AMD launches Zen+ 12nm Ryzen and X470 motherboards

[The Stilt]

Holy Hell! , different between 3200 & 3200 LL Is huge.

Exactly.
I'd personally choose 3200MHz with the tightest possible timings any day over the standard CL16+ 3600MHz.

 

[Gideon]

Ryzen 1K-series officially supports up to 2666MHz (1 DPC SR) and 2K-series is supposed to officially support 2933MHz.

Yes, I know that. I meant that Overclocked memory-kit makers have tested and validated memory kits for Pinnacle Ridge with 3400Mhz (@1.35v). Something that hasn't been done with Summit Ridge. I would much rather see a PR memory kit running @ 3600 MHz though. Confirming that PR can actually be stable on clocks higher than 3466 MHz.

 

[raghu78]

Nope.
And nope.

High speed memory won't automatically make Ryzen better in gaming.
The improvements from higher than 3200MHz MEMCLK are already diminishing.
At or above 3200MHz lower latency matters much more.

We are seeing lower memory latency from PR . We have to wait and see how well PR can run DDR4 4000 Mhz. It will be interesting to see the performance effects of various combinations of memory speeds and latency with PR.

https://videocardz.com/75504/cpc-hardware-amd-ryzen-7-2700x-review-goes-online

 

[wahdangun]

Pardon my ignorance, but where can one find that listing? The only place mentioning new geil memory and Ryzen 2 seems to be this article. While official 3400 MHz support is a nice improvement over Ryzen 1xxx (3200 MHz), where do you get so low latencies from?

In Tech power up, www.techpowerup.com/242797/

And the sniper x, will launch with 3200 and 3600. So maybe with some golden sample PR IMC can handle 4000 MT/s

Actually i7 8700K will have lower latency ~ >50ns with extreme timings and speeds ~>40 even down to 35ns. If AMD would give clock on IF (DATA fabric that connects Memory to CCXs) and we could clock it up to 3-3,5GHz then Ryzen 2000 would be clear winner in every game. There is no way that NB can ever match efficiency of IF.


i7 8700K has AUTO clock depends on memory to 4,3GHz on NB. Base clock is 3,2-3,7GHz for NB,

if INFINITY FABRIC (DATA FABRIC) would clock to 4GHz we actually could see some insane numbers below 35ns for memory latency.
http://www.overclock.net/forum/225-...icial-reviews-benchmarks-update-19-a-595.html

yeah,.. <35ns, but this is 5155MHz on NB.

But even with this insane number and clock it still cannot reach 60GB/s.

I think ZEN+ core is already ahead of INTEL for gaming, but AMD doesn't have the process and Ryzen architecture makes it inefficient for that kind of work.

Users, that can try do as I did, put 4-3,8GHz on i7 8700K and put NB to 1,6GHz and you will be amazed by tech from AMD.

I know that AMD is pushing hard, but they need to get business in servers... then they can focus on gaming GPU/CPU only. I bet 7nm from GF/IBM could give AMD a real chance.

But with intel ring bus topology the latency will not uniform and actually RR latency between core in the same ccx was lower than intel. And AMD with it's design can't decoupled IF frequency with it's IMC because it will introduce more latency.

 

[Bouowmx]

L3$ runs at same as core.

I just went through telling you that on Intel Haswell and later, they don't have to.

https://www.anandtech.com/show/6355/intels-haswell-architecture/10

The three clock domains in Haswell are roughly the same as what they were in Nehalem, they just all happen to be on the same die. The CPU cores all run at the same frequency, the on-die GPU runs at a separate frequency and now the L3 + ring bus are in their own independent frequency domain.

See also: https://en.wikichip.org/wiki/File:skylake_soc_clock_domain_block_diagram.svg

 

[wahdangun]

Nope.
And nope.

High speed memory won't automatically make Ryzen better in gaming.
The improvements from higher than 3200MHz MEMCLK are already diminishing.
At or above 3200MHz lower latency matters much more.

But wasn't RR scale with clock ? So the gain from using high speed ram will be exponentially more pronounced when RR clocked high enough like 4 GHz or more ?

 

[Markfw]

Holy Hell! , different between 3200 & 3200 LL Is huge.

Yes, but what is 3200ll and 3600ll ? I have 3200 cas 14, is this even lower ?

 

[PhonakV30]

Yes, but what is 3200ll and 3600ll ? I have 3200 cas 14, is this even lower ?

I think LL means low latency so Yes 3200 LL means 3200 Cl 14 , also CL 14 Is lowest latency among all memory kits

 

[CatMerc]

I think LL means low latency so Yes 3200 LL means 3200 Cl 14 , also CL 14 Is lowest latency among all memory kits

I remember the thread this pic came from, the LL is for 3200 CL12 or something to that tune.

