Discussion Intel current and future Lakes & Rapids thread

[dmens]

Tigerlake and 10nm SF greatly improves V/F curves:

-Tigerlake's Willow Cove can operate at 2.6GHz in the same voltage as Sunny Cove at 1.6GHz.
-3.6GHz in the same voltage as SNC at 2.6GHz.
-4.6GHz in the same voltage as SNC at 3.9GHz.
-Finally, it can go even higher with higher voltages.

Sunny Cove at 15W runs at 1.8-2GHz with MT workloads and 2.4-2.6GHz at 25W.

Gonna call total BS on those numbers.

 

[btarlinian]

Gonna call total BS on those numbers.

It's more like 2.5 GHz on SNC requires the same voltage as ~3.3-3.4 GHz on Willow Cove.

Slide 63 in this deck is pretty close to hard numbers. (https://newsroom.intel.com/wp-conte...Architecture-Day-2020-Presentation-Slides.pdf)

 

[IntelUser2000]

They just need to cut that IGP, leave 32 EUs and go 8 cores for desktop, overclockers will make it reach 5 GHz all core regardless if some mobile chip can do 4.7. At that point who needs Rocket lake, honestly?

I want to see a literal doubled up Tigerlake with 8 cores, and 192EUs, along with single 4GB HBM2 stack.

 

[TESKATLIPOKA]

My biggest problem with Tiger Lake is the low core count. For 1-4 threads It will be the best, but the more threads you will use the difference between Tiger Lake and Renoir will be smaller and smaller untill Renoir will be faster. I think in Cinebench MT Tiger Lake won't tie with Renoir 4700U, but lag noticeably behind in performance, because ~25% better IPC is not enough to erase 50% difference in performance. If It had at least 6 cores It would be much better in my opinion.

 

[dmens]

It's more like 2.5 GHz on SNC requires the same voltage as ~3.3-3.4 GHz on Willow Cove.

Slide 63 in this deck is pretty close to hard numbers. (https://newsroom.intel.com/wp-conte...Architecture-Day-2020-Presentation-Slides.pdf)

Where do these numbers come from? The plot on slide 63 does not even have any numbers on the voltage axis.

It's more like 2.5 GHz on SNC requires the same voltage as ~3.3-3.4 GHz on Willow Cove.

I can guarantee you that will not happen.

 

[dullard]

Where do these numbers come from? The plot on slide 63 does not even have any numbers on the voltage axis.

They come from Intel. Why do you need numbers on the X-axis? For scientific interest, I would like the numbers. But, to answer the question at hand, you do not need the numbers. Look at the voltage where Sunny Cove is 2.5 GHz. Then draw a line straight up (at whatever voltage that is). That hits about 3.3 GHz to 3.4 GHz on Willow Cove.

I can guarantee you that will not happen.

Where does your guarantee come from?

 

[Markfw]

They come from Intel. Why do you need numbers on the X-axis? For scientific interest, I would like the numbers. But, to answer the question at hand, you do not need the numbers. Look at the voltage where Sunny Cove is 2.5 GHz. Then draw a line straight up (at whatever voltage that is). That hits about 3.3 GHz to 3.4 GHz on Willow Cove.

Where does your guarantee come from?

With Intel's track record of lying about everything lately, I believe an ex-Intel employee more than them.

 

[dmens]

They come from Intel. Why do you need numbers on the X-axis? For scientific interest, I would like the numbers. But, to answer the question at hand, you do not need the numbers. Look at the voltage where Sunny Cove is 2.5 GHz. Then draw a line straight up (at whatever voltage that is). That hits about 3.3 GHz to 3.4 GHz on Willow Cove.

Oh I dunno, maybe because the x-axis might not be a linear scale. Little details like that. I know a marketing slide when I see one.

Where does your guarantee come from?

15 years in CPU design at Intel. Besides that, anyone with a passing knowledge of digital design reality would know those numbers don't pass the smell test.

 

[dullard]

With Intel's track record of lying about everything lately, I believe an ex-Intel employee more than them.

