Discussion Intel current and future Lakes & Rapids thread

[dullard]

Plug in any voltage you want for Sunny Cove at 3.9 GHz and then use your graph to estimate voltage for Willow Cove at 2.5 GHz and 2 GHz and let me know how realistic it seems.

Sure, as soon as you tell me the Sunny Cove voltage at 2.5 GHz and 2.0 GHz (because I honestly do not know those numbers and want this to be a valuable exercise).

 

[Hitman928]

Sure, as soon as you tell me the Sunny Cove voltage at 2.5 GHz and 2.0 GHz (because I honestly do not know those numbers and want this to be a valuable exercise).

You really don't need those values, if you set the 3.9 GHz voltage to 1.1 V (or really anywhere you want that's even somewhat realistic) you can get a very close estimate (based on the chart) for those values. You can go through it if you want, but you'll see that if you use your chart, the Willow Cove voltage at 2 GHz is something around like 0.25V which isn't even close to realistic.

 

[Ajay]

As an engineer, I thought you would be evidence based. What is your evidence that this data is wrong? Or are you just going on your anti-Intel emotions?

I'm NOT anti-Intel; I would like to see them execute better - but they are not. What I was pointing out, is that this slide is NOT an engineering slide, but a marketing slide.
From personal experience (and years of seeing this in the semiconductor biz), I do not trust marketing departments to produce graphs meeting rigorous engineering standards.
What engineer leaves off the X-Axis; none that I've met.

 

[dullard]

You really don't need those values, if you set the 3.9 GHz voltage to 1.1 V (or really anywhere you want that's even somewhat realistic) you can get a very close estimate (based on the chart) for those values. You can go through it if you want, but you'll see that if you use your chart, the Willow Cove voltage at 2 GHz is something around like 0.25V which isn't even close to realistic.

Yes, I need them. Because I need to know if Sunny Cove voltages are realistic for this discussion of Willow Cove to be of any value. It is already of limited value because voltages don't matter to me. Performance and somewhat power is what matters. Is the Sunny Cove 0.25 V frequency even close to realistic?

 

[dullard]

I'm NOT anti-Intel; I would like to see them execute better - but they are not. What I was pointing out, is that this slide is NOT an engineering slide, but a marketing slide.
From personal experience (and years of seeing this in the semiconductor biz), I do not trust marketing departments to produce graphs meeting rigorous engineering standards.
What engineer leaves off the X-Axis; none that I've met.

This still dances around the question. What specifically is incorrect in the graph? With this information, we can help the SEC form a lawsuit against Intel for misleading investors.

I agree that I would never leave off units and numbers on graphs that I put out. But, your insistence that the graph cannot be used needs some shred of actual evidence for your posts to matter on the subject. Otherwise your leaving out information about what is incorrect is just as bad as the marketing people leaving out the labels.

The graph says that Willow Cove can reach 3.25 GHz at the same voltage that Sunny Cove uses to reach 2.5 GHz. That is a 30% frequency improvement. Is that incorrect? How much?

The graph says that Willow Cove can reach 3.7 GHz at the same voltage that Sunny Cove uses to reach 3.0 GHz. That is a 23% frequency improvement. Is that incorrect? How much?

The graph says that Willow Cove can reach 4.2 GHz at the same voltage that Sunny Cove uses to reach 3.5 GHz. That is a 20% frequency improvement. Is that incorrect? How much?

The graph says that Willow Cove can reach 4.7 GHz at the same voltage that Sunny Cove uses to reach 3.9 GHz. That is a 21% frequency improvement. Is that incorrect? How much?

 

[Hitman928]

Yes, I need them. Because I need to know if Sunny Cove voltages are realistic for this discussion of Willow Cove to be of any value. It is already of limited value because voltages don't matter to me. Performance and somewhat power is what matters. Is the Sunny Cove 0.25 V frequency even close to realistic?

No, the Sunny Cove frequency at 0.25V is not even close. The cores would be off at that voltage.

Here's a sample polling of Tigerlake Vid and frequency at idle from someone at the techpowerup forums. I won't post the full table but you can go to the link if you want to see it.

 

[Hitman928]

This still dances around the question. What specifically is incorrect in the graph? With this information, we can help the SEC form a lawsuit against Intel for misleading investors.

I agree that I would never leave off units and numbers on graphs that I put out. But, your insistence that the graph cannot be used needs some shred of actual evidence for your posts to matter on the subject. Otherwise your leaving out information about what is incorrect is just as bad as the marketing people leaving out the labels.