It's highly tuned, not something every Ryzen can get and requires a very good set of B-Die.

Actually i7 8700K will have lower latency ~ >50ns with extreme timings and speeds ~>40 even down to 35ns. If AMD would give clock on IF (DATA fabric that connects Memory to CCXs) and we could clock it up to 3-3,5GHz then Ryzen 2000 would be clear winner in every game. There is no way that NB can ever match efficiency of IF.


i7 8700K has AUTO clock depends on memory to 4,3GHz on NB. Base clock is 3,2-3,7GHz for NB,

if INFINITY FABRIC (DATA FABRIC) would clock to 4GHz we actually could see some insane numbers below 35ns for memory latency.
http://www.overclock.net/forum/225-...icial-reviews-benchmarks-update-19-a-595.html

yeah,.. <35ns, but this is 5155MHz on NB.

But even with this insane number and clock it still cannot reach 60GB/s.

I think ZEN+ core is already ahead of INTEL for gaming, but AMD doesn't have the process and Ryzen architecture makes it inefficient for that kind of work.

Users, that can try do as I did, put 4-3,8GHz on i7 8700K and put NB to 1,6GHz and you will be amazed by tech from AMD.

I know that AMD is pushing hard, but they need to get business in servers... then they can focus on gaming GPU/CPU only. I bet 7nm from GF/IBM could give AMD a real chance.

The very act of decoupling the IF from DRAM would introduce latency that will need to be compensated for by pushing up clocks. Crossing clock domains is never cheap.

So no, matching Intel's ringbus and AMD's IF frequency isn't a fair IPC comparison either. They're built for completely different goals.

 

[The Stilt]

I remember the thread this pic came from, the LL is for 3200 CL12 or something to that tune.

It's highly tuned, not something every Ryzen can get and requires a very good set of B-Die.

3200LL was CL12, 3466LL CL14.
Both had the subtimings (which make the biggest difference tuned to the max).

3200MHz and everything north to that require B-die anyway.

Regardless 3200MHz was significantly easier to achieve than 3466MHz due to the frequency itself.
Not all Ryzens can do 3466MHz.

 

[.vodka]

Your timing sets make a world of difference and I'd like to thank you again for making these available to every Ryzen user. Maybe not everyone is able or would like to push things as far as 3200C12 and 3466C14, but your 3200C14/3466C15 sets with proper subtimings rock!

 

[IRobot23]

I just went through telling you that on Intel Haswell and later, they don't have to.

https://www.anandtech.com/show/6355/intels-haswell-architecture/10



See also: https://en.wikichip.org/wiki/File:skylake_soc_clock_domain_block_diagram.svg

Yes you are right.
Then correct testing should be (for me) : Coreclock 1,6GHz, taget cache 1,6GHz (intel) + ryzen ddr4 3200MHz (+intel same).

 

[Markfw]

3200LL was CL12, 3466LL CL14.
Both had the subtimings (which make the biggest difference tuned to the max).

3200MHz and everything north to that require B-die anyway.

Regardless 3200MHz was significantly easier to achieve than 3466MHz due to the frequency itself.
Not all Ryzens can do 3466MHz.

And for samsung b-die 3200 cas 14, what are the "tuned subtimings" ? All I know to set is 14,14,14,34

 

[IRobot23]

3200LL was CL12, 3466LL CL14.
Both had the subtimings (which make the biggest difference tuned to the max).

3200MHz and everything north to that require B-die anyway.

Regardless 3200MHz was significantly easier to achieve than 3466MHz due to the frequency itself.
Not all Ryzens can do 3466MHz.

I was unable to run your timings on my dims (samsung E die). Even safe 3466MHz wasn't stable at 3200MHz. Maybe I should give another try, but looks like samsung B die is just incomparable.

 

[IEC]

I was unable to run your timings on my dims (samsung E die). Even safe 3466MHz wasn't stable at 3200MHz. Maybe I should give another try, but looks like samsung B die is just incomparable.

Samsung E die isn't capable of the low latency timings of B die. It's why B die commands a large premium (especially post Ryzen)

 

[The Stilt]

And for samsung b-die 3200 cas 14, what are the "tuned subtimings" ? All I know to set is 14,14,14,34

These will do on any decent B-die:

If you're on TR then leave tRDWR and tWRRD to auto.

 

[CatMerc]

Yes you are right.
Then correct testing should be (for me) : Coreclock 1,6GHz, taget cache 1,6GHz (intel) + ryzen ddr4 3200MHz (+intel same).

