Dr Su will present Zen 2 and Navi at Computex 2019 (May 27th)

[mr moose]

7 minutes ago, Stefan Payne said:

No, because he could for example have used the Documents for Blackmail, to get a bit more money out of Intel.

In that case it wouldn't be Espionage, would it?

 

What he did was stealing documents and violated company policies.

But with those Documents, it seems there wasn't anything done.

He just had them. 

AMD worked with the FBI, they couldn't find any evidence that AMD knew about the Documents...

 

It's like I never made this post.

1 hour ago, mr moose said:

Given that many people in the industry have said it happens,  there are actual legal cases of it happening.

 

https://www.themarysue.com/intel-information-theft/

https://www.therichest.com/rich-list/10-of-the-most-infamous-cases-of-industrial-espionage/

 

Not to mention I said I witnessed it first hand.  But I am hardly going to throw ex employers under the bus.

 

So let me get this straight, If I am accusing Intel of espionage you agree but if I am accusing AMD of it then I am making false statements?  Typical.

 

Majority of what you said is just you not liking reality again.

 

 

Again, it happens, it has happened, there have been court cases, it will continue to happen.  Pretending it doesn't and trying to dismiss one case using personally assumed information doesn't make it less of a thing. 

[Stefan Payne]

1 minute ago, r2724r16 said:

As much as I'd like to believe all of this, I just don't think it's gonna happen. Especially at the rumoured pricepoints (~U$500 for the 16 core CPU).

Why??

It makes sense from a productpolitical point to do it like it was shown in the Reddit Leak:

https://www.eteknix.com/amd-ryzen-3000-pricing-leaks-look-great/

 

Also look at the Leaks closely:
The BASE Clock is 4.3GHz. That does sound plausible, does it?

The 5,1GHz is the Boost Clock.

 

Ryzen 7/2700x multicore Boost is somewhere around the 4GHz Area, Base is 3,7GHz, max Boost 4,3GHz

Ryzen 7/1700x multicore boost is somewhere around 3,5-3,6GHz or so. Base is 3,4GHz, max Boost is 3,8GHz

So that ~5GHz Boost is only SINGLE CORE boost.

 

And we are talking about a real shrink, something we haven't had in a long time within a socket/generation.

So that's comparable to 22nm Haswell -> 14nm Intel CPUs.

 

Haswell 1st gen 4770k had 3,5GHz Base, 3,9GHz Boost (4790k was 4,4GHz boost)

Kaby was 4GHz Base but only 4,2GHz Boost

Skylake was 4.2/4.5GHz

 

So 500MHz more Base and 300MHz Boost.

Or 15% more base but only +10% higher boost.

 

So lets assume that for AMD:

+15% base -> ~4,3GHz

+15% Boost -> ~5GHz

+10% Boost -> 4,73GHz

 

 

Do or do not 14nm Intel CPU reach 5GHz on some/many occasions?
Do or do not 22nm Intel CPU reach ~4,5GHz on some/many occasions?

 

And now why do you believe those increases will not happen, when AMD redesigned their cores and had a significant Shrink??

[Stefan Payne]

15 minutes ago, mr moose said:

It's like I never made this post.

 

Again, it happens, it has happened, there have been court cases, it will continue to happen.  Pretending it doesn't and trying to dismiss one case using personally assumed information doesn't make it less of a thing. 

Because it happened in some occasions, its proof that it happened with Ryzen, even though Intel didn't show any signs that it did happen?

Yeah, no. That's nonsense.

And what I see there:
Company secrets thrown into the trash/dumpster diving - sorry, but its YOUR job to responsibly get rid of your shit.

Many others might or might not have been shared voluntarily because they tried to work on something - that failed. Yeah, sucks.

 

The Case you shared about the ex-Intel Employee wasn't even proven Espionage. From the FBI Article, it looks like its only theft...

[r2724r16]

Just now, Stefan Payne said:

Why??