The graph says that Willow Cove can reach 3.25 GHz at the same voltage that Sunny Cove uses to reach 2.5 GHz. That is a 30% frequency improvement. Is that incorrect? How much?

The graph says that Willow Cove can reach 3.7 GHz at the same voltage that Sunny Cove uses to reach 3.0 GHz. That is a 23% frequency improvement. Is that incorrect? How much?

The graph says that Willow Cove can reach 4.2 GHz at the same voltage that Sunny Cove uses to reach 3.5 GHz. That is a 20% frequency improvement. Is that incorrect? How much?

The graph says that Willow Cove can reach 4.7 GHz at the same voltage that Sunny Cove uses to reach 3.9 GHz. That is a 21% frequency improvement. Is that incorrect? How much?

Without hard data, we don't know.

To avoid issues with the SEC, all Intel has to do is say the graphs are meant for illustrative purposes only as evidenced by the fact that there's no actual x-axis and the SEC will let it slide. Marketing teams do stuff like this all the time. I'm sure 10SF/Tigerlake gives a substantial performance/efficiency improvement over 10+/ICL. I'm also sure it is not nearly as big as the graph wants you to believe.

 

[dullard]

No, the Sunny Cove frequency at 0.25V is not even close. The cores would be off at that voltage.

Here's a sample polling of Tigerlake Vid and frequency at idle from someone at the techpowerup forums. I won't post the full table but you can go to the link if you want to see it.

View attachment 44077

Intel has the ability to test at voltages that CPU cannot hit.

 

[dmens]

Intel has the ability to test at voltages that CPU cannot hit.

0.25v is below threshold voltage...

 

[Hitman928]

Intel has the ability to test at voltages that CPU cannot hit.

Even if we assume that this is true for this voltage (it's not but let's just go with it), the graph then becomes based on some unknown structures that are somehow related to ICL and Tigerlake but with an unknown relationship or knowledge of how this graph would change when the full CPU is measured. Basically any way you try and draw useful data from this graph, you are going to run into major problems that destroy the integrity or usefulness of the data. Hence, it's a marketing slide meant to convey the impression that remarkable improvements were made but can't really be trusted due to the manipulation of the graph.

 

[Ajay]

This still dances around the question. What specifically is incorrect in the graph? With this information, we can help the SEC form a lawsuit against Intel for misleading investors.

I agree that I would never leave off units and numbers on graphs that I put out. But, your insistence that the graph cannot be used needs some shred of actual evidence for your posts to matter on the subject. Otherwise your leaving out information about what is incorrect is just as bad as the marketing people leaving out the labels.

The graph says that Willow Cove can reach 3.25 GHz at the same voltage that Sunny Cove uses to reach 2.5 GHz. That is a 30% frequency improvement. Is that incorrect? How much?

The graph says that Willow Cove can reach 3.7 GHz at the same voltage that Sunny Cove uses to reach 3.0 GHz. That is a 23% frequency improvement. Is that incorrect? How much?

The graph says that Willow Cove can reach 4.2 GHz at the same voltage that Sunny Cove uses to reach 3.5 GHz. That is a 20% frequency improvement. Is that incorrect? How much?

The graph says that Willow Cove can reach 4.7 GHz at the same voltage that Sunny Cove uses to reach 3.9 GHz. That is a 21% frequency improvement. Is that incorrect? How much?

Believe what you want to believe. Have a nice day.

 

[dullard]

Even if we assume that this is true for this voltage (it's not but let's just go with it), the graph then becomes based on some unknown structures that are somehow related to ICL and Tigerlake but with an unknown relationship or knowledge of how this graph would change when the full CPU is measured. Basically any way you try and draw useful data from this graph, you are going to run into major problems that destroy the integrity or usefulness of the data. Hence, it's a marketing slide meant to convey the impression that remarkable improvements were made but can't really be trusted due to the manipulation of the graph.

The slides are not talking about Ice Lake or Tigerlake. The slides do not speak of any specific CPU within those lines. Instead, the slides speak of the microarchitecture capabilities. The Sunny Cove microarchitecture could produce CPUs anywhere on the presented Sunny Cove curve. That does not mean Intel will produce CPUs anywhere on that curve. Some combinations just have no reasonable use case. For example, I could produce a car that tops off at 0.25 MPH on a flat road. That does not mean that I will actually do so.

