New Zen microarchitecture details

[William Gaatjes]

Brad used the 'HP' since Bobcat/Jaguar, but this is more advanced. He's also used it for the Samsung M1 http://www.androidauthority.com/closer-look-samsung-mongoose-cpu-712587/


Sent from HTC 10
(Opinions are own)

A, interesting. Thank you.
While traversing the fora, i noticed some people think the AMD design team, took a Bobcat/Jaguar design and replaced all performance limiting factors in that design while adding new functions like SMT and the uopcache to get as end result a high performance CPU.

If true, that reads like how core and Nehalem came to be after the P4, the Israeli design team of Intel took the P3 and replaced all performance limiting factors in that design while adding the best features of the P4 to the design.

The upcoming years are going to be exciting and boring at the same time. Blows will be given forth and back.

 

[biostud]

Cryptic message "ZenOC@Air=5G"

https://www.techpowerup.com/229090/amds-upcoming-ryzen-chips-to-reportedly-overclock-5-ghz-on-air

 

[iBoMbY]

There was this new patent from AMD released recently:

Thermal oscillator

Embodiments described herein improve on the state-of-the-art by being approximately 150 times smaller in area than current diode implementations for thermal sensors that are known to the inventor.

Could this be how they are managing to put so many (what did they say, hundreds?) sensors into Summit Ridge?

Edit: Ohh, and I really hope they will provide a way to query all these sensors per software, and a map of their position. This way it would be possible to create an interactive thermal (and maybe voltage) image of the die in monitoring software, which may not be useful for most users, but could look pretty cool.

 

[Doom2pro]

Ohh, and I really hope they will provide a way to query all these sensors per software, and a map of their position. This way it would be possible to create an interactive thermal (and maybe voltage) image of the die in monitoring software, which may not be useful for most users, but could look pretty cool.

That would be boss.

 

[William Gaatjes]

Brad used the 'HP' since Bobcat/Jaguar, but this is more advanced. He's also used it for the Samsung M1 http://www.androidauthority.com/closer-look-samsung-mongoose-cpu-712587/


Sent from HTC 10
(Opinions are own)

I am reading just for the fun of it about Jaguar from the realworld tech site :
http://www.realworldtech.com/jaguar/2/

Conditional near branches are implicitly predicted as not-taken, which saves space in the BTB. Once such a branch is taken, it is set to always taken in the BTB. Should the always taken branch subsequently fall through, it switches to a dynamic neural network predictor using 26-bits of global history. The Bobcat and Jaguar branch predictor proved to be so successful that it was later adopted for AMD’s big cores, particularly Piledriver.

Is this the only neural net that AMD Lisa Su talked about with respect to (Ry)zen?
I got the impression it was promoted as if it was revolutionary. But they have been using it since Jaguar.
Maybe they found a way to improve it ?

 

[DrMrLordX]

Maybe they found a way to improve it ?

I believe the implication is that they have made improvements, hence their decision to market the new branch prediction technology under the SenseMI label.

 

[KTE]

I am reading just for the fun of it about Jaguar from the realworld tech site :
http://www.realworldtech.com/jaguar/2/

Is this the only neural net that AMD Lisa Su talked about with respect to (Ry)zen?
I got the impression it was promoted as if it was revolutionary. But they have been using it since Jaguar.
Maybe they found a way to improve it ?

There are neural networks, and then predictors based on it. There was another paper: http://cseweb.ucsd.edu/~atsmith/rnn_branch.pdf
And a good book to read: https://books.google.co.uk/books?id=nxqrDAAAQBAJ&lpg=PA1&ots=jFaAsdw8nD&dq
http://www.solver.com/neural-network-prediction

Neural network prediction mimics the brain, and it's cutting edge AI work, most of it not even patented, so very secret and it's damn complex in the way it really works. Brads work on Bobcat and then on the Samsung M1 used a more iterative but perfected version (2 branches taken per cycle, bigger history and weighting tracking). The predictor models remain the same as found in research (smith/gshare/McFarling) but they become tweaked around the hardware budget of their implementation. You're missing 99% of its detail yet

Sent from HTC 10
(Opinions are own)

 

[William Gaatjes]

It makes sense that the engineers who designed Zen for AMD found ways to improve its accuracy. I have some basic understanding of what is going on with a neural network. It does always reminds me a bit of a PID control.

 

[.vodka]

Buildzoid quotes Flank3r, trying not to break any NDAs. I'm sure you've heard the name on the AMD overclocking scene. Both names, actually.

Start at 13 minutes. Basically, Zen as is now, unpolished, unfinished does 4.2-4.4GHz on all eight cores on average. 5GHz happens at unhealthy voltages... expected. Overclocks like Broadwell-E. Is at Broadwell-E's level more or less. Is cheap not only for the CPU but also for the motherboards, being less complex with two less memory channels, etc.

