New Zen microarchitecture details

[KTE]

Expecting 2.5x PD performance 8v8 core is a pipe dream. I'm left flabbergasted.

It'd be SPECTACULAR to hit 2.5x... 4C/4T PD vs 8C/16T Zen.

In highly MT workloads.

Clock for clock, I am expecting 2.3x performance of PD 4C with 8C Zen (non synthetics)

Once again I still feel their struggle will be like with Phenom: power and clocks.


Sent from HTC 10

 

[Abwx]

Once again I still feel their struggle will be like with Phenom: power and clocks.

Power wont be an issue, and actually AMD should even have a perf/Watt advantage process wise..


At same voltage 14nm LPP work at 20% higher frequency than Intel s 14nm, at 0.8V the former achieve 2.4GHz while the latter barely reach 2GHz.
This advantage extend up to the range at wich GF s transistors start to have a second knee, wich at this point is an unknown but it s 100% sure that it wont happen before 1.1V, so 4GHz shouldnt be a problem as much as the core count.

 

[itsmydamnation]

Expecting 2.5x PD performance 8v8 core is a pipe dream. I'm left flabbergasted.

It'd be SPECTACULAR to hit 2.5x... 4C/4T PD vs 8C/16T Zen.

In highly MT workloads.

Clock for clock, I am expecting 2.3x performance of PD 4C with 8C Zen (non synthetics)

Once again I still feel their struggle will be like with Phenom: power and clocks.

so your expect nothing much more then what AMD got going from bulldozer to piledriver, all the while they are implementing:

uop cache
loop buffer
double the width of execution units
prefetches and predictors become dedicated per core
much reduced L2 latency
increased L1 latency size
likely stack cache
+
all the normal incremental improvements to load store system, predictors, prefetches, schedulers, etc.

And in bulldozer to piledriver we had load store improvements, decode improvements and moving some instructions that had always lived on ALU's to AGLU because the ALU's are under such congestion.

Im amazed how some of you think Zen is just going to magically run slower just because. Actually look at what the likely uarch changes are and what it means for clocks. The frontend and backend of CON cores are already just as wide as ZEN, its execution resources and cache that are going to get the biggest internal changes. The idea that AMD suddenly forgot how to make the parts that already run at 4ghz+ run at 4+ghz is quite retarded. its not like the 6 pipeline STARS core on 45nm with a really short pipeline length couldn't hit 3.5+ghz..............

 

[The Stilt]

At same voltage 14nm LPP work at 20% higher frequency than Intel s 14nm, at 0.8V the former achieve 2.4GHz while the latter barely reach 2GHz.

You base your "Intel 14nm processes cannot operate higher than 2.0GHz at 0.8V" statement on? Still on a single screenshot with had Broadwell-U running > 0.8V @ 2.0GHz?
What is doable in lab (or paper) has nothing to do with what is doable either technically of cost wise in production.

 

[Abwx]

You base your "Intel 14nm processes cannot operate higher than 2.0GHz at 0.8V" statement on? Still on a single screenshot with had Broadwell-U running > 0.8V @ 2.0GHz?
What is doable in lab (or paper) has nothing to do with what is doable either technically of cost wise in production.

I m not talking of what is done in labs but of industrial standards for commercialisation of semiconductors.

If you want the whole picture the measured 2.4GHz is not at 0.8V but at 0.72V, wich is the former less 10%, wich is the minimal voltage margin of a commercial product, so we re talking of actual specs as they are find in finished designs.

Likewise power is not measured at 0.8V but at 0.88V, that is, with 10% excess voltage, all thoses datas are explicit on GF s slides..

 

[The Stilt]

Here is GF28A (Kaveri) and GF32SHP SOI (Vishera) voltage scaling (Vmin vs. Fmax) plotted against the 14nm LPP fantasy. The measurements were made quite long ago (in 2014) so neither of the chips represent the latest available production (process wise). They were both recorded on factory proofed (measured characteristics) silicon. There will be variation between specimen to specimen, however both of the parts were picked to represent the average of their own production variation (GF28A High SIDD, HS), (GF32SHP Low SIDD, LS).

Expecting a purely low power targeted manufacturing process to outperform a custom higher performance 28nm (GF28A) process by 1GHz @ 1100mV and a super high performance 32nm SOI (GF32SHP) by 900MHz @ 1100mV is much more than overly optimistic. It is unrealistic.

Even more so when both of the high performance processes were deployed under designs which were built purely for speed, while Zen is completely the opposite. We know how well the revamped Hound-family (12h, Stars) did even on the ultra high frequency capable 32nm SHP SOI process.

I added a fourth line to the chart to represent Vmin-Fmax scaling I would personally expect the 14nm LPP to have (with Zeppelin).