Keeping in mind that it is speaking of capabilities, you can gain a lot of useful information from the graph. Which conclusions are false here?

Discussion - Intel current and future Lakes & Rapids thread

Page 423 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[dullard]

Believe what you want to believe. Have a nice day.

Thanks. I believe in the most accurate and up-to-date data that is available to me. Thus I will believe that Willow Cove can hit ~20% higher frequencies than Sunny Cove--until someone can post any shred of actual evidence against it. If newer or more reliable data becomes available, then I'll update my belief. So far, ~20% higher frequency from that graph is the only data that I have to go on. A 20% gain on a process improvement actually seems logically achievable as well. It isn't like the graph was suggesting a 1000% frequency improvement or anything that is outlandish.

I hope you have a nice day too.

 

[Hitman928]

The slides are not talking about Ice Lake or Tigerlake. The slides do not speak of any specific CPU within those lines. Instead, the slides speak of the microarchitecture capabilities. The Sunny Cove microarchitecture could produce CPUs anywhere on the presented Sunny Cove curve. That does not mean Intel will produce CPUs anywhere on that curve. Some combinations just have no reasonable use case. For example, I could produce a car that tops off at 0.25 MPH on a flat road. That does not mean that I will actually do so.

No, this is false. You cannot produce a CPU on 10SF at 0.25V. Full stop.

You are getting so out in the weeds trying to justify these charts it's getting pretty ridiculous.

Keeping in mind that it is speaking of capabilities, you can gain a lot of useful information from the graph. Which conclusions are false here?

Discussion - Intel current and future Lakes & Rapids thread

Page 423 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

As I said before, without a reliable graph or actual hard data, I can't speak how accurate or inaccurate your statements are. That's the whole point of the discussion. You are asking me to disprove something that hasn't been proven and no reliable data has been presented.

 

[Exist50]

Basically any way you try and draw useful data from this graph, you are going to run into major problems that destroy the integrity or usefulness of the data.

How? We can point to direct frequency improvements, iso-voltage and Vmax to Vmax. What additional information do you think adding specific voltage points would provide that could change the conclusions?

but can't really be trusted due to the manipulation of the graph

What manipulation? Honestly don't see why this is an actual debate. We have real silicon for both Ice Lake/Sunny Cove and Tiger Lake/Willow Cove. We know their approximate voltage ranges, and the data matches everything we see in these graphs. You could ignore Intel's slide entirely and reproduce it purely with existing measurements, or just comparing SKUs.

 

[Hitman928]

How? We can point to direct frequency improvements, iso-voltage and Vmax to Vmax. What additional information do you think adding specific voltage points would provide that could change the conclusions?



What manipulation? Honestly don't see why this is an actual debate. We have real silicon for both Ice Lake/Sunny Cove and Tiger Lake/Willow Cove. We know their approximate voltage ranges, and the data matches everything we see in these graphs. You could ignore Intel's slide entirely and reproduce it purely with existing measurements, or just comparing SKUs.

If you have voltage tables for ICL or TGL, please share. I'd love to compare them to Dullard's chart.

The specific voltage points were meant as a sanity check. If you actually try to figure out the specific voltage points on Dullard's chart, you'll realize the chart is not in the realm of what is possible.

 

[Exist50]

If you have voltage tables for ICL or TGL, please share. I'd love to compare them to Dullard's chart.

The specific voltage points were meant as a sanity check. If you actually try to figure out the specific voltage points on Dullard's chart, you'll realize the chart is not in the realm of what is possible.

A sanity check on what though? If the point is frequency gains from the process, the voltage values don't really add much to the discussion. We already know, again from silicon, that all those frequency points are attainable.

Take the voltage they use to show the +1GHz iso-voltage improvement. Looks like towards the upper-mid end, which would generally be something like 1.0V or so. But for the purposes of evaluating the gains of 10++, what changes if the actual number is 0.9 or 1.1V?

And I'm not even touching on dullard's numbers here. The graph obviously doesn't start at 0V, since nothing will show up until the threshold voltage. Total range is probably something like 0.6-1.3V, give or take.

 

[dullard]

No, this is false. You cannot produce a CPU on 10SF at 0.25V. Full stop.

Intel's 10 nm threshold voltage is ~0.24 V. See page 16 here: https://hal.archives-ouvertes.fr/hal-01551695/document

Typical logic transistors have threshold voltages of 0.2 V to 0.3 V. Again, you won't get a great saleable CPU out of that, but it is something you can get data for to put on a graph of process capabilities.

