New Zen microarchitecture details

[Abwx]

Abwx: You're using an ambiguous marketing slide without figures as evidence for frequency gain at the same power.

That slide could be in the 2GHz range... Or the 1GHz.

Can you show?

It s clearly for the whole frequency range, this is visible on this graph by the way the curves are shaped,
the blue dotted line is the power in function on frequency of GF 28nm, it end at about 4.3GHz for frequency,
AMD made the same when they published XV comsumption in respect of Steamroller, the curves were also spanning the full range albeit being scaled on the axys.

FTR if it was in a restricted frequency range the curves would be close to straight lines, here it is clear that these are polynomials based curves with the knees when the curve transit to higher degrees also clearly visible..

Even if you think FO4 is the same or similar (it's isn't, it's worse for clocks, as BD was a speed demon design. Speed demon = low IPC x high clocks):

FO4 delay is smaller than with 28nm, for the simple reason that 14nm LPP transistor have 20% lower switching capacitance than 28nm, this should theorically allow 20% higher frequency just with this improvement..

Since transconductance has been increased by 1.5x this will add on the previous improvement and the total is theorically 45% lower FO4 delay, although this cant be achieved because the strain capacitances and routings resistances and inductances will take their toll and limit what can be actually realised, but still, working at frequencies comparable to Kaveri is trivial.

XV 3.5GHz 8C hits 130W - best possible case - based on the top DT model (in reality it is more).

The 4C 3.5 use 46W in Cinebench but with 8C it would be at 130W..?..
A 8C would consume 70% more due to some circuits not being duplicated..

I suggest that you make some reasearch before throwing random numbers whose only result is to produce useless discussions, see, i wouldnt have edited my post if i didnt notice this non sense..

 

[coercitiv]

Have you read the remaining of my post?
I gave solid evidence that that, was not an optimistic claim...

First, it was optimistic from the moment you wrote "AMD stated". Second, both AMD claims and your own deductions are partially based on 14nm benefits, which I included in my basic napkin exercise as a conservative 30% drop in power for the upper range of frequency spectrum. You're welcome to increase that even further, what value do you reckon is more representative for 14nm vs 28nm at 3.6Ghz and beyond?

 

[bjt2]

First, it was optimistic from the moment you wrote "AMD stated". Second, both AMD claims and your own deductions are partially based on 14nm benefits, which I included in my basic napkin exercise as a conservative 30% drop in power for the upper range of frequency spectrum. You're welcome to increase that even further, what value do you reckon is more representative for 14nm vs 28nm at 3.6Ghz and beyond?

As other in this thread said, 14nm FF has lower capacitance and higher transconductance. And leakage is 1/6. So this tells us that at same frequency and same transistors number, the current is lower, given the correct Vcore.

What is the correct Vcore? For 28nm it's about 1.35V. For 14nm it's about 1.05-1.1V. Let's assume the worst. It's more than 20% less. That is more than 40% less power from the Vcore alone.

But the current is also inferior.

So the power is more than 40% less than 28nm.

GF claim of -65% makes sense.

I would expect at least -50%.

Polaris has +15% frequency and -30/-40% power, so it's safe to assume -50% at same frequency. And the Vcore if i recall correctly it's about or below 1.1V, as hypotized.

 

[AtenRa]

Comparing an 8 Module 16 Threads XV to 8 Core 16 Threads ZEN to estimate the Fmax is pointless. Different mArchitectures, different lithographic process etc etc.

 

[bjt2]

Comparing an 8 Module 16 Threads XV to 8 Core 16 Threads ZEN to estimate the Fmax is pointless. Different mArchitectures, different lithographic process etc etc.

BD has 15 stage pipeline, Zen 19 stage. So in a sense you are right. Zen will be faster at same process. It is manufactured even in a better process, so...

It's logic. The fact that according to many of you it's physically impossible for the tiny AMD to make a cpu faster than the mighty INTEL, it's not a dogma...

Let's examine your statements:
Different uarch. You are right. 19 vs 15 stages is different. Zen arch is faster.
Different process. You are right. 14nm FF vs 28nm BULK. Zen process is faster.

How this sum up to Zen having clock lower of BD i don't get it.

The only reason (much more a hope) is to not surpass INTEL, given that IPC seems comparable, from the blender tests.