Also, Zen does not have a cold bug. The overclockers are going to have their fun with this. What will overclock better under LN2? Phenom II or Zen?

Intel's precious monopoly maintaining margins are about to take a dive, I think.


Come on CES, why does time have to pass so slowly...

 

[superstition]

How does one get that type of clockspeed with a low power process?

 

[bjt2]

How does one get that type of clockspeed with a low power process?

Lower FO4.

 

[cytg111]

4.2 to 4.4 from GloFo first slice of the wafer. Ill take it. Bodes well going forward, I am seeing ~5 in my crystal ball here.

 

[rtsurfer]

Maybe with Zen+...

Sent from my Nexus 5X using Tapatalk

 

[bjt2]

That guy in that video suppose just same OC than INTEL chips... Just to not disappoint INTEL fans. How can he say that? An unfinished ES has yet higher base clock than a retail product, consuming less. Why a final chip can't OC higher than BWE? Because it's a low power process? Because INTEL can't and so AMD, because it can't beat INTEL? Had he ever heard of FO4? BD still OC higher on an inferior process and Jim Keller stated that Zen has same "DNA" than Zen, so it's an high clock design...
This guy don't even know the pipeline differences between Zen and SKL and still rants about SMT efficiency that can't be superior to INTEL because noone can't beat INTEL. Well. A very well argumented thesis... Why INTEL SMT implemetation is superior? Because intel have SMT since pentium4? Quite convincing arguments... The truth is that we have 4int+4fp uop/cycle vs 4 int+FP (max 2 FP for INTEL). Even a low cost auto can beat a luxury car, if it has an engine powerful enough. And Zen engine is more powerful.

 

[rtsurfer]

He doesn't know all your theories, but he knows how the actual chip is Overclocking. He knows the practical stuff.

Prolly wrong about Intel's SMT being better though.

Sent from my Nexus 5X using Tapatalk

 

[Atari2600]

Start at 13 minutes. Basically, Zen as is now, unpolished, unfinished does 4.2-4.4GHz on all eight cores on average. 5GHz happens at unhealthy voltages... expected. Overclocks like Broadwell-E. Is at Broadwell-E's level more or less.

Our next question is then...

With AMD's built in "overclocking" utilities, does it somehow do 4.2-4.4 GHz within the 95W thermal envelope? Or is that with a relaxation of the 95W TDP? (I would lean to the latter, but I wonder just how far above - if it does that at <125W....)

 

[bjt2]

He doesn't know all your theories, but he knows how the actual chip is Overclocking. He knows the practical stuff.

Prolly wrong about Intel's SMT being better though.

Sent from my Nexus 5X using Tapatalk

He does not have Zen yet, and what the czech user have is still an ES, who knows how old, with an unfinished MB and BIOS and AGESA... If CanardPC source managed to obtain 5GHz on air, even if on one core, probabily for MB fault, why expect 4.2-4.4GHz, suspiciously near or under 6900K max clocks? As every overclocker knows, the power grows exponentially with frequency and even he says that, later in the video, saying that OC on 8, 6 and 4 core is not much higher than OC on 10 core 6950X. If someone managed to have 5GHz on one core, we can expect 4.7-4.8 ghz at least and not 4.2-4.4, just because INTEL can't...

 

[rtsurfer]

He does not have Zen still, and what the czech user have is still an ES, who knows how old, with an unfinished MB and BIOS and AGESA... If CanardPC source managed to obtain 5GHz on air, even if on one core, probabily for MB fault, why expect 4.3-4.4GHz, suspiciously near or under 6900K clocks? As every overclocker knows, the power grows exponentially with frequrency and even he says that later in the video, saying that OC on 8, 6 and 4 core is not much higher than OC on 10 core 6950X. If someone managet to have 5GHz on one core, we can expect 4.7-4.8 ghz at least and not 4.2-4.4, just because INTEL can't...

He quoted Czech user because outright stating the numbers that he has seen would put him in violation of NDA. Doesnt mean that the chips he saw overclocked were as old as Czech user's or run on the same BIOS/board. What matters is that the numbers he saw are same as the ones posted by Flanker & mine too.
It has nothing to do with AMD can't do it because Intel can't, but everything to do with the process AMD has to manufacture the chips on. We all would love to see AMD prosper, Bullzoid is quite the AMD fanboy himself. And while there might be some variation in the final retail silicon, 4.7-4.8 is too much to expect. I would LOVE to be wrong on this, but I see little chance of that happening.

I dunno what has made CanardPc claim what they did.