It makes sense from a productpolitical point to do it like it was shown in the Reddit Leak:

https://www.eteknix.com/amd-ryzen-3000-pricing-leaks-look-great/

 

Also look at the Leaks closely:
The BASE Clock is 4.3GHz. That does sound plausible, does it?

The 5,1GHz is the Boost Clock.

 

Ryzen 7/2700x multicore Boost is somewhere around the 4GHz Area, Base is 3,7GHz, max Boost 4,3GHz

Ryzen 7/1700x multicore boost is somewhere around 3,5-3,6GHz or so. Base is 3,4GHz, max Boost is 3,8GHz

So that ~5GHz Boost is only SINGLE CORE boost.

 

And we are talking about a real shrink, something we haven't had in a long time within a socket/generation.

So that's comparable to 22nm Haswell -> 14nm Intel CPUs.

 

Haswell 1st gen 4770k had 3,5GHz Base, 3,9GHz Boost (4790k was 4,4GHz boost)

Kaby Skylake was 4GHz Base but only 4,2GHz Boost

Skylake Kaby lake was 4.2/4.5GHz

 

So 500MHz more Base and 300MHz Boost.

Or 15% more base but only +10% higher boost.

 

So lets assume that for AMD:

+15% base -> ~4,3GHz

+15% Boost -> ~5GHz

+10% Boost -> 4,73GHz

I'm assuming you mixed Kabylake and Skylake...

 

The 4770K was released in mid-2013. The 7770K was released in early-2017. That's a gap of over 3 years, in which the boost clock increased by 0.6 GHz.

 

How do you expect AMD to increase their boost clock from 4.2/4.3 GHz to 5.1 GHz in just one year?

Just now, Stefan Payne said:

Do or do not 14nm Intel CPU reach 5GHz on some/many occasions?
Do or do not 22nm Intel CPU reach ~4,5GHz on some/many occasions?

Fair.

Just now, Stefan Payne said:

And now why do you believe those increases will not happen, when AMD redesigned their cores and had a significant Shrink??

Even if AMD is able to manufacture a 16 core CPU that runs up to 5.1 GHz, do you actually think they'll sell it for ~U$500?

[mr moose]

11 minutes ago, Stefan Payne said:

Because it happened in some occasions, its proof that it happened

 

 

That's all I am saying,  your need to go and create absolute specifics in order to have an argument is petulant at best.

11 minutes ago, Stefan Payne said:

with Ryzen, even though Intel didn't show any signs that it did happen?

Yeah, no. That's nonsense.

And what I see there:
Company secrets thrown into the trash/dumpster diving - sorry, but its YOUR job to responsibly get rid of your shit.

Many others might or might not have been shared voluntarily because they tried to work on something - that failed. Yeah, sucks.

 

The Case you shared about the ex-Intel Employee wasn't even proven Espionage. From the FBI Article, it looks like its only theft...

Corporate espionage happens whether you like it or not, and if you think AMD/Intel don't and haven't then you are naive.

[leadeater]

40 minutes ago, Stefan Payne said:

Scroll down, the "World's first 7nm GPUs" Part.

>1.25x higher performance at the same power.

Here something similar:

https://liliputing.com/2018/06/amds-first-7nm-gpu-coming-this-year-7nm-cpu-in-2019.html

 

max. Boost Clock of Ryzen 2700x is 4,3GHz. 

4,3*1,25 -> 5,375GHz

 

Even +15% higher clockrates than max Boost of Ryzen 2700X means 5GHz.

It's a bit more complicated than that, fab improvements aren't directly related to frequency increases as much as we'd hope. High performance at same power is a bit deceptive too because power increase at frequency steps isn't linear. Without architecture changes it's a bit easier to take a guess at frequency improvements but Zen2 is a lot different, on a different fab and on a different fab node technology altogether.

 

One of the more reliable, or more known, ways to increase the CPU frequency is more pipeline stages, AMD Piledriver for example on 32nm which could boost to 5Ghz.

 

I would err on the side of caution and hope for 5Ghz rather than expect it.