AMD has already done an 8 core at frequency much higher than your lovely BDW-E, but since it does not surpass it, meh... We can safely ignore it, even if it is made on the old 28nm bulk process and is anyway a great piece of engineering. But performance is all that count, so ok...

But now Zen has the IPC. Zen should have the frequency. And all of you talk of max 3-3.2GHz that is casually the same frequency of the unsurpassable BDWE of the mighty INTEL. This is your opinion unsupported by facts.

Facts supporting my prediction I have given a lot. And i don't want to repeat them.

But please do not throw them away with your unsupported statements.

Give some arguments other than "different process, different arch so you can't say anything"...

 

[Arachnotronic]

BD has 15 stage pipeline, Zen 19 stage. So in a sense you are right. Zen will be faster at same process. It is manufactured even in a better process, so...

It's logic. The fact that according to many of you it's physically impossible for the tiny AMD to make a cpu faster than the mighty INTEL, it's not a dogma...

Let's examine your statements:
Different uarch. You are right. 19 vs 15 stages is different. Zen arch is faster.
Different process. You are right. 14nm FF vs 28nm BULK. Zen process is faster.

How this sum up to Zen having clock lower of BD i don't get it.

The only reason (much more a hope) is to not surpass INTEL, given that IPC seems comparable, from the blender tests.

AMD has already done an 8 core at frequency much higher than your lovely BDW-E, but since it does not surpass it, meh... We can safely ignore it, even if it is made on the old 28nm bulk process and is anyway a great piece of engineering. But performance is all that count, so ok...

But now Zen has the IPC. Zen should have the frequency. And all of you talk of max 3-3.2GHz that is casually the same frequency of the unsurpassable BDWE of the mighty INTEL. This is your opinion unsupported by facts.

Facts supporting my prediction I have given a lot. And i don't want to repeat them.

But please do not throw them away with your unsupported statements.

Give some arguments other than "different process, different arch so you can't say anything"...

Believe whatever you'd like! Why is it so important to you to prove to a bunch of strangers on an internet message board that Zen will have massive perf/clock and clock like crazy?

Some of us are just giving our predictions, take them for what they're worth.

The silicon should be out soon, so we'll be able to put all of this back and forth to rest soon enough.

 

[AtenRa]

Personally I really want to see the 15W TDP Mobile APUs and what performance they can achieve vs 14nm Intel 15W TDP Mobile SKUs. Because FinFets are like FD-SOI, they really shine at low power.

Also its about time we get a real 4-5W TDP APU for x86 11.6" Two in One Tablet/Laptop market, leaving Intel alone just killed that segment also. Having competition with CoreM will ignite that market with much better products stilling marketshare from 10-12" Android Tablets.

 

[Arachnotronic]

Personally I really want to see the 15W TDP Mobile APUs and what performance they can achieve vs 14nm Intel 15W TDP Mobile SKUs. Because FinFets are like FD-SOI, they really shine at low power.

Also its about time we get a real 4-5W TDP APU for x86 11.6" Two in One Tablet/Laptop market, leaving Intel alone just killed that segment also. Having competition with CoreM will ignite that market with much better products stilling marketshare from 10-12" Android Tablets.

I don't follow your reasoning here. How does having an additional chip supplier impact demand for devices?

 

[coercitiv]

So the power is more than 40% less than 28nm.

GF claim of -65% makes sense.

I would expect at least -50%.

Just so we're clear, a minimum of -50% drop in power would put Zen 8C at 100W for 4Ghz. Do you stand by this prediction?

 

[Phynaz]

Comparing an 8 Module 16 Threads XV to 8 Core 16 Threads ZEN to estimate the Fmax is pointless. Different mArchitectures, different lithographic process etc etc.

This x10

 

[bjt2]

Believe whatever you'd like!

Believe is not the correct word. It's not faith. I gave you numbers, arguments, logical deductions.

If after that, YOU believe that Zen will have low IPC and/or clock max 3.2GHz, without argumenting, then this is you that is using faith.

You can believe that Zen will clock at maximum 3.2GHz, but i gave you solid arguments to think otherwise. It's up to you to change your mind, if you think that my arguments were solid enough.

Why is it so important to you to prove to a bunch of strangers on an internet message board that Zen will have massive perf/clock and clock like crazy?