Again, crossing clock domains is costly, and therefore any design that has the flexibility to decouple the fabric from DRAM speed will have an innate disadvantage per clock. Ryzen's fabric is locked to DRAM frequency to avoid that clock domain crossing penalty.

This isn't really a fair test.

 

[IRobot23]

these are mine

 

[IRobot23]

Again, crossing clock domains is costly, and therefore any design that has the flexibility to decouple the fabric from DRAM speed will have an innate disadvantage per clock. Ryzen's fabric is locked to DRAM frequency to avoid that clock domain crossing penalty.

This isn't really a fair test.

You are right, since one is design reach higher clock and eliminate latency. My intention was to see what happens when both have same DRAM latency at same core clock.

I don't think that Ryzen 2000 will run 4000MHz ddr4, if 3600CL16 is like 3200CL14 on Ryzen 1000 it should be great.

 

[Gideon]

Can confirm that Stilt's timings seem to work better compared to 3466 @ CL16. I had similar timings set up before and reverted to 3466, as AIDA64 showed a slight regression in bandwidth and the same latency. I figured that the higher Infinity Fabric speed offset the worse timings, but it seems i was wrong.

When testing more realistic usecases like Geekbench or games though, it seems to perform better with 3200 and tight timings (as the geekbench comparison result shows).

OT, but I still don't understand though, why my SGEMM result is consistently half of other Ryzens. I haven't really touched anything other than the memory clocks and timings (and the low score persists even with JEDEC timings & stock clocks). An earlier stepping?

 

[Markfw]

These will do on any decent B-die:

If you're on TR then leave tRDWR and tWRRD to auto.

Well, here are mine using the standard XMP profile.

So which ones are really out of whack ? and can I change all these in BIOS ?

 

[DisEnchantment]

About the famous Zen@5G that was hidden at the time and now used by fanboys to shitpost, It was not part of the preview. BTW, I probably know what happened. A feature planned for SMR seems to have been skipped later and is now included in PNR. Guess which one?

https://twitter.com/d0cTB/status/979433595528974346

Which one?

Don’t expect a very high boost due to “XFR2 Enhanced” in standard configuration. The overall Turbo boost is already quite high.

https://twitter.com/d0cTB/status/979429869812580353

We tried and benchmarked PNR on X370, X470 and A320 boards. We finally keep the A320 scores because it uses a retail BIOS with public AGESA for PNR. X470 BIOS was full of various “cheats”. A320 BIOS was conform with AMD specs. Both results were close.

https://twitter.com/d0cTB/status/979429170122981377

Right now, that what I see on games. Of course, with auto-overclocking or other cheats enabled (all-cores Turbo, etc.), you can have a 5-8% gain. They were enabled by default on our X470 boars...

https://twitter.com/d0cTB/status/979442416116817920

Gains over A320 results... stock 2700X

 

[CatMerc]

These will do on any decent B-die:

If you're on TR then leave tRDWR and tWRRD to auto.

For some reason my kit and motherboard refuse to POST with geardownmode disabled, no matter how loose the timings on my 3200.

Got a G.Skill TridentZ 3200 CL14 RGB 2x8, on a Gigabyte AORUS AX370 Gaming K7, and a 1700.

 

[The Stilt]

Well, here are mine using the standard XMP profile.

So which ones are really out of whack ? and can I change all these in BIOS ?

Set tRDRDSCL, tWRWRSCL to 2 CLKs and tRFC to 308 CLKs.
tWR you can set to 12.

Those will give you the biggest boost, which should be quite significant compared to those timings you posted.

These can all be found under the standard memory timings, each and every board has them.

For some reason my kit and motherboard refuse to POST with geardownmode disabled, no matter how loose the timings on my 3200.

Got a G.Skill TridentZ 3200 CL14 RGB 2x8, on a Gigabyte AORUS AX370 Gaming K7, and a 1700.

Try disabling GearDownMode first while leaving Cmd2t to "Auto".
Once the system has posted re-enter the bios and set Cmd2t to 1T.

I suggest that you use >= 1.05V VDDCR_SoC as these settings will put more strain on the memory controller, than the standard configuration.

 

[dnavas]

X470 BIOS was full of various “cheats”.
https://twitter.com/d0cTB/status/979429170122981377

Ah, so all-core turbo and memory overclocking are considered cheats? Interesting. Well, ignoring all-core turbo numbers would probably result in incomplete understanding of the node. But is it just all-core, because it's the single core boost getting pinned at the relatively low 4.3 mark which is most surprising.
I also don't understand the big mystery behind "5 Ghz". The chip "exaggerates" temperature and doesn't always report stats (when powered down), so is it hard to guess it might have mis-reported clocks on the high end at some point?