible for cache coherency changes let alone changes to the Data Fabric.
Are you aware of any AMD updates released, even via all the motherboard manufacturers?

But more importantly, no-one can say how AIDA works in terms of data set sizes (4MB onwards has the unusual behaviour in Ryzen) nor how they test if at all between CCX (which I doubt they do).

You need the bespoke tools like I mentioned in my previous post for the reasons I mentioned back then as well that comes down to difference in mechanism and behaviour with performance for thread dependency/data dependency/cache coherency.
Cheers

 

[CSI_PC]

Just to say if AMD could had created a fully coherent mesh at 8C (so 16 threads) with same behaviour up to 16MB as one CCX it would had been one mutha of a CPU.
Shame they had to go with the 4C CCX design and also the L3 split 2MB per core cache coherency (some of the early tests we seen so far from hardware.fr indicate this at times is not behaving fully as expected), but makes sense from a cost perspective and accessibility to more customers as it would had been a lot more expensive and technically very challenging to build and manufacturer.

Cheers

 

[OrangeKhrush]

And what updates do you think is going to improve this?
The motherboard update is to improve DDR4 behaviour/voltages/maybe frequency turbo performance-behaviour/timings/etc but not core CPU related functions, motherboard manufacturers would not be responsible for cache coherency changes let alone changes to the Data Fabric.
Are you aware of any AMD updates released, even via all the motherboard manufacturers?

But more importantly, no-one can say how AIDA works in terms of data set sizes (4MB onwards has the unusual behaviour in Ryzen) nor how they test if at all between CCX (which I doubt they do).

You need the bespoke tools like I mentioned in my previous post for the reasons I mentioned back then as well that comes down to difference in mechanism and behaviour with performance for thread dependency/data dependency/cache coherency.
Cheers

from what I have seen and things mentioned the stability in getting ram to 3400 has generally helped so in some way getting stable RAM at higher speed is helping somewhat to improving performance and consistency in performance.

i am not sure how AIDA tests cache, I assume it is on a ping rate.

Platform stability is more important at this point which is visible from the differences posted around and we are not a month removed yet.

All vendors have had multiple releases to bios and ucode updates, MSI already have a stable 3400+ A XMP

 

[CSI_PC]

from what I have seen and things mentioned the stability in getting ram to 3400 has generally helped so in some way getting stable RAM at higher speed is helping somewhat to improving performance and consistency in performance.

i am not sure how AIDA tests cache, I assume it is on a ping rate.

Platform stability is more important at this point which is visible from the differences posted around and we are not a month removed yet.

All vendors have had multiple releases to bios and ucode updates, MSI already have a stable 3400+ A XMP

But your post in response to mine was suggesting those bespoke tools should be done again because the motherboard updates would influence results.
As you point out now those updates are focused on other aspects such as DDR4 stability (so agreeing with my point) and so surely you agree there is no point running those tests again until AMD actually provides an update to all motherboard manufacturers relating to core Ryzen function updates.
These guys only have so much time/resources and so will not repeat tests until it is actually worth doing them.

There is a difference in the type of update needed as I mentioned and what we are currently seeing are just from the motherboard vendors for the factors I mentioned.
I assume we can agree that no-one knows how AIDA is working from a dataset perspective nor whether it fixes to one core or tests across CCXs (but then AIDA would not know that the CPU is technically 2 CCX).
Cheers

 

[OrangeKhrush]

Just to say if AMD could had created a fully coherent mesh at 8C (so 16 threads) with same behaviour up to 16MB as one CCX it would had been one mutha of a CPU.
Shame they had to go with the 4C CCX design and also the L3 split 2MB per core cache coherency (some of the early tests we seen so far from hardware.fr indicate this at times is not behaving fully as expected), but makes sense from a cost perspective and accessibility to more customers as it would had been a lot more expensive and technically very challenging to build and manufacturer.

Cheers

What makes you think they didnt' try that route and maybe that route what seems to you as working is actually worse, maybe that route you have a bunch of cores tripping over themselves trying to help themselves to unified cache pools. I really doubt AMD just decided...."hmmm boys plan a works but like lets just thow a CCX there and one there and just like troll the buyers". There is obviously limitations to your approach, my guess is chaos unfolding trying to stop cores cannibalizing another.

 

[CSI_PC]

What makes you think they didnt' try that route and maybe that route what seems to you as working is actually worse, maybe that route you have a bunch of cores tripping over themselves trying to help themselves to unified cache pools. I really doubt AMD just decided...."hmmm boys plan a works but like lets just thow a CCX there and one there and just like troll the buyers". There is obviously limitations to your approach, my guess is chaos unfolding trying to stop cores cannibalizing another.