 

[Abwx]

Here is GF28A (Kaveri) and GF32SHP SOI (Vishera) voltage scaling (Vmin vs. Fmax) plotted against the 14nm LPP fantasy. The measurements were made quite long ago (in 2014) so neither of the chips represent the latest available production (process wise). They were both recorded on factory proofed (measured characteristics) silicon. There will be variation between specimen to specimen, however both of the parts were picked to represent the average of their own production variation (GF28A High SIDD, HS), (GF32SHP Low SIDD, LS).

Expecting a purely low power targeted manufacturing process to outperform a custom higher performance 28nm (GF28A) process by 1GHz @ 1100mV and a super high performance 32nm SOI (GF32SHP) by 900MHz @ 1100mV is much more than overly optimistic. It is unrealistic.

Even more so when both of the high performance processes were deployed under designs which were built purely for speed, while Zen is completely the opposite. We know how well the revamped Hound-family (12h, Stars) did even on the ultra high frequency capable 32nm SHP SOI process.

I added a fourth line to the chart to represent Vmin-Fmax scaling I would personally expect the 14nm LPP to have (with Zeppelin).

What have 32nm and 28nm have to do here..?..

Are 14nm finfets manufactured using thoses as basis..?.

Actualy it s you who are using fantasis trying making points, the purple curve is indeed a exemple of wishfull thought, i guess that you engaged too much in claims about 2.8-3.0 frequencies, now you are stuck with irrelevancies, mixing datas that have nothing to do with the discussion to better hide the lack of arguments..

Edit / Your curves fo previous processes are with 0% voltage margin, the 0.8V at 2.4GHz include the 10% voltage margin, next time read better before answering that fast :

If you want the whole picture the measured 2.4GHz is not at 0.8V but at 0.72V, wich is the former less 10%, wich is the minimal voltage margin of a commercial product

 

[The Stilt]

All transitions to smaller process since 32nm have resulted in reduced Fmax, for both AMD and Intel. Why would it any different with 14nm LPP, especially when the design itself no longer targets high speeds?

 

[Arachnotronic]

All transitions to smaller process since 32nm have resulted in reduced Fmax, for both AMD and Intel. Why would it any different with 14nm LPP, especially when the design itself no longer targets high speeds?

The Stilt, check this out:

http://semiaccurate.com/forums/showpost.php?p=265281&postcount=3064

Looks like current models are at ~2.4GHz, target is 3.6GHz on the 8 core model (though the AMD Red Team member apparently thinks they'll hit closer to 3.3-3.4GHz).

 

[The Stilt]

I don't trust anything I read on that forum, but anything between 3.0 - 3.6GHz sounds quite right. Although 2.4GHz in the other hand sounds too low, at least for the final silicon :sneaky:

 

[itsmydamnation]

The Stilt, check this out:

http://semiaccurate.com/forums/showpost.php?p=265281&postcount=3064

Looks like current models are at ~2.4GHz, target is 3.6GHz on the 8 core model (though the AMD Red Team member apparently thinks they'll hit closer to 3.3-3.4GHz).

You know your reaching when you have to quote internet strong man. Notice how he never references anything.
Thats because he takes everything out of context just so he can be "right".

 

[Arachnotronic]

You know your reaching when you have to quote internet strong man. Notice how he never references anything.
Thats because he takes everything out of context just so he can be "right".

He's quoting an AMD Red Team+ member.

 

[USER8000]

I don't trust anything I read on that forum, but anything between 3.0 - 3.6GHz sounds quite right. Although 2.4GHz in the other hand sounds too low, at least for the final silicon :sneaky:

So 8C/16T Zen at 3GHZ to 3.6GHZ at a 95W TDP seems comparable in clockspeeds to the 8C/16T 6900K running at a 140W TDP.

 

[itsmydamnation]

He's quoting an AMD Red Team+ member.

Reread what I said, it's pretty simple. He will constantly remove context, so without a reference (which he never provides) you never know what is meant by the person he is quoting.

 

[USER8000]

Reread what I said, it's pretty simple. He will constantly remove context, so without a reference (which he never provides) you never know what is meant by the person he is quoting.

I think it is funny how he is saying if AMD is not getting above 4GHZ on an 8C/16T part at 95W TDP,does not have over 8 cores and does not beat Skylake it is a fail.

Yet,he is so caught up in his negativity that he forgets parts like the Core i7 6900K are running at well under 4GHZ for 8C/16T parts and Intel has the part rated at a 140W TDP.

If AMD can get to Haswell levels of IPC that will massively close the gap with Intel especially if power consumption drops quite a bit too. The Core i7 6900K is a $1089 part too.

 

[Abwx]

All transitions to smaller process since 32nm have resulted in reduced Fmax, for both AMD and Intel. Why would it any different with 14nm LPP, especially when the design itself no longer targets high speeds?

The limitation is in the transistor power dissipation, if shrinked 1.4x they will have 2x less area, so their perf/watt should be increased by 2x if you want to operate them at same frequency.