[Stefan Payne]

10 minutes ago, r2724r16 said:

I'm assuming you mixed Kabylake and Skylake...

You are right. Was planning with going with the younger one but then decided to go with the older one but forgot to change the names to the right ones...

 

10 minutes ago, r2724r16 said:

The 4770K was released in mid-2013. The 7770K was released in early-2017. That's a gap of over 3 years, in which the boost clock increased by 0.6 GHz.

That's not the important part.

The Important part is that the 4770K was an earlier 22nm Processor.

The 7700k was a later 14nm processor.

I think we should compare it with the 6700k because its an ealry 14nm and that already gave you a good boost of around 500MHz Core Clock.

10 minutes ago, r2724r16 said:

How do you expect AMD to increase their boost clock from 4.2/4.3 GHz to 5.1 GHz in just one year?

Its not the Time that's important.

Its the manufacturing process. THat's why I chose those because they were the last processors where that happened.

With the Jump from Sandy -> Ivy Bridge only the TDP sank a bit. But the Clockrate did not change.

I expect the 7nm Parts to clock higher than the 14nm and improved 12nm parts.

 

10 minutes ago, r2724r16 said:

Even if AMD is able to manufacture a 16 core CPU that runs up to 5.1 GHz, do you actually think they'll sell it for ~U$500?

They might.

That has two advantages:

a) agressive pricing -> more sold, so higher marketshare.

b) the Competition has to react eventually, wich also tanks their revenue due to lower sales prices.

 

But we don't know what Lisa and her team has planned, if they want to increase ASP or if they want to go for marketshare.

[Stefan Payne]

4 minutes ago, leadeater said:

It's a bit more complicated than that, fab improvements aren't directly related to frequency increases as much as we'd hope.

True, because we're already at the top end and can't shrink everything these days. Only a couple of thinks are smaller.

 

4 minutes ago, leadeater said:

High performance at same power is a bit deceptive too because power increase at frequency steps isn't linear.

With the same voltage it is more on the linear side but with higher clockrates we tend to increase the voltage, wich dramatically increases the Power Consumption.

I can't link you the Article from the Russians because their Site is offline
(Xbitlabs), who did a couple of years ago a test of Clockrates vs. OC. 

Though the CPUs used at the time were only 65nm and 45nm.

4 minutes ago, leadeater said:

Without architecture changes it's a bit easier to take a guess at frequency improvements but Zen2 is a lot different, on a different fab and on a different fab node technology altogether.

Problem is:
They did work on the core and changed somethings because they had to anyway because the libaries they used aren't compatible with the TSMC 7nm design rules.

The Problem is:
We do not know what AMD did change and what their design goals were.

4 minutes ago, leadeater said:

One of the more reliable, or more known, ways to increase the CPU frequency is more pipeline stages, AMD Piledriver for example on 32nm which could boost to 5Ghz.

 

I would err on the side of caution and hope for 5Ghz rather than expect it.

AMD has to increase the CPU Clockrates to really compete with Intel, so I doubt that they didn't adress this "Problem" of the Ryzen design.

The Question is how much did they do to increase the clockrates.

 

Remember, we aren't talking about high percentages, we are talking about 10-15% higher clockrates to get to the 5GHz and also not on all cores but "2 Core Boost".

 

In the End, I'd expect somewhere around +1GHz on either the Ryzen 7-1700 or 1700X.

Base is, as said above, 3,4GHz.

Normal Multicore Boost is around 3,5-3,6GHz

Maximum Boost I've seen is 3,9GHz for short periods/2 Core Loads.

 

Or around +10-15% on 2700x.

 

[hayden1881]

 

[leadeater]

8 hours ago, Stefan Payne said:

With the same voltage it is more on the linear side but with higher clockrates we tend to increase the voltage, wich dramatically increases the Power Consumption.

I can't link you the Article from the Russians because their Site is offline
(Xbitlabs), who did a couple of years ago a test of Clockrates vs. OC. 