The cost would be very high, the coherent mesh aspect would be horrendous technically and I doubt they could do the split 2MB (so 2x1MB) per core in that setup as well.
They would probably have to charge the same as Intel, who IMO would have more wriggle room on pricing as AMD would have a tighter margin as well.
It is a design us enthusiasts would had loved but AMD would lose accessibility to the mainstream, they made the right choice IMO even though I would prefer the 'mutha of a all CPUs' myself and I bet quite a few here would as well but the cost would put many off even with the consistent and high performance it would have.
Cheers

 

[OrangeKhrush]

But your post in response to mine was suggesting those bespoke tools should be done again because the motherboard updates would influence results.
As you point out now those updates are focused on other aspects such as DDR4 stability (so agreeing with my point) and so surely you agree there is no point running those tests again until AMD actually provides an update to all motherboard manufacturers relating to core Ryzen function updates.
These guys only have so much time/resources and so will not repeat tests until it is actually worth doing them.

There is a difference in the type of update needed as I mentioned and what we are currently seeing are just from the motherboard vendors for the factors I mentioned.
I assume we can agree that no-one knows how AIDA is working from a dataset perspective nor whether it fixes to one core or tests across CCXs (but then AIDA would not know that the CPU is technically 2 CCX).
Cheers

It is still probably work in progress until later this year but if AMD can systematically work problems away that is fine, it means that the teething of summit ridge will be ironed out by Pinnacle ridge. From testers with B2 and B3 silicon have said from numerous sources taht behaviour of B1 is not as pronounced or largely removed in B2

 

[lolfail9001]

I guess cache latency is not a problem anymore or never was

Intra-CCX cache was never the issue, and new version of AIDA64 measures precisely intra-CCX.

Inter-CCX (i.e. memory access) is where things get dicey.

 

[OrangeKhrush]

Intra-CCX cache was never the issue, and new version of AIDA64 measures precisely intra-CCX.

Inter-CCX (i.e. memory access) is where things get dicey.

I am still following the improvements made to inter communication, the thing is it works well often enough to highlight that it is not a fatal issue and the few isolated incidents it doesn't work it seems to still be related to how fast the memory bus works. At this stage there is a lot of tight lips about later revisions to know exactly how it is being addressed. From AMD's internal statement they are not concerned and seem to know the limitations.

 

[CSI_PC]

Intra-CCX cache was never the issue, and new version of AIDA64 measures precisely intra-CCX.

Inter-CCX (i.e. memory access) is where things get dicey.

I would like to see more tests like those done by hardware.fr where they show the latency doubles when the data set in Cache went just above 4MB on same CCX.
Considering the 2MB per core is split as 2x1MB, may suggest there is also something unusual at times (stress not always) happening even on the same CCX in terms of cache coherency.

Cheers

 

[OrangeKhrush]

I would like to see more tests like those done by hardware.fr where they show the latency doubles when the data set in Cache went just above 4MB on same CCX.
Considering the 2MB per core is split as 2x1MB, may suggest there is also something unusual at times (stress not always) happening even on the same CCX in terms of cache coherency.

Cheers

The issue is pronounced in gaming over workload which talks more towards the behaviour of game code in particularity to Ryzen uArch. It is trite now that faster ram improves the situation as much as 20% going from a 2133 kit to 3200 kit, after that it is still not a lineated gain, can be as high as 35% goin from 2133 to 3600mhz so is the issue still the memory bus speed holding the system back. In linear loads it is less relevant so Ryzen does what it does crushes.

 

[Formula.350]

I would like to see more tests like those done by hardware.fr where they show the latency doubles when the data set in Cache went just above 4MB on same CCX.

I can try to do this for you, if you want? I just need to update AIDA as I see in that one screenshot above that they're using build 4200 and mine is 4089. While not as bad as my original 5.80 release, which didn't even call it Ryzen just "16 Core Summit-Ridge", 5.80.4089 still will mention after latency tests that "AIDA is not fully optimized for this CPU".

Nevertheless, there is a hidden test in AIDA that runs the Cache & Memory latency from like 512bytes up to 256MB, in 4 ways, and plots it all out on a graph for you. As such, it'll take a few reboots, and I won't be able to "fully" test with SMT disabled since I'm not ready to update my BIOS to add that option, but I can at least set Affinity to just 4 of the threads after configuring the system as 4+0.

