Computex 2019 - AMD teased new Ryzens 3. gen (launch 7.7.2019)

[FlanK3r]

SUMMARY NOTES
AMD CPU launch will be 7.7. 2019


New chipset name is AMD X570 withPCIe 4.0. (on the photo is Crosshair VIII motherboard)


IPC improvement is high15%! It was supposed to be originally 10%. Doubled FP and also doubled size ofcache.


CPUs can be up to16C/32T in theory, AMD showedCPU with two chiplets.At launch day minimal 3 models are ready.

Ryzen 7 3700X, CPU with high performance and clocks are3.6 GHz baseclock and turbo up to4.4 GHz andTDP only65W (it is 8C/16T CPU)!

-much better performance than current flagship Ryzen 7 2700X (propably close to 9900K in performance)

Ryzen 7 3800X is again eightcore withSMT, baseclock 3.9 GHz and turbo up to4.5 GHz, TDP 105W

Surprise is the 3800X defeati9-9900K not only in multithreadCinebenchR20 benchmark, but also in single thread. Andi9-9900K has much higher clocks for single thread (5 GHz).

Finally, newRyzen 9 is here! Information about the incoming flagship Ryzen 9 3900Xwith12C/24T. Default clocks is3.8 GHz and turbo up to 4.6 GHz. TDP is still solid105W!

performance demo test against Intel HEDT i9-9920X ...

Price list.

[madcho]

sounds great.

Is TPM possible to disable ?

[SADS]

i'll be interested to see how these compete single thread. I'm big on my games, so interested to know if they will over take intel now in those applications.

[vario]

i'll be interested to see how these compete single thread. I'm big on my games, so interested to know if they will over take intel now in those applications.

Well, amd showed some game numbers vs 2700x.
So , i took those percentage gains, and applied them to 2700X scores in the same games vs intel counterparts.
It seems ryzen 3000 will get around 8600K/8700K performance. So, good.But still worse than the 9700/9900K.
Which was not that surprising, if they would stomp intel in games, they would show that .They lack the gigahurtzes for that
Still, im getting back into the warm and fuzzy amd fold this year with a 3900X or a new threadripper if they release it.Its gonna be a solid upgrade from 5960x.

[zanzabar]

sounds great.

Is TPM possible to disable ?

Do you mean secure boot? No way they make TPM mandatory or enable by default.

[madcho]

no hack on this yet ?

[AliG]

Well, amd showed some game numbers vs 2700x.
So , i took those percentage gains, and applied them to 2700X scores in the same games vs intel counterparts.
It seems ryzen 3000 will get around 8600K/8700K performance. So, good.But still worse than the 9700/9900K.
Which was not that surprising, if they would stomp intel in games, they would show that .They lack the gigahurtzes for that
Still, im getting back into the warm and fuzzy amd fold this year with a 3900X or a new threadripper if they release it.Its gonna be a solid upgrade from 5960x.

There's a massive amount of L3 cache on these chips. I would wait to see how the overclocking tests look - there's a good chance the gaming performance will be very competitive.

[vario]

There's a massive amount of L3 cache on these chips. I would wait to see how the overclocking tests look - there's a good chance the gaming performance will be very competitive.

You have to remember, they dont have 32MB L3, they have 2x16MB L3, and thats not the same, CCX`s are still 4 core ones, whenever a thread migrates from a core on CCX 1 to a core on CCX 2 l3 has to be filled again.
But yes, they are 2x bigger.But you have to understand that they are not having better latency, and thats why intels are fast(well that and clocks)

[AliG]

You have to remember, they dont have 32MB L3, they have 2x16MB L3, and thats not the same, CCX`s are still 4 core ones, whenever a thread migrates from a core on CCX 1 to a core on CCX 2 l3 has to be filled again.
But yes, they are 2x bigger.But you have to understand that they are not having better latency, and thats why intels are fast(well that and clocks)

Gotta remember games are 4 threads at most usually; I'll take a 30% improvement in cache in those cases (as compared to 12 Mb on 9700k). Yes, latency is slower than Intel. But in general the cost of cache misses is much higher than the read time. As long as the cache management is good I'll bet a notable portion of the IPC improvement came from double the size.