 

[bjt2]

He quoted Czech user because outright stating the numbers that he has seen would put him in violation of NDA. Doesnt mean that the chips he saw overclocked were as old as Czech user's or run on the same BIOS/board. What matters is that the numbers he saw are same as the ones posted by Flanker & mine too.
It has nothing to do with AMD can't do it because Intel can't, but everything to do with the process AMD has to manufacture the chips on. We all would love to see AMD prosper, Bullzoid is quite the AMD fanboy himself. And while there might be some variation in the final retail silicon, 4.7-4.8 is too much to expect. I would LOVE to be wrong on this, but I see little chance of that happening.

I dunno what has made CanardPc claim what they did.

Maybe in the first batches there are huge variations on silicon quality or the Canard PC source has a different step with different scaling...
You are an overclocker, so you should know that 5GHz on air, even on one core only, if true, would mean 4.7-4.8GHz on all cores... And that was at least a june ES... I don't think that newer ES would ever go outside AMD labs. For electrical MB validation even an old ES would suffice. What B&C said on twitter is that retail chips have higher frequencies than expected and since they are 6 months newer, at least, of june ES, I think that 5GHz on one core should be feasible on most chips... And conversely 4.7-4.8 all core...
Even Bristol Ridge 65W overclock at 4.9GHz on air and on 28nm! I don't buy that Zen could not go over this...

 

[F-Rex]

How does one get that type of clockspeed with a low power process?

Ask the_stilt. Hé Will explain you how zen on 14lpp is limited to 2.8/3.2ghz turbo.
Best process analyst ever.

 

[cytg111]

He does not have Zen yet, and what the czech user have is still an ES, who knows how old, with an unfinished MB and BIOS and AGESA... If CanardPC source managed to obtain 5GHz on air, even if on one core, probabily for MB fault, why expect 4.2-4.4GHz, suspiciously near or under 6900K max clocks? As every overclocker knows, the power grows exponentially with frequency and even he says that, later in the video, saying that OC on 8, 6 and 4 core is not much higher than OC on 10 core 6950X. If someone managed to have 5GHz on one core, we can expect 4.7-4.8 ghz at least and not 4.2-4.4, just because INTEL can't...

Why are you on about Intel? ~4.4 is a range we have heard before and going by baysian reasoning, this is where I put my three fiddy.

 

[bjt2]

Why are you on about Intel? ~4.4 is a range we have heard before and going by baysian reasoning, this is where I put my three fiddy.

Someone is not saying the truth or the ES that are going around are very different. IF IT'S TRUE that someone got a 5GHz air overclock, even on one core, there is no way that the air OC is limited to 4.4GHz. Every overclocker knows that the power will exponentially rise with clock, so 5GHz on air, even if it's a suicide run, means at least 4.6-4.7GHz on all cores. This IF Canard PC source said the truth.
And that was a buggy early ES. According to B&C newer steps clocks even better... So... I don't think that on retail chips they will be limited to 4.4GHz OC. Maybe early ES. And even this is questionable if Canard PC source said the truth...

 

[Dygaza]

What kind of power consumption would you guys expect when overclocked to 4.2Ghz with all cores, or let's say even 4,7Ghz with all cores.

 

[JoeRambo]

You are an overclocker, so you should know that 5GHz on air, even on one core only, if true, would mean 4.7-4.8GHz on all cores... And that was at least a june ES... I don't think that newer ES would ever go outside AMD labs. For electrical MB validation even an old ES would suffice. What B&C said on twitter is that retail chips have higher frequencies than expected and since they are 6 months newer, at least, of june ES, I think that 5GHz on one core should be feasible on most chips... And conversely 4.7-4.8 all core...

So Your claim is that AMD is incompetent in only releasing 3.4+Ghz chip, when they could release an uber winner @4ghz or so at higher TDPwattage? Instead of releasing a winner, they are sticking with 3.4Ghz base and leaving >1Ghz on the table???

It is way more likely 8C Zen will top out @some 4.2-4.3Ghz on top air and power usage will be rather high.

Note that those are still very respectable clocks for performance they are showing us, but zealotry is doing no good for Zen.

 

[iBoMbY]

To quote Samsung:

14LPP (Performance boosted edition) is the 2nd FinFET generation which the performance is enhanced up to 10%. 14LPP is the single platform for every application designs with the improved performance for computing/Network designs and the lowered power consumption for Mobile/Consumer designs. 14LPP will be the main process technology offering in 2016 and after.

If you clock it within the golden range, it uses less power, but that doesn't mean you can't clock it higher with the cost of efficiency.

Edit: It could clock about 30% higher than 28nm, with not much changed GCN cores, and they are not a high-frequency design. It all depends on the exact design of Ryzen, but from what I can tell they didn't want to build a CPU which can only be low-power.