EDIT: Newest AIDA acquired. It's not just the changed build, but a minor-version increase as well, to 5.90 (build 4200). So far it seemed to have resolved the L3 issue with sporadic results, which definitely helps me in figuring out if changes I've made in the BIOS are the result of result changes in AIDA heh Otherwise the overall performance hasn't changed drastically from 5.80.4089. However, I can say that going from the VERY unsupported version 5.70.3800 (where it only recognizes it as a 16 Core Summit-Ridge chip), performance has been impacted. BUT I don't know if that's legitimate or if it's something hinky... They claim they don't let benchmarks compare between updates to their 'BenchDLL' code, due to "optimizations", so it's curious that it's optimized and ends up not performing as well lol On the Memory Read, it's 1000mb/s lower as an example.

But anyways, 4+0 is booted now, so will re-run the test(s) to compare. I can say that Win10 appeared to have loaded faster... I was more focused on typing here but the loading screen wasn't up for as long as usual. *shrug* Will re-edit shortly with results.

 

[CSI_PC]

The issue is pronounced in gaming over workload which talks more towards the behaviour of game code in particularity to Ryzen uArch. It is trite now that faster ram improves the situation as much as 20% going from a 2133 kit to 3200 kit, after that it is still not a lineated gain, can be as high as 35% goin from 2133 to 3600mhz so is the issue still the memory bus speed holding the system back. In linear loads it is less relevant so Ryzen does what it does crushes.

Faster RAM works on both Kaby Lake and Ryzen equally, linked that data not long ago and several times now.
Cache coherency cannot be underestimated in its importance and how shared L3 well used can help performance even with games, more indirectly but just look at the gaming performance of Broadwell 5775C with EDRAM as an L4 Cache even though it is only 65W and clocks for shit.

Cheers

 

[OrangeKhrush]

Faster RAM works on both Kaby Lake and Ryzen equally, linked that data not long ago and several times now.
Cache coherency cannot be underestimated in its importance and how shared L3 well used can help performance even with games, more indirectly but just look at the gaming performance of Broadwell 5775C with EDRAM as an L4 Cache even though it is only 65W and clocks for shit.

Cheers

I was speaking to someone today, im interested in a 1600, i was told it is a good option as it is B2 revision of R7, i wonder if AMD tuned the fabric between revisions knowing that B1 had to be sold out or sit with surplus on B1 chips. I think it was planned that way. Sell all old stocks on B1 then drop the B2 fbomb

 

[CSI_PC]

I was speaking to someone today, im interested in a 1600, i was told it is a good option as it is B2 revision of R7, i wonder if AMD tuned the fabric between revisions knowing that B1 had to be sold out or sit with surplus on B1 chips. I think it was planned that way. Sell all old stocks on B1 then drop the B2 fbomb

Possible, will be interesting as the sites will test it with their tools.
Cheers

 

[Formula.350]

Decided make a new post instead of update since network issues slowed me down...

RESULTS: For starters, I can't override AIDA from using All Threads through setting affinity in Task Manager due to it opening a separate process to run benchmarks. Additionally, the Parameters option to disable HT only applies to CPU and FPU benchmarks. As mentioned, I lack the SMT option in my BIOS currently.

System Specs:
Ryzen 7 1700X @ Stock Speeds, Stock Voltages (except CPU-NB running at 1V instead of 0.90V for no honestly particular reason)
MSI X370 Titanium BIOS v1.10 (shipping version)
DDR4-3200 14-15-15-35 1.36V (Set at 1.35V) G.Skill TridentZ F4-3200C15D-16GTZKY (See attached image for complete timings).
Scythe Katana 3 (Small HSF but doing just fine so far)
.... And the irrelevant hardware for these tests can be found in my sig. (Though currently HDDs are a 160GB SATA-II Velociraptor and a 2.5" 500GB SATA-II WD Black that usually live in a USB3 enclosure)

What's most surprising here is that the cache performance has gone down while Latency remaining unchanged! I suppose one can look at that either way for being good or bad. So far look like 'good' as far as the upcoming 6C/12T models are concerned, but less thrilling for single module 4C/8T models. Also, non-cache CPU/FPU tests where AIDA is set to HT Off, one or two show gains and those results are consistent between runs. (even at full 4+4)
(The Cache & Mem Latency test on the bottom is too big for side-by-side)

------------------- 8C/16T------------------------------------- 4+0 4C/8T ----------------

 

[CSI_PC]

Decided make a new post instead of update since network issues slowed me down...