[zanzabar]

They are adding support for 4000mhz ram without overclocking the memory controller. That should really help with the latency.

They also may have a different cache structure it looks like it has cache on the IO controller as well, and that there are different io chip configurations for 1 or 2 cpu die configurations. They should have much more info closer to launch.

[vario]

They are adding support for 4000mhz ram without overclocking the memory controller. That should really help with the latency.

They also may have a different cache structure it looks like it has cache on the IO controller as well, and that there are different io chip configurations for 1 or 2 cpu die configurations. They should have much more info closer to launch.

I didnt see any talk about that, could you link to the rumours of 2 different IO dies on ryzen platform ? (epyc will have its own and that we know).
As for the IO size, if you subtract cores from zeppelin die, you are left with pretty much exactly the die size that IO has, so no space really for L4 cache.I jope im wrong tho

[kromosto]

What i saw from AMD presantation is single cıre zen2 is still weak. I dont have any numbers but comparing a 8c16t with a 8c8t in cinebench and showing you win with a margin that is not enough for 2x threads states this.

[vario]

What i saw from AMD presantation is single cıre zen2 is still weak. I dont have any numbers but comparing a 8c16t with a 8c8t in cinebench and showing you win with a margin that is not enough for 2x threads states this.

They were comparing 9700K to the 3700X tho and not 3800X.
I think they only lack like 300mhz , its gonna get very close.

[kromosto]

Their boost delta is only 100mhz. I dont know about their all core boost but if amd intentionally didnt limited 3700x they should be similar. And imo if it was remarkle then AMD will emphasize performance by comparing 9900k with 3800x. Instead they choose to emphasize price/performance by comparing 3700x with 9700k.

[vario]

Their boost delta is only 100mhz. I dont know about their all core boost but if amd intentionally didnt limited 3700x they should be similar. And imo if it was remarkle then AMD will emphasize performance by comparing 9900k with 3800x. Instead they choose to emphasize price/performance by comparing 3700x with 9700k.

But they were comparing them in cinebench 20, all core boost what is important there, 3800X is a 105W part and 3700X is a 65W part, that HAS to hamper all core boost, with that comparison they just looked way better, they have higher performance at much lower power consumption AND price.
9900K is gonna get compared to 3900X because again, they will be able to vastly outperform it (in cinebench) so thats the comparison they want to put out.
As for 9900K vs 3800X, i dont know man, it looks to me like they are gonna be neck and neck in cinebench, but thats just a lot less "flashy" so its better to show off with 3900X.
9900KS is gonna be faster (than 3800X), but at 5ghz all core, its gonna eat like 200W.

[zanzabar]

I think that KF is going to be over 200W for 5ghz all core on any reasonable load. You also have to consider how turbo works on both platforms. XFR does not respect any TDP so long as the board/cooler are good enough, and in spec turbo is almost useless. The KF is still a 95W part on ARK, so other than the warranty on 5ghz all core you are not really getting anything without overclocking. Anyone doing a proper bench without overclocking and in official spec then would have a hamstrung KF doing about 105W, and the 3800x on a good board with a good cooler will be rocking 250-300W in spec if it needs it, but on a garbage board or cooler will be doing the same kind of much lower clocks and ~100W that the KF is doing. The 3800x is shipping a 160W wraith prism cooler, so it would never really work out to compare them with both being "in spec" and "not overclocked."