 

[dmens]

Intel's 10 nm threshold voltage is ~0.24 V. See page 16 here: https://hal.archives-ouvertes.fr/hal-01551695/document

Typical logic transistors have threshold voltages of 0.2 V to 0.3 V. Again, you won't get a great saleable CPU out of that, but it is something you can get data for to put on a graph of process capabilities.

What is this? It is some generic document on the "10nm node", even though Intel processes are geared towards high speed ASIC design and therefore has no bearing to this survey.

And even if it were, the keeper circuitry inside sequentials don't work reliably at threshold voltage, they need margin above that to function. As do the SRAM and any kind of memory element. You are talking nonsense.

 

[Hitman928]

Intel's 10 nm threshold voltage is ~0.24 V. See page 16 here: https://hal.archives-ouvertes.fr/hal-01551695/document

Typical logic transistors have threshold voltages of 0.2 V to 0.3 V. Again, you won't get a great saleable CPU out of that, but it is something you can get data for to put on a graph of process capabilities.

No. You aren't getting any kind of real CPU to operate at 0.25V at 2 GHz. It's not going to happen. Your Idrive is so low there's no way you can drive the necessary gate capacitances+wires quickly enough to get a reliable signal through the logic and out of the chip. Your eye diagram would be non-existent.

 

[dullard]

What is this? It is some generic document on the "10nm node", even though Intel processes are geared towards high speed ASIC design and therefore has no bearing to this survey.

And even if it were, the keeper circuitry inside sequentials don't work reliably at threshold voltage, they need margin above that to function. As do the SRAM and any kind of memory element. You are talking nonsense.

Yes, the memory portions of a CPU need 0.6 V or more. But that is irrelevant to this discussion about logic frequencies vs voltage.

 

[dullard]

No. You aren't getting any kind of real CPU to operate at 0.25V at 2 GHz. It's not going to happen. Your Idrive is so low there's no way you can drive the necessary gate capacitances+wires quickly enough to get a reliable signal through the logic and out of the chip. Your eye diagram would be non-existent.

Again, this isn't a graph of CPUs. It is a graph of process capabilities. The process can go down to those voltages and you can see what frequencies you'll get. You won't make a CPU out of those extreme though since CPUs have other transistors that require different voltages than just logic transistors.

 

[dmens]

Yes, the memory portions of a CPU need 0.6V or more. But that is irrelevant to this dicussion.

Actually, it is. Because on top of memory structures not working, combinational logic transistor performance is so inconsistent at those voltages, that trying to make a claim of generic performance increase is pure fantasy.

 

[Hitman928]

A sanity check on what though? If the point is frequency gains from the process, the voltage values don't really add much to the discussion. We already know, again from silicon, that all those frequency points are attainable.

Take the voltage they use to show the +1GHz iso-voltage improvement. Looks like towards the upper-mid end, which would generally be something like 1.0V or so. But for the purposes of evaluating the gains of 10++, what changes if the actual number is 0.9 or 1.1V?

And I'm not even touching on dullard's numbers here. The graph obviously doesn't start at 0V, since nothing will show up until the threshold voltage. Total range is probably something like 0.6-1.3V, give or take.

The fact that 10SF improved over 10+ is not in debate. Trying to use the charts Intel provided to show how much it is improved is the unknown. They left the x-axis unlabeled on purpose to make the improvements seem better than they are. Because of that, I can't tell you how accurate the chart is in either direction and the values aren't even consistent between charts, so what I am supposed to get from the chart other than 10SF is some kind of improvement over 10+? Any attempt to get real hard values to gauge actual improvement (outside of basically max frequency which is already known) runs into issues. Since we have reviews of ICL and TGL, why not reference those instead of these ambiguous charts?

 

[Hitman928]

Again, this isn't a graph of CPUs. It is a graph of process capabilities. The process can go down to those voltages and you can see what frequencies you'll get. You won't make a CPU out of those extreme though since CPUs have other transistors that require different voltages than just logic transistors.

You are not going to get any kind of real logic to operate at 2 GHz at 0.25V. If you are arguing that they are basically showing what a handful of specific transistors can do, then who cares? The graph becomes just as meaningless as there are so many steps between that and a functional logic block, let a lone any kind of CPU that their graph should just be thrown away.