RESULTS: For starters, I can't override AIDA from using All Threads through setting affinity in Task Manager due to it opening a separate process to run benchmarks. Additionally, the Parameters option to disable HT only applies to CPU and FPU benchmarks. As mentioned, I lack the SMT option in my BIOS currently.

System Specs:
Ryzen 7 1700X @ Stock Speeds, Stock Voltages (except CPU-NB running at 1V instead of 0.90V for no honestly particular reason)
MSI X370 Titanium BIOS v1.10 (shipping version)
DDR4-3200 14-15-15-35 1.36V (Set at 1.35V) G.Skill TridentZ F4-3200C15D-16GTZKY (See attached image for complete timings).
Scythe Katana 3 (Small HSF but doing just fine so far)
.... And the irrelevant hardware for these tests can be found in my sig. (Though currently HDDs are a 160GB SATA-II Velociraptor and a 2.5" 500GB SATA-II WD Black that usually live in a USB3 enclosure)

What's most surprising here is that the cache performance has gone down while Latency remaining unchanged! I suppose one can look at that either way for being good or bad. So far look like 'good' as far as the upcoming 6C/12T models are concerned, but less thrilling for single module 4C/8T models. Also, non-cache CPU/FPU tests where AIDA is set to HT Off, one or two show gains and those results are consistent between runs. (even at full 4+4)
(The Cache & Mem Latency test on the bottom is too big for side-by-side)

View attachment 20621 View attachment 20622

------------------- 8C/16T------------------------------------- 4+0 4C/8T ----------------
View attachment 20623 View attachment 20624

Thanks for doing the work.
I am not surprised about the latency figures doing the 8C vs 4C (4+0) as both would still be using the coherent L3 cache in same way.
You prove that something is rather unusual as hardware.fr noted after 4MB data set and this should not happen as a CCX complex should mean consistent latency for whole of the shared 8MB, I do wonder if it is down to the 2MB per core being split as 2x1MB *shrug* and has an influence on cache coherency.
Separately I would had expected more of a distinction around the 24-32MB range as it should have its own latency not similar to either lower or higher sizes neighbouring it.

Thanks again and also for looking at how it uses threads to test the cache.

 

[Formula.350]

Thanks for doing the work.
I am not surprised about the latency figures doing the 8C vs 4C (4+0) as both would still be using the coherent L3 cache in same way.
You prove that something is rather unusual as hardware.fr noted after 4MB data set and this should not happen as a CCX complex should mean consistent latency for whole of the shared 8MB, I do wonder if it is down to the 2MB per core being split as 2x1MB *shrug* and has an influence on cache coherency.
Separately I would had expected more of a distinction around the 24-32MB range as it should have its own latency not similar to either lower or higher sizes neighbouring it.

Thanks again and also for looking at how it uses threads to test the cache.

Most welcome.

First, I doubt it does, but if I'm understanding how the test works, it's using 512Byte chunks of data (or whatever), which while fiddling with UEFI BIOS programs earlier today the one detailed the Cache, and all three showed (accuracy questionable) that their cache memory is structured in 1KB blocks. It'd be interesting to see what the results would be with "1024-byte Stride" if one could set that. Like with storage tests they test that way to illustrate the highest it can perform in ideal situations. Granted, most situations aren't ideal, hence just the curiosity to see what it'd at least show

As for why there's the shift in performance, I simply think it comes down to the die configuration on these Ryzens, as the core modules are setup as:
2C-Cache-2C
and then each module on the die is arranged
2C-Cache-2C 2C-Cache-2C
which to me means there is going to be an inherent latency when communicating between L3's.
Had they been oriented "vertical" instead of horyzontally *snickers*, looking more like..
2C 2C
L# L#
2C 2C
(L# being L-cache 1,2,3) That'd have allowed for much much quicker communication between the L3, providing for a far more efficient CCX structure as a result.
Can only hope that's how they arrange it on Naples, as 4 rectangle modules make a rather compacted square (I know, that's horrible logic lol)

 

[CSI_PC]

Most welcome.