Though the CPUs used at the time were only 65nm and 45nm.

You have to increase the voltage to increase the clocks, that's how (current) transistors work and a characteristic of them. The increase in voltage allows quicker transitions from off state to on state and also makes the state change easier to distinguish as there is a greater voltage difference between the states.

 

Not all transistors are the same though so depending on type and makeup those voltages and voltage differentials are different but the fundamentals of how it works is the same.

 

The transistors used in CPUs today are known as High-k type and use metal oxide Gate material instead of silicon oxide, as well as metal Gate electrode. Intel first introduced this in 45nm and is the primary reason why there was such a huge performance increase over the previous generation and also why there hasn't been a big leap like it again because we still use High-k type today.

 

As transistors get smaller you have to decrease the gate oxide thickness and that starts to bring in problems like leakage/tunneling which makes the gate less efficient. This also increases power consumption along with another issue of clearly knowing if the transistor is off or on, one of the ways to combat that is to increase the voltage but that also increases the leakage hence non linear power increase to increase frequency at the upper end of the transistor capabilities.

 

What's changing as we develop new fabrication technology is the metal oxides and the accuracy in which we can print the transistors, EUV is a big change to the latter.

 

So a test of clocks/frequency in overclocking a CPU for given voltages really only tells you about that fab node and transistor. Once you start changing gate oxide material, gate pitch, fin pitch etc the relationship between fab nodes starts to drift quite far. You can see that by comparing GloFo 14nm to Intel 14nm, how similar are those?

 

With the introduction of EUV and the increased accuracy/resolution that comes with it we could start to see more differences between products and SKUs. Like higher base clocks because of the increased transistor efficiency, or lower for more compelling low power options.

 

Chiplets is also another major factor because we are now talking about isolation of power delivery of the cores to other parts of the CPU and their transistors. We gain a lot by doing this, heat reduction, heat source size reduction, true isolation. There's other potential gains like being able to increase the core voltage with less potential to effect other areas and transistors in the CPU, we have some control of that today already but there is the possibility for that to be further improved. This is also carried across to multi-chiplet packages where you (should?) have individual voltage regulators for each chiplet meaning greater control and less voltage drop under high power conditions.

 

The most interesting and important area for Zen2 isn't actually Ryzen it's Threadripper and EPYC. Either way I'd still keep a fair amount of caution in frequency expectation for Zen2/Ryzen 3000 because it's 7nm, EUV which is very new with unknown challenges and characteristics relating to the first part of this post. 

[GoldenLag]

7 hours ago, r2724r16 said:

Even if AMD is able to manufacture a 16 core CPU that runs up to 5.1 GHz, do you actually think they'll sell it for ~U$500?

Why wouldnt they?

 

Okay id say 500$ is a bit optimistic. Say 550 or 600$. Though they can do 500$ aswell.

 

I do not see a good reason why they wouldnt be selling it at the previous pricepoint of the 1800x. 

[GoldenLag]

Just now, VegetableStu said:

Highly doubt they'd price a 16-core like an 8-core ._. my guess would be 600-750-ish

(I'm giving myself some emotional headroom there, LOL)

I dont see them break the 600$ mark. It would be too much of a HEFT price. They will probably drop the price of the 2950x to about 100-250$ more than the new 7nm part.

 

While AMD will be uncontested, they still want marketshare and reqognition. Going far above the price of 9900k wouldnt do them any good.

[GoldenLag]

1 minute ago, VegetableStu said:

I were to make a parametric guess, I'd say it'll be about $150 less than a 16-core Threadripper. the 2950X if TR 3000 doesn't start at 16 cores

Its a propper HEDT plattform afterall. It sort of needs to cost a bit extra.