Some of us are just giving our predictions, take them for what they're worth.

A technical forum is a place to argument. With numbers, reasoning, solid arguments.
There is an italian proverb that says (maybe exist also in english language) "math is not an opinion"
If one casually read this thread, he read two/three crazy elements that says Zen will be faster, with arguments, and a noisy majority with assert otherwise, often without argumenting. The mean reader can't distinguish this and probabily think that Zen will be a fiasco, with max 3-3.2GHz clock, and this it's not sure.

The silicon should be out soon, so we'll be able to put all of this back and forth to rest soon enough.

This is a correct agnostic sentence. Different from "Zen will not clock over 3.2GHz and will have an IPC lower than BDW. Stop." without arguments.

 

[bjt2]

Just so we're clear, a minimum of -50% drop in power would put Zen 8C at 100W for 4Ghz. Do you stand by this prediction?

Yes. This was ever my prediction, even in other forums. If i will be proven wrong, then i will not take it as tragedy (except for competition), but I genuinely think that it's possible, and gave my reasoning multiple times: equal or lower FO4, Zen about the transitors of an XV module, GF promises of -65% power, AMD statements of energy/cycle, early 14nm tests on NEON FPU, estimations based on apple A9 SoCs, etc...

 

[ShintaiDK]

Personally I really want to see the 15W TDP Mobile APUs and what performance they can achieve vs 14nm Intel 15W TDP Mobile SKUs. Because FinFets are like FD-SOI, they really shine at low power.

Also its about time we get a real 4-5W TDP APU for x86 11.6" Two in One Tablet/Laptop market, leaving Intel alone just killed that segment also. Having competition with CoreM will ignite that market with much better products stilling marketshare from 10-12" Android Tablets.

Looking at Polaris even when going low power, it was a very bleak experience. Nothing shining there.

So this would have to be very different. In other words, the uarch matters a lot!

 

[CentroX]

Next zen news coming in january according to sweclockers. They also think it might get delayed.

 

[jpiniero]

Next zen news coming in january according to sweclockers. They also think it might get delayed.

Another delay past February is fine but it better be released before Kaby/Skylake-X is. If it does get delayed again it would be because they are unhappy with the clocks.

 

[DrMrLordX]

AMD claims the performance benefit from core to core compare. Not module to core.

Go AMD marketing? There's no reason why we have to look at things the same exact way and claim that an A12-9800 (for example) is the equivalent of an i5-6500, which is a bit of a farce.

Today, even a single Haswell core with HT can match an XV module in total throughput in a LOT of benchmarks. Sometimes it exceeds XV. It's not like Intel cores are wanting for width, and Zen will apparently be the same way. So heck yeah I would compare an 8m/16t XV to an 8c/16t Broadwell-E, especially when I know that in many benchmarks the Broadwell-E would win!

Looking at Polaris even when going low power, it was a very bleak experience. Nothing shining there.

Go on posting that. Hopefully you can bring down the prices so Madpacket can afford another 6x 480 rig . . . heh

 

[AtenRa]

I don't follow your reasoning here. How does having an additional chip supplier impact demand for devices?

Competition

It will increase the variety of products currently on the market, it will lower prices and increase TAM.

 

[lolfail9001]

Believe is not the correct word. It's not faith. I gave you numbers, arguments, logical deductions.

If after that, YOU believe that Zen will have low IPC and/or clock max 3.2GHz, without argumenting, then this is you that is using faith.

You can believe that Zen will clock at maximum 3.2GHz, but i gave you solid arguments to think otherwise. It's up to you to change your mind, if you think that my arguments were solid enough.

The problem with your deductions is that you keep on ignoring the simple fact that Piledriver clocks way higher than SB-E on LN2, even though according to you it should be other way around. That alone should make you reconsider your argumentation. So, yes, it is a matter of faith, simply some like to justify their faith with what they call logic. At last, it is scientific sort of belief, as in few months we will learn who was right and who was square.

 

[bjt2]

The problem with your deductions is that you keep on ignoring the simple fact that Piledriver clocks way higher than SB-E on LN2, even though according to you it should be other way around. That alone should make you reconsider your argumentation. So, yes, it is a matter of faith, simply some like to justify their faith with what they call logic. At last, it is scientific sort of belief, as in few months we will learn who was right and who was square.