First, I doubt it does, but if I'm understanding how the test works, it's using 512Byte chunks of data (or whatever), which while fiddling with UEFI BIOS programs earlier today the one detailed the Cache, and all three showed (accuracy questionable) that their cache memory is structured in 1KB blocks. It'd be interesting to see what the results would be with "1024-byte Stride" if one could set that. Like with storage tests they test that way to illustrate the highest it can perform in ideal situations. Granted, most situations aren't ideal, hence just the curiosity to see what it'd at least show

As for why there's the shift in performance, I simply think it comes down to the die configuration on these Ryzens, as the core modules are setup as:
2C-Cache-2C
and then each module on the die is arranged
2C-Cache-2C 2C-Cache-2C
which to me means there is going to be an inherent latency when communicating between L3's.
Had they been oriented "vertical" instead of horyzontally *snickers*, looking more like..
2C 2C
L# L#
2C 2C
(L# being L-cache 1,2,3) That'd have allowed for much much quicker communication between the L3, providing for a far more efficient CCX structure as a result.
Can only hope that's how they arrange it on Naples, as 4 rectangle modules make a rather compacted square (I know, that's horrible logic lol)

Well hardware.fr had similar trend results in places to you apart from once they reached 24-32MB dataset size and the 6MB (although same behaviour with 4-core setup), and the timings are moderately aligned as well for the most part up to 4MB.
But then even yours is not consistent up to 8MB on the 4-core setup but it is not fully equal to the 8-core in test.

Any chance you can look at the 8-core test and with the window-data to the maximum of 128MB just like the 4-Core?
Yeah being paranoid here I know but curious the 4-core setup is behaving like hardware.fr for 6MB.

Thanks

 

[Raptorium]

Hi, I'm new here but found this thread very useful, I'm just buying my first new pc build in quite a while and I decided to go with Ryzen because of the multi-threading for 3d and video editing (plus it just feels far more exciting than intel at the moment). Can anyone recommend a good 3600 DDR4 ram set? I was looking at the Corsair Vengeance LPX, any experience with that one? (http://www.corsair.com/en-gb/vengea...600mhz-c16-memory-kit-red-cmk32gx4m4b3600c16r) Is there a benefit for keeping it at 2x8, or would 4x8 still work fine? Thx

 

[Formula.350]

Any chance you can look at the 8-core test and with the window-data to the maximum of 128MB just like the 4-Core?

Whoa I just realized what you're referring to... I hadn't noticed that there was a test size discrepancy! That's interesting and peculiar... I have no clue as to why that happened other than to suspect as a result of overall Cache Size changing due to disablement of a 4C module.

I've done my absolute best to modify the first 8C/16T to only be the 128MB worth of data and still retain the correct overall size as the 4C/8T for ease of back-and-forth comparisons.

 

[spine]

At the 1:20 mark in the video he explains the reason he didn't run 3600MHz RAM speed on the i7 7700k, reason being that 3600MHz made little to no difference on the 7700k in "these" benches.
So beyond that, I would say they were max vs max, with Ryzen at 4GHz(3.97) and the i7 at 5GHz.
Some would say that's fair considering the superior IPC on the i7.

No, it shows obvious bias by deliberately hiding data and presenting interpreted results.

As soon as he explained that 3600 part I just had to turn it off because I knew at that point I couldn't believe any of whatever data he'd show, simply because I don't know what he's not showing.

And that's assuming I believe his testing was both perfect and devoid of his obvious desire for AMD to do well.

 

[CAD4466HK]

Hi, I'm new here but found this thread very useful, I'm just buying my first new pc build in quite a while and I decided to go with Ryzen because of the multi-threading for 3d and video editing (plus it just feels far more exciting than intel at the moment). Can anyone recommend a good 3600 DDR4 ram set? I was looking at the Corsair Vengeance LPX, any experience with that one? (http://www.corsair.com/en-gb/vengea...600mhz-c16-memory-kit-red-cmk32gx4m4b3600c16r) Is there a benefit for keeping it at 2x8, or would 4x8 still work fine? Thx

Good luck getting 3600MHz RAM to run @3600MHz.
The upper "norm" seems to be 2933MHz(3000) for most, 3200MHz speeds are certainly doable depending on what mobo you have.
For now Ryzen seems to operate better with only 2 banks populated if your going for higher RAM speeds, seen a few peeps with 4 DIMMs hitting 2666MHz.
But once again it depends on your mobo and the BIOS flavor of the week.
Once you get your mobo, comb through it's QVL to get the best RAM certified for you that fits your needs.

 

[CAD4466HK]

No, it shows obvious bias by deliberately hiding data and presenting interpreted results.

As soon as he explained that 3600 part I just had to turn it off because I knew at that point I couldn't believe any of whatever data he'd show, simply because I don't know what he's not showing.

And that's assuming I believe his testing was both perfect and devoid of his obvious desire for AMD to do well.

I wouldn't go as far to call him biased, but the problem I am having with this "benchmark" after it being some time since I posted it,
is where the fuck are all the other people hitting 3600MHz on their builds?!?
I figured by now we would of had a shit ton worth of benches and other user reviews showing 3600MHz speeds in the wild.
Leaves me believing that it was a hoax or a fluke.

