New Zen microarchitecture details

[Rifter]

Or Bulldozer at 8.4GHz?

yes or BD, or any other cpu arch.

 

[CentroX]

I am a bit surprised there hasn't been more leaks about ryzen at this point. AMD are protecting their stuff well.

 

[cytg111]

..So, what other things make you think that Zen will overclock less?

- Just one thing. One little tiny miniscule thing. Lack of evidence. You provice conjecture, not evidence. Plus you HAVE to consider the source. It may all be true, and then some, but common, dont go making 2, 3, 4 assumptions add them up and pass the sum as fact. Am not buying.

 

[Atari2600]

There are areas of silicon that are of low quality: notably the border of the wafer. They promised >=3.4 base clock in 95W. What if there are some chips that are not defective but leak a lot? They are harvested as quad core and the leaky CCX fused off. But the other CCX is near the bad one, so has the same impurities and probabily does not clock well.

That assumes there won't be (albeit with reduced probability) local flaws toward the centre of the wafer.

So, your lower grade (harvested as opposed to 4C mask) quads come from the wafer edge as you indicate, the top grade come from toward the centre wafer where a rogue flaw has ingrained itself.

If there are 95W 4c then there is hope for high clocked quads, but i heard only of 65W quads...

With Kaby Lake having impressive clocks out of the gate (existing 7700K is 4.2B/4.4ACT/4.5SCT I believe), and if AMD feel the need, then a quick and dirty response from AMD would be to release quads at 95W. Still within AM4 spec, so nothing of the existing infrastructure changes - but it'd mean 4C Zen can have a real go at matching 4C KL clocks.


B - Base
ACT - All Core Turbo
SCT - Single Core Turbo

 

[Atari2600]

CanardPC 4-5 month before launch, on an ES, very possibly not even final silicon... using an average load. Hmm.

Reminds me of that Phenom 3.0GHz a year before launch.

So this has pretty much the same issue as the AMD power figure.

Now that is utter FUD (worthless opinions are own).

The Barcelona 3.0 GHz was at the level of a fudzilla rumour.


We've now got independently verified CPUIDs of 3.6/3.9 8C Zen which are supposed to have a 95W TDP. If a 3.6/3.9 8C can fit under 95W, why is it a surprise that a 3.4 GHz chip (even with mild overvolting to ensure stability) also comes in under 95W?

If the 3.6/3.9 wouldn't have a hope of fitting inside the 95W AM4 TDP, then AMD would not be building them! (If you think otherwise then your really into the tinfoil hat faked moon landings level of conspiracy theorist.)

 

[Doom2pro]

I am a bit surprised there hasn't been more leaks about ryzen at this point. AMD are protecting their stuff well.

It's been quiet lately... Too quiet... The calm before the storm? I just hope official review samples are sent and released BEFORE launch.

 

[CentroX]

It's been quiet lately... Too quiet... The calm before the storm? I just hope official review samples are sent and released BEFORE launch.

Fury X benchmarks was embargoed to the same day it went on sale.

 

[Dresdenboy]

This very interesting for sure. But is there a way to deduce the ST IPC from the numbers if we assume that the front end if not a bottleneck? I can see that the latencies are the same more or less, but I am not sure if the TP is multithreaded or ST.

These lat/tput numbers only give an indication for the core's performance. One might apply a weight based on their typical frequency in real world code, but this wouldn't help much (too many variables at play).

These results are actually obtained by running one thread.

 

[cytg111]

It's been quiet lately... Too quiet... The calm before the storm? I just hope official review samples are sent and released BEFORE launch.

Probably that F4 3.6/4.0 is the final stepping? (ie. no more ES is sent out, no new tricks).
While everyone is going bonkers over plus/minus 2% IPC I am really curious at what clocks the hex and the quad comes in at... We have had zero leaks on those.

 

[Abwx]

what clocks the hex and the quad comes in at... We have had zero leaks on those.

Canard PC said that the 6C/12T are within 65W but they dont specify at wich fequency, likely 3.4GHz if we are to extrapolate from the 8C/16T 3.6GHz frequency.

https://twitter.com/CPCHardware/status/818429683633127424

 

[cytg111]

Canard PC said that the 6C/12T are within 65W but they dont specify at wich fequency, likely 3.4GHz if we are to extrapolate from the 8C/16T 3.6GHz frequency.

https://twitter.com/CPCHardware/status/818429683633127424

I think it would be weird for them not to produce a higher clocking 4c/6c in the same 95W tdp ..

 

[Rifter]

I think it would be weird for them not to produce a higher clocking 4c/6c in the same 95W tdp ..

