Speculation: Ryzen 3000 series

[Yotsugi]

while there is no reason to be that close

You're no packaging engineer to say that.
Neither am I, and we both gotta wait for HC'19 for maybe proper Rome disclosure.

 

[Kedas]

To be possible to have direct access between chiplets there has to be cache-coherency aware northbridge in chiplets.....

I don't think we call core to core communication part of the "northbridge"

 

[Kedas]

You're no packaging engineer to say that.
Neither am I, and we both gotta wait for HC'19 for maybe proper Rome disclosure.

Are you sure? and if you read further you will read that I leave the option open for parameters I don't know

 

[Yotsugi]

I don't think we call core to core communication part of the "northbridge"

What's directory and snoop filters 101?
Cache coherency is some hard stuff.

Are you sure?

Well yes, I am.
We don't know a thing about reasons of Rome being packaged the way it is.

 

[naukkis]

I don't think we call core to core communication part of the "northbridge"

Even with Zen1 inter CCX communication is done through northbridge. And probably in Zen2 even chiplets inter CCX accesses goes through IO chip.

As AMD's last level cache isn't inclusive snooping into other cpu cache is costly and isn't done but instead northbridge snoop filtering is used and request to other cpu is done only when hit is confirmed.

 

[Gideon]

about core die connections, there is also the fact that the EPYC core dies are very close together, while there is no reason to be that close

There might be plenty of other reasons. Similarity to the Zen-1 layout being one (pins going to the same place on package, etc). One can't just place chiplets randomly on the package, without there being considerable added complexity/cost - especially when you must be socket-compatible with the older chip.

Also - pairwise die-connections (on EPYC at least) are a ludicrously useless idea. It really gives you nothing except added complexity and wasted resources. Even on Matisse it quite pointless. And when considering that lion's share of the chiplets are going to Rome, wasting considerable die-space for that nonsense would be downright criminal

 

[Gideon]

By the way, There was another Sisoft leak of a 4-core mattisse sample
It's has been taken down now , but luckily I took a screenshot yesterday (cropped unfortunately).

Particularily interesting were the "Processor Arithmetic" and "Processor Multimedia" (AVX2) workloads.

These are of the taken down scores of a Matisse (Zen 2) 4-core engineering sample running at 3.8 GHz
(Arithmetic above, Multimedia below):

Here are the average scores of a stock I7 6700K for reference, along with the listed clock-speeds (400 MHz faster than Zen 2):

These are the average scores of a stock R5 1500X

Now obviously take the following with mountains of salt, as the sample size for comparison is a single engineering sample vs the average of other stock processors, but

Compared to Skylake (6700K) the Matisse sample had a 10% clock speed deficit. Normalising the clock speed the Matisse sample had:

• ~5% higher IPC in the arithmetic test
• ~12,5% higher IPC in the vectorised multimedia test
• To be fair, there are other stock 6700K results where the difference is smaller (3% and 6%), but Matisse is still is a bit faster per-clock

Key takeaways:

• The AVX2 performance has indeed doubled compared to zen 1, rivalling that of a 2700X on a lower clocked 4-core Engineering Sample
• IPC, in this small subset of tests is neck-to-neck to Skylake, outperforming it very slightly

In the end it seems, it will be all about the clockspeeds
Overall AMD should be very competitive in professional workloads even with a 100-200 Mhz clock speed deficit

 

[coercitiv]

In the end it seems, it will be all about the clockspeeds

Almost all about the clockspeeds: they will still have the added advantage of lower power and core count flexibility, which may enable them to pull ahead in throughput. ST or light threaded perf will indeed count massively on max clocks.

Anyways, nice catch and ty for the short and sweet comparison.

 

[beginner99]

There might be plenty of other reasons. Similarity to the Zen-1 layout being one (pins going to the same place on package, etc). One can't just place chiplets randomly on the package, without there being considerable added complexity/cost - especially when you must be socket-compatible with the older chip.