 

[Formula.350]

Hi, I'm new here but found this thread very useful, I'm just buying my first new pc build in quite a while and I decided to go with Ryzen because of the multi-threading for 3d and video editing (plus it just feels far more exciting than intel at the moment). Can anyone recommend a good 3600 DDR4 ram set? I was looking at the Corsair Vengeance LPX, any experience with that one? (http://www.corsair.com/en-gb/vengea...600mhz-c16-memory-kit-red-cmk32gx4m4b3600c16r) Is there a benefit for keeping it at 2x8, or would 4x8 still work fine? Thx

There's a few things you, and we, will need to know before you go about buying stuff. (Sorry if the following comes across at all patronizing, I just make it a habit to assume the other person may not have as much knowledge as another, so I try to provide as much detail as possible. Plus it helps for anyone else looking for info )

First and foremost, the Ryzen platform is in its infancy and currently getting memory running at 3200MHz is hit and miss. To get above 3200Mhz, assuming you can get there to begin with, will require getting a motherboard that has the ability to change the Base Clock speed (aka BCLK Frequency). However, raising the BCLK can have undesirable side effects as well (with PCIe speeds and also potentially causing the chipset to totally flake out).

Second, keeping the above's intro in mind, the systems also have trouble running large size modules at decent speeds, but also have just the same kind of issue if running with 4 sticks of memory. So, indeed, a bit of a precarious situation this can leave a person in! If you NEED a lot of memory, no matter how you slice it, getting Samsung B-die equipped memory is going to leave you with the best chances of running them at the highest speeds. Whether you opt to get 2x16GB or 4x8GB, the results seem to be hit and miss for either. If you DO get 8GB sticks, even then you'll want to make sure that they are "1-Rank" modules, and determining that is not very easy until after you get them, so research on forums and looking at CPU-Z screenshots of the "SPD" tab will help.

Third, comes down to telling us what model of Motherboard you were thinking of getting, as that as well will pay a role in all of this. Lastly, that if you're using a kit with only 2 sticks, to place them in the memory slots so it looks like this, where o is "open" and | is the stick of RAM: x-|-x-|

While I haven't looked around much... Here on [H], of those I know for certain, we have Topweasel who has an ASRock X370 Taichi and is running 32GB using the G.Skill Ripjaws 5 F4-3200C15D-32GVK 2x16GB 15-15-15-35 kit. His speeds and timings I'm pretty sure are in the Ryzen Memory Speeds thread, but has gotten over 3000MHz.
I had also seen one person on the XtremeSystems forum accomplish 32GB using an ASRock X370 Taichi and the G.Skill TridentZ F4-3300C16Q-32GTZ 4x8GB 16-16-16-26 CR2 kit, but at DDR4-2666 14-14-14-34.

The best "guarantee" (which even still may not result in 100% success) would be getting either of G.Skill's Ryzen intended kits of Fortis or Flare X. The Flare X is the better choice of them, but due to being 'specially for Ryzen' means it's price will be higher than an identical kit of TridentZ.

Ryzen runs with a Command Rate of 1, which sometimes will show up as 1N at the end of the timings (like in the above links, but they are both rated at 2N) or listed elsewhere as CR1. So far the Samsung B-die seem to be handling that just fine, and I suspect that's probably why a lot of other kits with SK-Hynix chips aren't able to run very fast.

AND WHATEVER YOU DO, STAY AWAY FROM ANY MEMORY WITH RGB LIGHTING!

To finally answer your question on that specific kit... the only Corsair kit I've come across so far that has B-Die (I've not been looking though), is the 3600 that has 16-18-18 timings. BUT most B-die have timings like that one does of 16-16-16 (all the same number), so there's a good chance they have them. Again, research will have to be done... :\

 

[Formula.350]

I figured by now we would of had a shit ton worth of benches and other user reviews showing 3600MHz speeds in the wild.
Leaves me believing that it was a hoax or a fluke.

My interpretation of this, being it's in the AMD thread and by a person who has also been playing around with Ryzen, is was what he has managed to get one of his magical kits to run at:
http://www.xtremesystems.org/forums...-of-the-Jedi&p=5256468&viewfull=1#post5256468

3600 @ 12-11-11-21 is pretty impressive lol

 

[Raptorium]

Good luck getting 3600MHz RAM to run @3600MHz.
The upper "norm" seems to be 2933MHz(3000) for most, 3200MHz speeds are certainly doable depending on what mobo you have.
For now Ryzen seems to operate better with only 2 banks populated if your going for higher RAM speeds, seen a few peeps with 4 DIMS hitting 2666MHz.
But once again it depends on your mobo and the BIOS flavor of the week.
Once you get your mobo, comb through it's QVL to get the best RAM certified for you that fits your needs.