I agree, i believe that they will release a 95w quad highly clocked to compete directly with KL. Maybe not at launch if they are supply constrained, but within a few months.

 

[Dresdenboy]

With 128bit fp workload is there any difference in efficiency between a 128 or 256 capable fpu? And for 256bit code?

Efficiency for 128b code should be about the same. A 256b FPU could be power gated in part.

256b code might be anything between just as fast as 128b to two times faster (corner cases with dense fp ops).

 

[bjt2]

CanardPC 4-5 month before launch, on an ES, very possibly not even final silicon... using an average load. Hmm.

Reminds me of that Phenom 3.0GHz a year before launch.

So this has pretty much the same issue as the AMD power figure.

But you've still dodged my question. Can you equate your AC deducted power figures to measured DC values, for older CPUs?


You know, I think you might even be right with the approximate ballpark figure but my objection is for way you present it, your deducing process, and the lack of qualifying the factors that could make your figure very inaccurate. So and so is... Because, However. It's the latter part that you leave incomplete.

Example. http://www.guru3d.com/articles-pages/amd-phenom-x4-9850-black-edition-review,10.html

170W idle, 246W load.

Your method;
246-170= 76W AC
76 * .88 * .90 = 60W DC

5-35W idle? Fine.

95W DC worst case.

That's for Phenom 9850BE

And thats MILES out.

Sent from HTC 10
(Opinions are own)

Phenom does not have any sleep state and any clock gate.
It's made in pure dynamic logic, that can't be shut down.
If you look at page 2, there is the CPUID screen.
As you can see, the multiplier and Vcore is fixed.
So in idle is sitting at 2500MHz as under load.
Probabily the power consumption in idle is very high.
170W for MB, NB, SB, GPU all in IDLE seems quite high, even factoring the efficiency.
With Ahtlon 64 was introduced the IDLE at 800 or 1000MHz and reduced Vcore, and there started to draw much less...
Only recently AMD has put enough attention to energy saving.
They are not comparable.
Since only very old reviews agree vith your vison, have you ever had a doubt that maybe that results were not comparable, because are of a very OLD CPU (9 years ago) that don't have any energy saving feature?

 

[bjt2]

You cant compare different archs and assume it will yield the same results. The P4 was probably one of the best OC'ing chips in history with some hitting 8+ghz on suicide runs, something im not even sure has been broken to this day. This doesnt mean that all intel CPU's that came after can obtain the same OC's.

The record is of the FX8370: https://hwbot.org/benchmark/cpu_frequency/halloffame
@The Stilt it's you?

Keller said that Zen has the same DNA of bulldozer. Meaning that is an high frequency design. It has 19 stages integer pipeline, moreless like BD. It reaches 3.6GHz base with less power than the 6900K. +400MHz base with -45W TDP. Why it can't overclock more than 6900K?

 

[bjt2]

- Just one thing. One little tiny miniscule thing. Lack of evidence. You provice conjecture, not evidence. Plus you HAVE to consider the source. It may all be true, and then some, but common, dont go making 2, 3, 4 assumptions add them up and pass the sum as fact. Am not buying.

There is not only AMD, but also CanardPC. 5GHz on air on an early MB and ES, limited only by the VRMs.
Zen has more stages than intel architectures.
Zen already proved to clock higher and with lower power than 6900K. And it's not even finished.
It's physically impossible that the CPU suddenly draw 300W after 4GHz. And I already posted a graph with 14nm frequency scaling.
If a 3.6GHz 95W CPU can't clock over 4.4Ghz, why a 3.2GHz 140W CPU can? Because it have the INTEL magic process?

 

[cytg111]

Keller said that Zen has the same DNA of bulldozer. Meaning that is an high frequency design.

- Really? See thats what I am talking about as conjecture.
Keller said that some of Zen has similarities with some of BD and from that you conclude that he means its a high hz design?
You familiar with erasmus montanus?

https://en.wikipedia.org/wiki/Erasmus_Montanus

A rock cant fly, my mother cant fly, thus my mother must be a rock. (dont tell her that though)

 

[bjt2]

That assumes there won't be (albeit with reduced probability) local flaws toward the centre of the wafer.

So, your lower grade (harvested as opposed to 4C mask) quads come from the wafer edge as you indicate, the top grade come from toward the centre wafer where a rogue flaw has ingrained itself.

I made a probabilistic statement. In the past they didn't thoroughly tested all the chips. Too costly. Probabily they test the max frequency at 4 and 8 cores (and maybe 6), the power consumption in that condition and choose the best assignable SKU. In the old days of athlon x3, many full functioning X4 were underclassified. In part for the high demand, in part to lower testing costs.