Where am I saying that? SB-E works at 3.2GHz. Piledriver has a version that works at 4.7-5GHz at default (even if at 220W) and over 4GHz (i don't remember precisely) at a lower 125W... On 32nm SOI...

A12 9800 in 65W, INCLUDING a >1GHz 512SP GPU has 4 XV cores at 3.8-4.2GHz on the 28nm BULK...

In 130W 8 XV cores PLUS 1024SP GPU can be made. With more clock of INTEL. On the 28nm BULK.
1024SP at >1GHz draw about 70-80W, so we can safely assume that in 130W we can put 16 XV cores, or 8 modules, that are moreless the same transistors of 8 Zen cores (+-10%).

If you think that in a process with 1/6 of leakage, -20% capacitance, 1.5x conductance and 20% less vcore, you can't do an 8 core CPU in 95W with over 3.2GHz, well, I am not trying anymore to convince you... I quit. BELIEVE (this is a correct term here) what you want...

We are not talking of a magic process... We know that the 14nm FF of INTEL it's better.

Here we have an architecture with a lower FO4. The first iteration, BD, has an awful IPC, we can safely say this. It's a sort of AMD's Pentium 4 (even if the reasons for the awful IPC are different).

But now Zen seems to have even more pipeline stages than BD and so the FO4 is at most equal to BD.

The process it's better.

How come that Zen should loss 25% of clock compared to BD/XV? Why? Give me one reason...

 

[jpiniero]

The process it's better.

How come that Zen should loss 25% of clock compared to BD/XV? Why? Give me one reason...

GloFo's process is a bastardization of Samsung's; which in the first place is a mobile-first process. I doubt Samsung ever intended it to reach 3.2 Ghz. Between that, AMD's cost cutting and the focus on Server, it's pretty easy to see why the clocks would not be so great.

 

[bjt2]

GloFo's process is a bastardization of Samsung's; which in the first place is a mobile-first process. I doubt Samsung ever intended it to reach 3.2 Ghz. Between that, AMD's cost cutting and the focus on Server, it's pretty easy to see why the clocks would not be so great.

A process that has less leakage, less capacitance, higher transconduttance, low vcore can't be worst than 28nm BULK.
Polaris has +15% clock and less power consumed. So 14nm FF it's better than 28nm BULK. Stop.
Now 19 vs 15 stages makes me think that the relative FO4 of Zen is not superior to XV's.
So I can't see why Zen should lose 25% of clock...

 

[lolfail9001]

Where am I saying that? SB-E works at 3.2GHz. Piledriver has a version that works at 4.7-5GHz at default (even if at 220W) and over 4GHz (i don't remember precisely) at a lower 125W... On 32nm SOI...

You are saying that Zen should clock at least as good because it has longer pipeline than Piledriver. Well, Sandy has longer pipeline than Con cores as well, where are my 4Ghz stock 2600ks with 7Ghz Cinebenches at?

A12 9800 in 65W, INCLUDING a >1GHz 512SP GPU has 4 XV cores at 3.8-4.2GHz on the 28nm BULK...

And actual power consumption and GPU clocks under load? Where may i see that?

If you think that in a process with 1/6 of leakage, -20% capacitance, 1.5x conductance and 20% less vcore, you can't do an 8 core CPU in 95W with over 3.2GHz, well, I am not trying anymore to convince you... I quit. BELIEVE (this is a correct term here) what you want...

I have seen all those cool process characteristics flop with Polaris 10 , when 20% less vcore was compensated by AMD overvolting by 35% instead. And that overvolt not even resulting in meaningful OC gains at first.

The process it's better.

Name a single high clocking device on GloFo's 14nm LPP. Get it yet? In fact, consider for a second that rumors of 1050 Ti have it made on Samsung's 14nm LPP. Lo and behold, the rumors also have it clocked quite low out of the box. Way too low, if we consider The Stilt's mention of 100W TDP and the fact that 1060@1500Mhz consumes about 60 Watts according to Tom's.

 

[leoneazzurro]

I have seen all those cool process characteristics flop with Polaris 10 , when 20% less vcore was compensated by AMD overvolting by 35% instead. And that overvolt not even resulting in meaningful OC gains at first..