They've missed deadlines--terribly. But "lying about everything" is certainly more incorrect than even Intel is. Intel's marketing has been pretty accurate (other than missed deadlines). For example, Intel said their first 10 nm would be poor performing, and it was.

 

[dullard]

Oh I dunno, maybe because the x-axis might not be a linear scale. Little details like that. I know a marketing slide when I see one.

Yes, it would be nice to have the x-axis numbers. Yes, it is a marketing slide. But, if you were an engineer (or similar occupation), how come you don't realize that x-axis values are not needed to answer this question from the slide? See below were a few red lines were drawn to get the answer, all without any x-axis values:

 

[dmens]

Yes, it would be nice to have the x-axis numbers. Yes, it is a marketing slide. But, if you were an engineer (or similar occupation), how come you don't realize that x-axis values are not needed to answer this question from the slide? See below were a few red lines were drawn to get the answer, all without any x-axis values:
View attachment 27985

The lack of X-axis numbers renders left-right arrow means your left-right arrow is absolutely meaningless for comparative purposes.

You can try to draw conclusions from the vertical axis since it is supposedly linearly scaled, and that is where this slide utterly fails to sniff test: if the claim is they managed to achieve a general 40% increase in frequency on the same voltage (just ignore power for now) on a typical, average sampled part from two very similar core architectures, we are talking about gains that used to take multiple process generations to achieve. That is not happening, unless Intel completely abandoned low power operational characteristics, which while possible, won't cheat third-party power benchmarks no matter how much they re-define TDP, PL* windows, whatever. You can add on top the troubled development history of the 10nm process, but really that is not even necessary to render these claims odorously offensive.

My professional opinion, this is borderline fraudulent marketing. My take is that they can avoid legal liability by simply claiming they are not comparing average parts.

 

[itsmydamnation]

Yes, it would be nice to have the x-axis numbers. Yes, it is a marketing slide. But, if you were an engineer (or similar occupation), how come you don't realize that x-axis values are not needed to answer this question from the slide? See below were a few red lines were drawn to get the answer, all without any x-axis values:
View attachment 27985

I hope you realise the irony of your own post

 

[Markfw]

They've missed deadlines--terribly. But "lying about everything" is certainly more incorrect than even Intel is. Intel's marketing has been pretty accurate (other than missed deadlines). For example, Intel said their first 10 nm would be poor performing, and it was.

So they deceived the stock holders so much that they are getting sued, but they have not been lying ? You don't get held legally liable for missing deadlines, just lying about them.

 

[jpiniero]

You can add on top the troubled development history of the 10nm process, but really that is not even necessary to render these claims odorously offensive.

My guess is that it's the trash tier Icelake versus the best binned Tiger Lake. Did seem like Intel was rather generous on voltage on Icelake too, just in general.

 

[dullard]

I hope you realise the irony of your own post

No, I do not. Could you please explain how x-axis values are needed to calculate something that only depends on y-axis values? This is science 101. Draw a straight line up. That is all you need.

 

[dullard]

The lack of X-axis numbers renders left-right arrow means your left-right arrow is absolutely meaningless for comparative purposes.

Clearly, you have no clue how to read this graph. Let me help you. The x-axis value could be 1 volt, 1 uV, or 1 kV. The x-axis could be linear, quadratic, log, or any other random scale. It doesn't matter when the only dimension we are talking about is the y-axis direction.

 

[dmens]

Clearly, you have no clue how to read this graph. Let me help you. The x-axis value could be 1 volt, 1 uV, or 1 kV. The x-axis could be linear, quadratic, log, or any other random scale. It doesn't matter when the only dimension we are talking about is the y-axis direction.
View attachment 27986

Did you even read my post? The y-axis stinks, the x-axis is not labeled. I read the graph and I am calling BS.

If you don't understand what I said about transistor characteristics, that is fine, it is not for laypersons. I don't care if you believe me.

 

[dullard]

Did you even read my post? The y-axis stinks, the x-axis is not labeled. If you don't understand what I said about transistor characteristics, that is fine, it is not for laypersons. I don't care if you believe me.

What stinks about the y-axis: 0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5? Really back it up. Give me a calculation that shows these intervals stink and are not reliable. I'll wait. GHz labels would make a better graph. Voltage values and labels would make a better graph. But neither are needed for this topic of discussion.