With Kaby Lake having impressive clocks out of the gate (existing 7700K is 4.2B/4.4ACT/4.5SCT I believe), and if AMD feel the need, then a quick and dirty response from AMD would be to release quads at 95W. Still within AM4 spec, so nothing of the existing infrastructure changes - but it'd mean 4C Zen can have a real go at matching 4C KL clocks.


B - Base
ACT - All Core Turbo
SCT - Single Core Turbo

I was only saying that there was 65W 4c ES and i didn't heard of 95W 4c ES. But this does not mean there will not be retail 95W 4c chips...

 

[bjt2]

- Really? See thats what I am talking about as conjecture.
Keller said that some of Zen has similarities with some of BD and from that you conclude that he means its a high hz design?
You familiar with erasmus montanus?

https://en.wikipedia.org/wiki/Erasmus_Montanus

A rock cant fly, my mother cant fly, thus my mother must be a rock. (dont tell her that though)

The only thing that can mean Keller statement is that Zen is a low FO4 design. This is confirmed by 19 stage integer pipeline, higher clock and lower power on an inferior process and 5GHz overclock on air.
If you still think that a 3.6GHz 95W CPU will overclock less than a 3.2GHz 140W CPU, then it's you that are a rock...
And erudite citations does not change reality...

 

[cytg111]

There is not only AMD, but also CanardPC. 5GHz on air on an early MB and ES, limited only by the VRMs.
Zen has more stages than intel architectures.
Zen already proved to clock higher and with lower power than 6900K. And it's not even finished.
It's physically impossible that the CPU suddenly draw 300W after 4GHz. And I already posted a graph with 14nm frequency scaling.
If a 3.6GHz 95W CPU can't clock over 4.4Ghz, why a 3.2GHz 140W CPU can? Because it have the INTEL magic process?

Two sources.. semi sources.. easter egged "5ghz on air" binary printed, maybe for clickbait? 7 cores disabled, during winter at -20 degrees C.. who knows.. Lets not forget haswell ES that clocked 5GHz+ while retail bummed out.
You Dont Know Yet. You cant invent new data from nothing, there is laws in place to prevent such things, like the 2nd law of thermodynamics.. no free lunches..

 

[cytg111]

If you still think that a 3.6GHz 95W CPU will overclock less than a 3.2GHz 140W CPU, then it's you that are a rock...

I dont think thats likely. It is just not a FACT yet!

 

[itsmydamnation]

- Really? See thats what I am talking about as conjecture.
Keller said that some of Zen has similarities with some of BD and from that you conclude that he means its a high hz design?
You familiar with erasmus montanus?

https://en.wikipedia.org/wiki/Erasmus_Montanus

A rock cant fly, my mother cant fly, thus my mother must be a rock. (dont tell her that though)

Except for the part where Jim Kellar talks about taking the DNA from there "high frequency designs" and their "low power designs" and mixing the best of both. he also talks about how amd know how to do high frequency circuit design and leveraging that for the next generation........

https://youtu.be/SOTFE7sJY-Q?t=4m45s

 

[cytg111]

Except for the part where Jim Kellar talks about taking the DNA from there "high frequency designs" and their "low power designs" and mixing the best of both. he also talks about how amd know how to do high frequency circuit design and leveraging that for the next generation........

https://youtu.be/SOTFE7sJY-Q?t=4m45s

Yea, sounds good. What was he going to say, that it would clock low and use alot of power?
may 2014 .. wonder if Keller knew how 14nm FF at glofo was gonna mature? Fab deals. Upcoming Errata.
I want Zen to clock to the skies but I dont see that you have the evidence to support the claim at this time, only conjecture.

 

[bjt2]

Two sources.. semi sources.. easter egged "5ghz on air" binary printed, maybe for clickbait? 7 cores disabled, during winter at -20 degrees C.. who knows.. Lets not forget haswell ES that clocked 5GHz+ while retail bummed out.
You Dont Know Yet. You cant invent new data from nothing, there is laws in place to prevent such things, like the 2nd law of thermodynamics.. no free lunches..

I am not inventing. 3.6GHz Base clock ES still at 95W and retail 3.2GHz base clock at 140W TDP are a reality. I can't believe that a chip with +400MHz base clock and less TDP will overclock less...

 

[bjt2]

I dont think thats likely. It is just not a FACT yet!

The OC potential is not a fact, ok, but the higher base clock, with less TDP (and actual consumption) it is a fact. And this will count for the vast majority of users... You may concede me that overclocker are not the vast majority of customers...