Sorry, but this argument is invalid. The capability to go high in frequencies is not a function of the process only, and comparing Polaris with 28 nm GCN parts results in a much better Power/Perf ratio. Especially considering that the power/performance issues of GCN at 28nm were on TSMC process, when Nvidia did not encounter the same issues on the SAME technology...

 

[bjt2]

You are saying that Zen should clock at least as good because it has longer pipeline than Piledriver. Well, Sandy has longer pipeline than Con cores as well, where are my 4Ghz stock 2600ks with 7Ghz Cinebenches at?

I am not convinced that BD pipeline is 15 stages. No official informations are out. Another user here said that. I gave an interval of 15-20. Anyway Zen has 19, more than INTEL's that is about 17.
And anyway if INTEL isn't capable to do an high frequency design it's not my fault. AMD has demonstrated to be capable even on the awful 28nm BULK: 4.3GHz of Fmax and 4GHz base on a 95W XV part. With an HDL low power library. This is ever forgotten. And many assert also that it's the L2 the frequency limiting factor...

And actual power consumption and GPU clocks under load? Where may i see that?

Tonga with 2048SP at 1GHz has a TDP of 190W. These SPs have more than 1GHz. I don't think that 512 SPs at a greater frequency draw less than 35W.

I have seen all those cool process characteristics flop with Polaris 10 , when 20% less vcore was compensated by AMD overvolting by 35% instead. And that overvolt not even resulting in meaningful OC gains at first.

This was answered by @leoneazzurro.
I can add that OC is not what we are considering at this stage. But Polaris is the first chip. After 6 months we hope that there is some improvement. Anyway a GPU has a much higher FO4 and much less pipeline stages, so frequencies are not comparable.

Again Fiji tops at 1050 on the 28nm, polaris at 1266 on the 14nm. XV tops at 4.3GHz on the same 28nm BULK. Zen will top at what? Less than 3.2GHz? Come on...

Name a single high clocking device on GloFo's 14nm LPP. Get it yet? In fact, consider for a second that rumors of 1050 Ti have it made on Samsung's 14nm LPP. Lo and behold, the rumors also have it clocked quite low out of the box. Way too low, if we consider The Stilt's mention of 100W TDP and the fact that 1060@1500Mhz consumes about 60 Watts according to Tom's.

Again the FO4 of CPU and GPU is completely different. And the 1050 Ti is a consumer part, which can be harvested for high yield like polaris...

 

[lolfail9001]

Sorry, but this argument is invalid. The capability to go high in frequencies is not a function of the process only, and comparing Polaris with 28 nm GCN parts results in a much better Power/Perf ratio. Especially considering that the power/performance issues of GCN at 28nm were on TSMC process, when Nvidia did not encounter the same issues on the SAME technology...

Sure, power/perf is better, but clocks is what we're talking about.

I am not convinced that BD pipeline is 15 stages. No official informations are out. Another user here said that. I gave an interval of 15-20. Anyway Zen has 19, more than INTEL's that is about 17.
And anyway if INTEL isn't capable to do an high frequency design it's not my fault. AMD has demonstrated to be capable even on the awful 28nm BULK: 4.3GHz of Fmax and 4GHz base on a 95W XV part. With an HDL low power library. This is ever forgotten. And many assert also that it's the L2 the frequency limiting factor...

That's just what Con core design does: clock high.

I can add that OC is not what we are considering at this stage. But Polaris is the first chip. After 6 months we hope that there is some improvement. Anyway a GPU has a much higher FO4 and much less pipeline stages, so frequencies are not comparable.

Frequencies are not comparable, but rest of characteristics you mentioned are mostly are, especially since i extended comparison to GCN3.

Again the FO4 of CPU and GPU is completely different. And the 1050 Ti is a consumer part, which can be harvested for high yield like polaris...

This is all irrelevant, assuming 1050 Ti is actually a full GP107. If the rumors turn out true [as in ,1050 Ti will overclock worse by a significant margin than 1060 and it is built on 14nm LPP], then we pretty much have a solid confirmation that 14nm LPP is simply bad for high clocking parts, since we have all seen clocking ability of Pascal (namely the ease of hitting 2Ghz and difficulty of going further). If they don't, great, wait for Zen keeps on.