1 minute ago, VegetableStu said:

AMD markets this as a small-end HEDT processor, I'd think they might have the "leverage" to price it higher o_o 16 cores is unheard of in regular desktop

 

then again we had 6-core and 8-core Ryzen in the first place, but I don't think anyone on a current 8-core system would really utilise 16 cores, unless they sincerely do rendering work on their systems. Pushing it as a low-cost HEDT would make the case for the 16-core part on AM4

AMD want Intel to Join in the game of cores. Something they have managed to do with the 9900k. AMD cant be beaten in a game of cores, and if 16 cores is the future, then they are set (because honestly i have a hard time imagining more than 16 cores being any useful)

[leadeater]

23 minutes ago, VegetableStu said:

if AMD markets this as a small-end HEDT processor, I'd think they might have the "leverage" to price it higher o_o 16 cores is unheard of in regular desktop

 

then again we had 6-core and 8-core Ryzen when they first came out, but I don't think anyone on a current 8-core system would really utilise 16 cores, unless they sincerely do rendering work on their systems. Pushing it as a low-cost HEDT would make the case for the 16-core part on AM4

 

if I were to make a parametric guess, I'd say it'll be about $150 less than a 16-core Threadripper. the 2950X if TR 3000 doesn't start at 16 cores.

There's two routes AMD could do with Ryzen 3000, launch with 2 chiplet product or only 1.

 

I can see a decent delay happening to make it easier to introduce new products and phase out old like low end TR plus give time for motherboards to up spec, plus will give an increased sense of longevity to existing platform owners liked promised though without specific commitment that literally everything would be supported. I can only see this happening if Zen2 performance increase is great enough it's better than a 9900k.

 

The other path is a bit more impactful, probably better for the brand too. Just lay everything out all at once, Ryzen and TR, and just home run hit. There might be a little bit of resentment for a small section of people, mainly TR buyers within the last year. There is one product(s) I think will be exempt from launch day and that's TR above 32 cores, I think they'll save that for later.

[Mihle]

I think there was a rumour that AMD changed their plans and that 16 core zen 2 won't be out the same time as the less core ones?

[porina]

8 hours ago, Stefan Payne said:

With the same voltage it is more on the linear side but with higher clockrates we tend to increase the voltage, wich dramatically increases the Power Consumption.

The behaviour has been well understood by overclockers. If you run a CPU to its limit in a given condition, then generally more clock needs more voltage, and it isn't a linear relationship. In particular, there is the dreaded voltage wall, where for small gains in clock you have to use a lot more voltage. To try to overcome that people resort to extreme cooling like LN2. FWIW in my limited attempts with chilled water (close to 0C), it gets me about 100 MHz more than ambient water (on Intel), hardly worth the effort.

 

This is in part why we generally see lower clocks on higher core count CPUs. We can run them in a more efficient part of their clock range to keep the total power consumption down. You still gain from having more cores. That's not to say you can't hit similar clocks on higher core count parts, if you can throw enough power into it, and deal with the resulting heat. A stock 1700 might be rated 65W TDP, but in my attempts to push it to 4 GHz and beyond I hit a CPU reported 180W before reaching the thermal limits of a 280mm AIO and had to stop. I think stock ran at 3.2 GHz all cores and was super cool. So nearly 3x power consumption for 25% more clock isn't a great tradeoff.

 

8 hours ago, Stefan Payne said:

AMD has to increase the CPU Clockrates to really compete with Intel, so I doubt that they didn't adress this "Problem" of the Ryzen design.

The Question is how much did they do to increase the clockrates.

AMD don't necessarily have to increase clock. They've already gone way down the "more cores" road and are marketing it well to enthusiasts and power users who might be able to use it (as well as some enterprise uses of less interest here). That said, it doesn't mean I don't want more clock from Ryzen, but it isn't necessarily the first priority. They also have opportunity to improve performance in consumer use cases by other methods, for example, from cache and IMC improvements. Zen 2 will be different from what we know and it remains to be seen what happens there.

 

It wasn't recent, but Intel had previously mentioned they may design less peak performing CPUs in order to make greater gains elsewhere like in power efficiency. Now that I say it, I wonder what their mix of CPU sales are between laptops and desktops. The laptop market as a whole is supposed to be much bigger, but how does that translate into CPU sales? Point is, they may optimise more for power than performance if that's the biggest part of the market. Desktop CPUs do have the side benefit of not being power constrained so they can still push them into a high performance lower efficiency operating region.