I ordered the Gigabyte Gaming 5 with Ryzen 1700x, as it seemed a popular option and has everything i'd need (http://www.gigabyte.us/Motherboard/GA-AX370-Gaming-5-rev-10#kf). According to the gigabyte QVL, it says that corsair vengeance lpx would be supported, but just not sure if i should go for a higher clock 3600 lpx like that or stick with the 3200 speed like you mentioned. The G-skill Flare-X is also on the qvl, but do you think there would be a benefit to go with the higher clock 3600 over that memory's 3200? Both are price similarly, in fact the 3600 corsairs are a little cheaper. As far as i can tell also, it seems that its better to keep the 2 DIMMs to 8gb and not 16gb, correct ?

 

[DooKey]

Flare X 3200 is back in stock at Newegg.

 

[CAD4466HK]

My interpretation of this, being it's in the AMD thread and by a person who has also been playing around with Ryzen, is was what he has managed to get one of his magical kits to run at:
http://www.xtremesystems.org/forums...-of-the-Jedi&p=5256468&viewfull=1#post5256468

3600 @ 12-11-11-21 is pretty impressive lol

Good find, and magical is right, in the Apple sense mabey, but interesting thread non the less.
Me thinks I'll read through the whole thread to make my own opinion.
God knows I haven't been to XS since Fugger was a thing.

 

[Raptorium]

There's a few things you, and we, will need to know before you go about buying stuff. (Sorry if the following comes across at all patronizing, I just make it a habit to assume the other person may not have as much knowledge as another, so I try to provide as much detail as possible. Plus it helps for anyone else looking for info )

First and foremost, the Ryzen platform is in its infancy and currently getting memory running at 3200MHz is hit and miss. To get above 3200Mhz, assuming you can get there to begin with, will require getting a motherboard that has the ability to change the Base Clock speed (aka BCLK Frequency). However, raising the BCLK can have undesirable side effects as well (with PCIe speeds and also potentially causing the chipset to totally flake out).

Second, keeping the above's intro in mind, the systems also have trouble running large size modules at decent speeds, but also have just the same kind of issue if running with 4 sticks of memory. So, indeed, a bit of a precarious situation this can leave a person in! If you NEED a lot of memory, no matter how you slice it, getting Samsung B-die equipped memory is going to leave you with the best chances of running them at the highest speeds. Whether you opt to get 2x16GB or 4x8GB, the results seem to be hit and miss for either. If you DO get 8GB sticks, even then you'll want to make sure that they are "1-Rank" modules, and determining that is not very easy until after you get them, so research on forums and looking at CPU-Z screenshots of the "SPD" tab will help.

Third, comes down to telling us what model of Motherboard you were thinking of getting, as that as well will pay a role in all of this. Lastly, that if you're using a kit with only 2 sticks, to place them in the memory slots so it looks like this, where o is "open" and | is the stick of RAM: x-|-x-|

While I haven't looked around much... Here on [H], of those I know for certain, we have Topweasel who has an ASRock X370 Taichi and is running 32GB using the G.Skill Ripjaws 5 F4-3200C15D-32GVK 2x16GB 15-15-15-35 kit. His speeds and timings I'm pretty sure are in the Ryzen Memory Speeds thread, but has gotten over 3000MHz.
I had also seen one person on the XtremeSystems forum accomplish 32GB using an ASRock X370 Taichi and the G.Skill TridentZ F4-3300C16Q-32GTZ 4x8GB 16-16-16-26 CR2 kit, but at DDR4-2666 14-14-14-34.

The best "guarantee" (which even still may not result in 100% success) would be getting either of G.Skill's Ryzen intended kits of Fortis or Flare X. The Flare X is the better choice of them, but due to being 'specially for Ryzen' means it's price will be higher than an identical kit of TridentZ.

Ryzen runs with a Command Rate of 1, which sometimes will show up as 1N at the end of the timings (like in the above links, but they are both rated at 2N) or listed elsewhere as CR1. So far the Samsung B-die seem to be handling that just fine, and I suspect that's probably why a lot of other kits with SK-Hynix chips aren't able to run very fast.

AND WHATEVER YOU DO, STAY AWAY FROM ANY MEMORY WITH RGB LIGHTING!