And please, thinly veiled insults are not needed on this forum. I understand transistors quite fine. I don't understand your lack of ability to read this graph though.

 

[btarlinian]

The lack of X-axis numbers renders left-right arrow means your left-right arrow is absolutely meaningless for comparative purposes.

You can try to draw conclusions from the vertical axis since it is supposedly linearly scaled, and that is where this slide utterly fails to sniff test: if the claim is they managed to achieve a general 40% increase in frequency on the same voltage (just ignore power for now) on a typical, average sampled part from two very similar core architectures, we are talking about gains that used to take multiple process generations to achieve. That is not happening, unless Intel completely abandoned low power operational characteristics, which while possible, won't cheat third-party power benchmarks no matter how much they re-define TDP, PL* windows, whatever. You can add on top the troubled development history of the 10nm process, but really that is not even necessary to render these claims odorously offensive.

My professional opinion, this is borderline fraudulent marketing. My take is that they can avoid legal liability by simply claiming they are not comparing average parts.

(The left-right arrow in that graph has nothing to do with comparing frequency at same voltage, that was in the original slide.) I don't really understand why this is so implausible. 20% drive current improvements have been seen before. If we assume that the 10 nm yield problems are patterning/interconnect related and the people who work on transistor technology were not just sitting around twiddling their thumbs for several years, this much improvement doesn't really seem that implausible. Add in the capacitance reduction from the looser gate pitch in high performance libraries and the other described improvements, I don't really see why this is so implausible.

I mean if the argument is that the 10 nm process used in ICL is a bad baseline, that's at least reasonable. But there we've seen base clocks on Tigerlake are significantly higher than Icelake, 1165G7 at 2.8 GHz vs. 1.3 GHz on 1065G7 (even if we take the cTDP up base frequency on the 1065G7 we get 1.5GHz -> 2.8 GHz, which is more than the improvement shown in this graph. I'm assuming the rest comes from headroom gained from power reduction in always on components).

 

[dmens]

What stinks about the y-axis: 0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5? Really back it up. Give me a calculation that shows it stinks and is not reliable.

And please, thinly veiled insults are not needed on this forum. I understand transistors quite fine. I don't understand your lack of ability to read this graph though.

Here, I will summarize it: claiming a 40% delay improvement at the same supply across a large voltage range on the average sampled part stinks to high heaven. Because while it is technically possible, it can only be achieved with a crippling concession on some other metric.

 

[dullard]

So they deceived the stock holders so much that they are getting sued, but they have not been lying ? You don't get held legally liable for missing deadlines, just lying about them.

So you are saying that a court found to be legally liable for lying lately on marketing documents? Please link. Or, are you jumping to the conclusions that you want to jump to?

 

[dullard]

Here, I will summarize it: claiming a 40% delay improvement at the same supply across a large voltage range on the average sampled part stinks to high heaven. Because while it is technically possible, it can only be achieved with a crippling concession on some other metric.

So, you can't back up your claim that the y-axis is "where this slide utterly fails to sniff test "?

Sunny Cove is crap on frequencies (what we are talking about). Not unusable crap, but crap. 40% better than crap is not impossible.

 

[dmens]

I don't really understand why this is so implausible. 20% drive current improvements have been seen before.

On-current does not scale linearly with digital gate delay. Here's a more common way to increase current: using a bigger cell. Ever wonder why power is hardly mentioned in these slides? There's your answer.

 

[dmens]

So, you can't back up your claim that the y-axis is "where this slide utterly fails to sniff test "?

Sunny Cove is crap. Not unusable crap, but crap. 40% better than crap is not impossible.

It is called engineering reality. Like I said, I don't care if you believe me. I am on a first name basis with many of the people who worked on this project and they would be embarrassed by this trash marketing.

 

[Hitman928]

So, you can't back up your claim that the y-axis is "where this slide utterly fails to sniff test "?

Sunny Cove is crap on frequencies (what we are talking about). Not unusable crap, but crap. 40% better than crap is not impossible.

What units is the y-axis in?