 

8 hours ago, Stefan Payne said:

Remember, we aren't talking about high percentages, we are talking about 10-15% higher clockrates to get to the 5GHz and also not on all cores but "2 Core Boost".

And then overclockers will make all cores do that  

[cj09beira]

9 hours ago, r2724r16 said:

I'm assuming you mixed Kabylake and Skylake...

 

The 4770K was released in mid-2013. The 7770K was released in early-2017. That's a gap of over 3 years, in which the boost clock increased by 0.6 GHz.

 

How do you expect AMD to increase their boost clock from 4.2/4.3 GHz to 5.1 GHz in just one year?

Fair.

Even if AMD is able to manufacture a 16 core CPU that runs up to 5.1 GHz, do you actually think they'll sell it for ~U$500?

we dont actually have to imagine how big a jump 7nm is in terms of base and boost clock, we just have to look at what amd has achieved with radeon VII, stock performance (similar to all core boost) went from 1500 to 1800 (a little bit lower than that on both cases mostly because of power throttling, though that is not important here), max frequency went from 1750, to 2050+ (radeon VII being a low end bin has very high variability in the silicon so people have got everything from 1900 ish mhz to 2200 mhz),

this is going from 14nm to 7nm, applying the same improvement to zen it would be 4.3ghz base (how interesting the exact same as the leak's base) and 4.8ghz max clocks, with the better binning of chiplets going even further is possible, but 4.8 is a good baseline for what to expect

[leadeater]

37 minutes ago, porina said:

The laptop market as a whole is supposed to be much bigger, but how does that translate into CPU sales? Point is, they may optimise more for power than performance if that's the biggest part of the market. Desktop CPUs do have the side benefit of not being power constrained so they can still push them into a high performance lower efficiency operating region.

In a way laptop usage aligns closer with the server market than desktop does as well, performance efficiency comes first. If Intel biases away their architecture and fab design from performance to power we might see a fair reduction in overclocking, very much doubt any performance reduction though because that would be far too hard to sell.

[Stefan Payne]

7 hours ago, leadeater said:

You have to increase the voltage to increase the clocks, that's how (current) transistors work and a characteristic of them.

Yes but the power consumption rises/falls linear with the frequency.

Resistive loads increase the power with the power of 2 -> Double Voltage = double the Current -> 4 times the power consumption.

 

With Transistors it might be even worse than that for various reasons.

 

 

So with an increase in frequency of 20% at the same voltage, the Power Consumption also rises by around 20%.

With Voltage it is different, obviously.

7 hours ago, leadeater said:

The increase in voltage allows quicker transitions from off state to on state and also makes the state change easier to distinguish as there is a greater voltage difference between the states.

Yeah also there is a minimal voltage for semiconductors that they need to work. ANd we are darn close to those voltages.

 

7 hours ago, leadeater said:

As transistors get smaller you have to decrease the gate oxide thickness and that starts to bring in problems like leakage/tunneling which makes the gate less efficient.

Yeah and you can't necessarily shrink the transistors these days, as seen with the PHY or other stuff that needs to drive signals over the Motherboard.

In that case you have a current requirement for that circuit that prevents it from beeing shrunk with the processes wich is why AMD threw out the Northbridge/Chipset part of the CPU and put it into a new Die because the size difference between 14nm and 7nm is probably rather negligable (maybe 10%, maybe 15% smaller)

 

7 hours ago, leadeater said:

This also increases power consumption along with another issue of clearly knowing if the transistor is off or on, one of the ways to combat that is to increase the voltage but that also increases the leakage hence non linear power increase to increase frequency at the upper end of the transistor capabilities.

Leakage was one of the things that killed the Pentium 4 back in the day and that was with 90nm.

Its also one of the main problems with modern chipdesign.