Good point. And not just socket compatible but drop-in compatible meaning the cooler must also fit and don't have some uneven pressure. And mabye packaging lines themselves will need little adaption only.

In the end it seems, it will be all about the clockspeeds

Which IMHO intel will win. I simply don't see TSMC process reaching >5 Ghz. regardless of power consumption. I can imagine single-core turbo close to 5 ghz and especially much lower power use than 9900k but I expect 9900k will remain single-threaded king.

 

[coercitiv]

I expect 9900k will remain single-threaded king.

Intel needs more than that though, otherwise their desktop supremacy will literally be hanging on by a thread.

They still have the gaming advantage, and Zen 2 may yet have an unforeseen Achilles' heel, but my guess is they'll get uncomfortably close on all fronts.

 

[Zucker2k]

Good point. And not just socket compatible but drop-in compatible meaning the cooler must also fit and don't have some uneven pressure. And mabye packaging lines themselves will need little adaption only.



Which IMHO intel will win. I simply don't see TSMC process reaching >5 Ghz. regardless of power consumption. I can imagine single-core turbo close to 5 ghz and especially much lower power use than 9900k but I expect 9900k will remain single-threaded king.

A clock speed disadvantage could be offset by significant ipc advantage. My pessimism for why Zen2 won't sustain a 5GHz fmax with 8 cores/16 threads, for example, is heat. The denser these processes get the more prone the cores would become to heatspotting as cooling efficiency diminishes.

 

[Markfw]

A clock speed disadvantage could be offset by significant ipc advantage. My pessimism for why Zen2 won't sustain a 5GHz fmax with 8 cores/16 threads, for example, is heat. The denser these processes get the more prone the cores would become to heatspotting as cooling efficiency diminishes.

Denser yes, and I am not saying I think it will sustain 5 ghz either, but you forget that its on 7nm.

I think we need to wait and see on the 5 ghz thing.

 

[Thunder 57]

...I think we need to wait and see on the 5 ghz thing.

As far as 5GHz goes, I don't expect to see that on TSMC 7nm or Intel's 10nm. Remember that 14nm is so refined that I expect a clockspeed regression. Intel may of course make up for it with IPC improvements.

 

[Gideon]

As far as 5GHz goes, I don't expect to see that on TSMC 7nm or Intel's 10nm. Remember that 14nm is so refined that I expect a clockspeed regression. Intel may of course make up for it with IPC improvements.

Well, judging by the numerous Geekbench leaks and what Intel has presented about the architecture, it's pretty clear that Icelake finally has significant IPC improvements (at the very least more than Haswell -> Skylake), so that should be a given.
As for intel 10nm, IMO the Ultra High Performance Cell Libraries look insane (at a significant density cost, twice that of 14nm to 14nm++).

If they ever get this to work and yield (which is the real million dollar question), I have no trouble believing they can hit 5GHz on that process. I doubt it would do much more though, and even that would be well into 2020, at best.

As for AMD - I would really like to see Zen 2 it at 5 Ghz (as some rumors stated that it can do that, but not on all cores), but it will certainly be hard. For instance, we don't even know for sure if the chiplet desing can be soldered (Epyc samples seen so far surely aren't, there were pictures on Twitter showing paste between chiplets)

 

[Kedas]

Key takeaways:

• The AVX2 performance has indeed doubled compared to zen 1, rivalling that of a 2700X on a lower clocked 4-core Engineering Sample

Nice, if I'm not mistaken this is the first test that is showing this 2X that they were talking about, I was starting to fear that it was only design talk but not much in reality due to other bottlenecks.

 

[beginner99]

Nice, if I'm not mistaken this is the first test that is showing this 2X that they were talking about, I was starting to fear that it was only design talk but not much in reality due to other bottlenecks.

On the otherhand only niche application really use it. Software is way behind here. I mean even common python packages like numpy don't use it for "compatibility" unless you compile it yourself. Good luck with that on windows.