I didnt see any talk about that, could you link to the rumours of 2 different IO dies on ryzen platform ? (epyc will have its own and that we know).
As for the IO size, if you subtract cores from zeppelin die, you are left with pretty much exactly the die size that IO has, so no space really for L4 cache.I jope im wrong tho

I dont have a source for that, but I did talk a rep about epyc last year and they were talking about next gen having some cache on a routing chip to help with cross die latency. It seems like this would be the application if it is going to be a thing. I want to go through the white papers when they come out with the launch.

[vario]

I think that KF is going to be over 200W for 5ghz all core on any reasonable load. You also have to consider how turbo works on both platforms. XFR does not respect any TDP so long as the board/cooler are good enough, and in spec turbo is almost useless. The KF is still a 95W part on ARK, so other than the warranty on 5ghz all core you are not really getting anything without overclocking. Anyone doing a proper bench without overclocking and in official spec then would have a hamstrung KF doing about 105W, and the 3800x on a good board with a good cooler will be rocking 250-300W in spec if it needs it, but on a garbage board or cooler will be doing the same kind of much lower clocks and ~100W that the KF is doing. The 3800x is shipping a 160W wraith prism cooler, so it would never really work out to compare them with both being "in spec" and "not overclocked."



I dont have a source for that, but I did talk a rep about epyc last year and they were talking about next gen having some cache on a routing chip to help with cross die latency. It seems like this would be the application if it is going to be a thing. I want to go through the white papers when they come out with the launch.

I dont know man, i dont think they were comparing them overclocked, remember this showcase few months back when they were showing 8 core zen 2(no model number then) vs 9900K in cinebench ?
They had the same scores, but power usage on zen 2 and 9900K was in spec , and lower on zen2 than 9900K.
There are limits to auto oc using xfr/pbo and on the intel sie mce also there are turbo durations that SHOULD be respected.
TDP on intels site is coolers capacity more than real power consumption tho.

As for the IO die, yes, the one for EPYC is diffrent, its so big it can have cache on it.But it would simply not fit on am4 package. From what ive seen, for Am4 there is one IO die, the same for 1 and 2 chiplet cpus.

[zanzabar]

As for the IO die, yes, the one for EPYC is diffrent, its so big it can have cache on it.But it would simply not fit on am4 package. From what ive seen, for Am4 there is one IO die, the same for 1 and 2 chiplet cpus.

It sounded like epyc and ryzen 2 would the same but double the parts (4 cpu die, 2 IO die, with them being the same or very similar.) That is why I want those white papers. It might turn out that the cores have the same cache as before, but the IO chip doubles it for each cpu die connected.

If it was like that then I dont see how you get the 64/128 lane PCI-e.

[vario]

It sounded like epyc and ryzen 2 would the same but double the parts (4 cpu die, 2 IO die, with them being the same or very similar.) That is why I want those white papers. It might turn out that the cores have the same cache as before, but the IO chip doubles it for each cpu die connected.

If it was like that then I dont see how you get the 64/128 lane PCI-e.

Im not sure i understood you correctly.So im just gonna post pics .
Thats Epyc 2, it has 8 chiplets and one huge IO die:
https://images.anandtech.com/doci/13...78_678x452.png

Thats Ryzen 3000 8 Core one:
https://cdn.wccftech.com/wp-content/...en-3rd-Gen.jpg

And thats Ryzen 3000 12 core one base on 2x eight core chiplets.
https://images.anandtech.com/doci/14...ar_678x452.jpg

The dimensions of RYZEN`s IO die, are pretty much very similar what you would get if you took zen zeppelin die and removed cores+cache, as it is still based at 14nm, no shrinkage.So i really doubt that they managed to cram more cache in there.
But as you see, the epyc IO die is massive in comparison, so i dont know .

[zanzabar]

That makes more sense.

[madcho]

They use the specialised process for the hub, I guess.

They want improve yield, and increase performance of the cores by using specialised process for the chiplet. Great move.

[AliG]

https://www.guru3d.com/news-story/ry...-overview.html

Motherboard compatibility in case anyone was interesting. Nice that AMD isn't forcing you to change boards every single CPU lol