 

7 hours ago, leadeater said:

What's changing as we develop new fabrication technology is the metal oxides and the accuracy in which we can print the transistors, EUV is a big change to the latter.

Still silicon on Insulator is dead for whatever reasons.

AMD Used that the last time with the Original Bulldozer in 32nm.

After that the development of SOI regressed.

 

Its said that that helps with some of the Problems of modern CPU Design.

7 hours ago, leadeater said:

You can see that by comparing GloFo 14nm to Intel 14nm, how similar are those?

Ähm, how do we compare the two?
We need a design on both fabs - wich we don't have.

But what we have is a Design on Samsung 14nm and TSMC 16nm. More or less the same Chip. And that Chip showed that TSMC 16nm had a lower power consumption as the Samsung process.

 

Yes, it was an Apple CPU. The A9 to be precise wich was dual sourced on Samsung and TSMC. Though the question that remains:
How well was it optimized on either process??

 

Anyway, that's the only example for us to compare foundrys.

 

7 hours ago, leadeater said:

With the introduction of EUV and the increased accuracy/resolution that comes with it we could start to see more differences between products and SKUs. Like higher base clocks because of the increased transistor efficiency, or lower for more compelling low power options.

Yes and that is the unknown of the Zen2 die.

Is it designed/optimized for highest power efficiency at lower clocks or is it optimized for high frequency?
Or somewhere in between?

 

That is the big unknown right now. But its possible that AMD's design goal for Zen2 was to increase the frequency by 15%. That's not too much increase with an optimized design and a shrink with changing the fab as well.

GF had licensed the Samsung Process, and we already have a comparisation between Samsung and TSMC.

And in this case there was a ~5°C Difference in a mobile device outer shell according to the Article I've found....

 

7 hours ago, leadeater said:

Chiplets is also another major factor because we are now talking about isolation of power delivery of the cores to other parts of the CPU and their transistors. We gain a lot by doing this, heat reduction, heat source size reduction, true isolation. There's other potential gains like being able to increase the core voltage with less potential to effect other areas and transistors in the CPU, we have some control of that today already but there is the possibility for that to be further improved. This is also carried across to multi-chiplet packages where you (should?) have individual voltage regulators for each chiplet meaning greater control and less voltage drop under high power conditions.

I agree with you.

But the Chiplets is a necessity du to physical limitations of lower manufacturing processes. And also the financial ones.

It just makes more sense to throw anything out of the 7nm die for various reasons.

The things you mentioned here is one of the reasons to do it, another is cost. Its just cheaper to do it because according to bits and chips the yield of the 7nm process with AMD is at only around 70%.

 

7 hours ago, leadeater said:

The most interesting and important area for Zen2 isn't actually Ryzen it's Threadripper and EPYC. Either way I'd still keep a fair amount of caution in frequency expectation for Zen2/Ryzen 3000 because it's 7nm, EUV which is very new with unknown challenges and characteristics relating to the first part of this post. 

Yeah for EPYC the frequency isn't important as for the consoles.

So that is two out of 4 applications where the frequency is of no interest of the design. Even a 4GHz max. clock would be more than enough for those situations.

 

 

However it is different on the Ryzen products, wich include Threadripper.

To compete with Intel they need high frequencys of at least close to 5GHz especially if they want to have some headroom for the upcoming 10nm Intel CPUs (though AMD could at that time switch to 6nm)

 

[leadeater]

54 minutes ago, Stefan Payne said:

Yes but the power consumption rises/falls linear with the frequency.

No it doesn't, but it's hard to generalize like this at all because transistor type and makeup changes this too.

 

Rough, not CPU transistor example:

 

Also:

Quote

As switching frequencies increase, it becomes of paramount importance to reduce the switching losses in the converter. These are the losses associated with the transition of the switch from its on-state to its off-state, and back. The higher the switching frequency, the greater the number of times the switch changes state per second. Therefore, these losses are proportional to the switching frequency. Further, of these frequency-dependent loss terms, the most significant are usually those that take place within the switch itself. Therefore, understanding the underlying sequence of events in the switch during each transition, and thereby quantifying the losses associated with each of these events, has become a key expectation of any power supply designer.