 

[inf64]

Zen2 is a really nice core. Right now it's hard now to gauge all the benefits it brings to market but this core will definitelly be more successful than Zen/Zen+. AMD has overshot (will be seen) the IPC targets again and this core has no security issues Skylake (derivatives) have. It skales great and AMD already has some HPC wins. TSMC process is not perfect but they will get upper 3Ghz on server products and higher than that on worsktation parts. Desktop will likely be significant hit to intel as now AMD will have IPC parity and or advantage while clocking almost as high as top intel 14nm products while having significant perf/watt advantage. SunnyCove core will be a good uplift but I personally believe it will be win/lose Vs Zen2/Zen3 lineup. AMD has big plans for Zen4 generation so next few years will be very interesting with regards to CPU core wars.

 

[Gnyueh]

Some unreliable leak from my country about x570 chipset.
https://www.chiphell.com/forum.php?mod=viewthread&tid=1979902&extra=page=2&mobile=2

1.扩展性大幅度提升 起码会做2条全速的nvme x4 3.0 某些主板可能支持3-4条
2.由于有某个不可说的产品 所以供电加强幅度你懂的
3.外围搭配设备有明显区别 原生支持新的usb40gbps 但是取决于主板厂做不做
只能说asmedia那个大hub太垃圾了 zen的成功让小aa有更多人手开发新的芯片组
ps:550似乎没消息 应该会延迟发布
低端还是a320 只能说pcie4.0大法好！

(google translate)
1. Scalability is greatly improved. At least 2 full-speed nvme x4 3.0 will be used. Some motherboards may support 3-4.

2.Because there is something that cannot be said, the power supply is enhanced.

3. There are obvious differences between the peripheral equipment. The original support of the new usb40gbps but depends on the motherboard factory do not do

Can only say that asmedia that big hub is too junk zen's success makes small aa have more people to develop new chipsets

Ps:550 seems to have no news should be delayed

The low end is still a320. It can only be said that pcie4.0 Dafa is good!

 

[moinmoin]

As far as 5GHz goes, I don't expect to see that on TSMC 7nm or Intel's 10nm. Remember that 14nm is so refined that I expect a clockspeed regression. Intel may of course make up for it with IPC improvements.

Intel's 14nm is much refined, and they still are refining 10nm to make it work. In AMD's case frequency was hard limited by GloFo's licensed node originally being optimized for low power usage. Unless anything changed Zen 2 will use TSMC's high performance 7nm node, so TSMC has an inherent interest in its foundry customers being able to push high frequencies on it from the get-go compared to the power/mobile optimized 7nm variant.

 

[Gideon]

Wow, looncraz has designed a really nice Memory Latency calculator for Zen 2 (based on the 2 tests we had, that had latency numbers):

https://www.reddit.com/r/Amd/comments/b9jlsy

 

[Veradun]

I see rumors on x570 boards coming very soon.

May 1st (AMD 50th anniversary) halo product (paper) launch?

 

[french toast]

I see rumors on x570 boards coming very soon.

May 1st (AMD 50th anniversary) halo product (paper) launch?

I think it would be stupid to miss this opportunity imo.

 

[Veradun]

I think it would be stupid to miss this opportunity imo.

I definitely agree, it's a strong opportunity to "make some noise"

 

[Fanatical Meat]

I see rumors on x570 boards coming very soon.

May 1st (AMD 50th anniversary) halo product (paper) launch?

I definitely agree, it's a strong opportunity to "make some noise"

Been watching X470 boards at Microcenter. Nearly all have extra discounts applied to the and there is an additional $20-30 off select motherboards that rotates thru Microcenter inventory. Tiachi is now out of stock.
Appears that Microcenter has been thinning it’s X470 stock.

4-9 edit: Microcenter is down one more X470 board (13 in stock was 14 models)
B450 is one less in stock too.

 

[fleshconsumed]

Any news on the B550 motherboards? I have a few PCs that do not need more than B450/B550 chipset. I do not want to buy B450 when new generation is supposedly so close around the corner. So, anybody heard of B550 release date?