Intel s 14nm didnt improve much their perf/Watt compared to their 22nm, i once stated it ad nauseam but seems people are keeping using them as if they were the end of all, 14nm LPP work at lower voltage for the same frequency, so the individual transistors dissipation will be less in a square proportion of the voltage, that s physics laws and is not much subject to any debate set apart correcting absolute values of various parameters.

 

[Arachnotronic]

Reread what I said, it's pretty simple. He will constantly remove context, so without a reference (which he never provides) you never know what is meant by the person he is quoting.

It seems pretty clear what the Red Team+ member is saying. It seems to me that you are just attacking juanrga because you don't like him.

 

[naukkis]

So 8C/16T Zen at 3GHZ to 3.6GHZ at a 95W TDP seems comparable in clockspeeds to the 8C/16T 6900K running at a 140W TDP.

Probably 6900K without 256bit AVX could also have 95W tdp.....

 

[itsmydamnation]

It seems pretty clear what the Red Team+ member is saying. It seems to me that you are just attacking juanrga because you don't like him.

And yet all he has to do is provide a link and it would be clear as day, but it never happens.

If I just started quoting stuff like he does would you give it the same credit and backing, we all know the answer.....

 

[KTE]

Power wont be an issue, and actually AMD should even have a perf/Watt advantage process wise..

At same voltage 14nm LPP work at 20% higher frequency than Intel s 14nm, at 0.8V the former achieve 2.4GHz while the latter barely reach 2GHz.
This advantage extend up to the range at wich GF s transistors start to have a second knee, wich at this point is an unknown but it s 100% sure that it wont happen before 1.1V, so 4GHz shouldnt be a problem as much as the core count.

What Idd do the two corresponding process hit at 1.0V?

Even if 3GHz is possible process and yield wise, how can you be certain of the leakage and power at such frequencies?

At 0.8V, it still doesn't preclude the process variability and Vdd scaling drop-offs that common.

The Stilt, check this out:

http://semiaccurate.com/forums/showpost.php?p=265281&postcount=3064

Looks like current models are at ~2.4GHz, target is 3.6GHz on the 8 core model (though the AMD Red Team member apparently thinks they'll hit closer to 3.3-3.4GHz).

Who is that person quoting an anonymous source?

Sent from HTC 10

 

[AtenRa]

All transitions to smaller process since 32nm have resulted in reduced Fmax, for both AMD and Intel. Why would it any different with 14nm LPP, especially when the design itself no longer targets high speeds?

There is no reduced fmax, there is not increased fmax with lower power.

But, we have increased frequencies at lower power.

Examples,

20nm LPM at the same power is 42% faster (higher frequency) than 28nm SLP

14nm LPP is 47% faster at the same power than 28nm HPP.

16nm FF is 35% faster than 28nm HK/MG

 

[Arachnotronic]

Probably 6900K without 256bit AVX could also have 95W tdp.....

Yeah, the latest Intel chips have usually consumed less power at full (non-AVX) load than what their TDPs are rated at, but AVX2 is really a power hog.

 

[Abwx]

There is no reduced fmax, there is not increased fmax with lower power.

But, we have increased frequencies at lower power.

Examples,

20nm LPM at the same power is 42% faster (higher frequency) than 28nm SLP

14nm LPP is 47% faster at the same power than 28nm HPP.

16nm FF is 35% faster than 28nm HK/MG

What Idd do the two corresponding process hit at 1.0V?

Even if 3GHz is possible process and yield wise, how can you be certain of the leakage and power at such frequencies?

At 0.8V, it still doesn't preclude the process variability and Vdd scaling drop-offs that common.

Who is that person quoting an anonymous source?

Sent from HTC 10

All is in the second slide of Aten Ra s post, FTR dynamic power was measured at 0.880V, wich tell us that at this voltage everything is perfect, they wouldnt use value that would crush the metrics, wouldnt they..?..

Beside if it cant do 1V it would be a first in the industry, dont count on such irrationnal theories..

 

[SAAA]

Help me here: were was ever staded max 95W TDP for Zen? Because everyone says that but I can't find the original source, it almost seems like a fixation...

95W sounds about right for 6-cores at high IPC and frequency but it's quite limited if you have 8-cores. Unless the turbo speeds are much higher than stock, then it could work but it would need decent cooling.

In regard to that... if wraith cooler is any indication 125W TDP is a more plausible estimation for the top Zen. Why design a new cooler and don't make it compatible with an incoming product?

 

[Arachnotronic]

Help me here: were was ever staded max 95W TDP for Zen? Because everyone says that but I can't find the original source, it almost seems like a fixation...

95W sounds about right for 6-cores at high IPC and frequency but it's quite limited if you have 8-cores. Unless the turbo speeds are much higher than stock, then it could work but it would need decent cooling.

In regard to that... if wraith cooler is any indication 125W TDP is a more plausible estimation for the top Zen. Why design a new cooler and don't make it compatible with an incoming product?