And that's before bringing in to the mix electron tunneling which for CPUs is a major factor and is one of the big challenges in node shrinks.

[leadeater]

23 minutes ago, Stefan Payne said:

I agree with you.

But the Chiplets is a necessity du to physical limitations of lower manufacturing processes. And also the financial ones.

It just makes more sense to throw anything out of the 7nm die for various reasons.

The things you mentioned here is one of the reasons to do it, another is cost. Its just cheaper to do it because according to bits and chips the yield of the 7nm process with AMD is at only around 70%.

I'm not sure why but I think you're taking what I said as counter points or reasons not to do it? It should have been obvious those were all positives, reasons to do, but I was just more talking about the situation itself.

[Stefan Payne]

52 minutes ago, leadeater said:

No it doesn't, but it's hard to generalize like this at all because transistor type and makeup changes this too.

 

Rough, not CPU transistor example:

Ähm, isn't that a logarythmic diagram??

 

 

52 minutes ago, leadeater said:

Just like to correct this because it's something you shouldn't carry forward, it's not correct. Doubling the voltage doesn't double the current, you can change the voltage and the current doesn't have to change at all but power must, or current has to drop or resistance increase.

It is correct as I was talking about Resistive Loads.

Aka a Resistor.

 

IF you double the Voltage on the Resistor, you double the current as well.

You can also calculate it.

Just take a 100 Ohm Resistor.

 

Voltage = Resistance * Current

 

Voltage

------------ =  Current

Resistance 

 

So for a 100 Ohm Resistor, at 12V we are talking about 0,12A or 1.44W

So for a 100 Ohm Resistor at 6V we are talking about 0,06A or 0,36W

 

I know that with silicon semiconductors its not as easy and that there are also other effects that influence it for example the thickness of the insulation layer.

But in General, those aren't Ohm'sch Resistors, they are mostly capacitors with a bit of inductors so its not as cut and dry as with a Resistor.

[leadeater]

13 minutes ago, Stefan Payne said:

It is correct as I was talking about Resistive Loads.

Aka a Resistor.

Yea sorry I removed it just before you posted this as the situations I was going to reference honestly didn't matter at all, nor that related to what you were saying.

 

13 minutes ago, Stefan Payne said:

Ähm, isn't that a logarythmic diagram??

No, exponential. Point is you said it's linear and it's not, and that's only one factor for overall power for a CPU with billions of transistors. If switching loss is not linear then it's impossible for power draw for a CPU in relation to frequency to be linear.

[Stefan Payne]

40 minutes ago, leadeater said:

I'm not sure why but I think you're taking what I said as counter points or reasons not to do it? It should have been obvious those were all positives, reasons to do, but I was just more talking about the situation itself.

No, absoluitely not. I just added some additional arguments to go with the chiplet/MCM Approach instead of the Monolithic die.

 

I was just talking about a different side of the chiplet design, that allows other positive effects. Though I was more arguing on the financial aspect and the size of the I/O Die.

We both agree that the chiplet design makes the most sense with 7nm, just for different reasons.

 

I didn't  really think about the Power Consumption and design aspect.

That are very interesting points you made there.

 

 

 

[leadeater]

8 minutes ago, Stefan Payne said:

I didn't  really think about the Power Consumption and design aspect.

That are very interesting points you made there.

You may have not made the connection but why I said TR and EPYC are the more interesting ones is because I think with the chiplet design and potential power/voltage controls I think we might see the frequency penalties of high core CPUs all but disappear. Think about a situation on a 64 core EPYC where only 8 cores are active and the software/scheduler is properly optimized so only 1 chiplet is active/in high power state, is there any reason why the clocks would have to be any lower than on say a single chiplet Ryzen 3000? I'd say no they should be able to be the same, which is awesome if possible.