To finally answer your question on that specific kit... the only Corsair kit I've come across so far that has B-Die (I've not been looking though), is the 3600 that has 16-18-18 timings. BUT most B-die have timings like that one does of 16-16-16 (all the same number), so there's a good chance they have them. Again, research will have to be done... :\

Thanks for the detailed reply, that was actually super useful... I ordered the Gigabyte Gaming 5 mobo and I'm planning to put in a Ryzen 1700X (found it for a close enough price to the 1700 so went with that)... I've been checking out Gigabyte's QVL list as with everything i've been reading it seems like this is the best way to go for compatibility. Now that you mentioned to look for the 1-Rank modules is actually very useful, thx - the QVL for that board does list the number of ranks as well as which are Samsung based. I was mulling getting the G Skill Flare Xs, but thought maybe going to a higher 3600 clock might be better (both are on the qvl - http://www.gigabyte.us/Motherboard/GA-AX370-Gaming-5-rev-10#support-doc). The 3600 corsair is a little cheaper, but would you recommend the Flare X's as they seem to be specially made for Ryzen? And both of those do not have LEDs

 

[CAD4466HK]

I ordered the Gigabyte Gaming 5 with Ryzen 1700x, as it seemed a popular option and has everything i'd need (http://www.gigabyte.us/Motherboard/GA-AX370-Gaming-5-rev-10#kf). According to the gigabyte QVL, it says that corsair vengeance lpx would be supported, but just not sure if i should go for a higher clock 3600 lpx like that or stick with the 3200 speed like you mentioned. The G-skill Flare-X is also on the qvl, but do you think there would be a benefit to go with the higher clock 3600 over that memory's 3200? Both are price similarly, in fact the 3600 corsairs are a little cheaper. As far as i can tell also, it seems that its better to keep the 2 DIMMs to 8gb and not 16gb, correct ?

Well the Flare-X is still new and I have no data on them as of yet, getting 3600MHz could be a crapshoot now, but in a few months people could be kicking themselves in the ass for not getting higher speeds because of newer BIOS goodies coming to the table.
If it was me, I would get as much RAM as I needed in the least amount of DIMMs as possible, that way you can upgrade to more later.
See a lot of users going the 2x8 route, not as many going 2x16 though.

 

[gigaxtreme1]

Single rank seems to be the ticket for higher speeds now.

 

[lolfail9001]

Single rank seems to be the ticket for higher speeds now.

Single rank with Samsung B-die on a board with BCLK generator, if we are to lay out all conditions.

 

[cybereality]

The Flare X seem like a surefire bet. Says made for Ryzen and it's on at least MSI's QVL list. 3200 seems to be the max (without overclocking) so I'll probably just stick with that unless the situation drastically changes.

Only thing that could be better is I had to drop to 16GB from the 32GB I'm at now. 16 is still "enough" but I'd have preferred the full 32GB if given a choice (at desired speed).

 

[Formula.350]

Instead of just quoting everyone, and so he still gets an Alert... name drop! Raptorium

The higher speed question is always the biggest one you have to face with memory. Lots of stuff to weigh out...
For instance, how much about computing do you know? Have you ever overclocked? Sure we all start somewhere, so I'm not discouraging it at all, simply that Ryzen is being a fickle mistress and thus you need to not take it as a "certainty" at all.
However, if everything does play nice right from the start like the luck I had with my system, then on the plus side you'll at least be able to tighten the timings at a lower memory speed (aka lower that 16-16-16 to likely 14-14-14), which also will help performance in some areas.
Yet it does come back to IF everything plays nice... If you opt for the 4x8GB or the 2x16GB, there's still the chance of not even getting anything above 2133MHz (though I think that's unlikely if using B-Die it seems). Ryzen IS capable of pretty good memory bandwidth even at 2133. In my instance I was getting 33GB/s Read/Write at 2133 15-15-15-35 and 30GB/s Copy, but 3200 same timings nets me 47.7GB/s Read, 46.5GB/ Write, and 41.8GB/s Copy.

I don't think there are any 16GB single sticks that are currently Single Rank, but I may be wrong, the 64GB Kit of Flare X may be *shrug* Either way. Only YOU know how much RAM you need for your usage. If you can get away with 16GB, I think a 2x8GB kit is a smart, safe buy right now. The likelihood of being able to utilize it above 3200MHz will also go up in that instance, either through raising the BCLK or if higher options are added by AMD later which are then added to the BIOS natively (as I believe they've said Ryzen is capable, there are just Microcode issues to iron out before they feel confident in allowing for them). It'd definitely save you some money by doing 2x8 right now.