Discussion Intel current and future Lakes & Rapids thread

[moinmoin]

What's exactly problematic about it?

As I already answered in your quote: The lack of values for voltage.

 

[Exist50]

As I already answered in your quote: The lack of values for voltage.

What additional information would specific voltage labels provide you towards showing representative frequency improvement?

 

[Hitman928]

What additional information would specific voltage labels provide you towards showing representative frequency improvement?

How much more voltage does Sunny Cove take to reach 3.9GHz compared to Willow Cove? What happens if you continue to allow for more voltage on Sunny Cove to hit higher frequencies like they allowed on Willow Cove?

 

[moinmoin]

What additional information would specific voltage labels provide you towards showing representative frequency improvement?

The actual curve. As is the curve is completely arbitrary and easy to tweak to make look nice.

 

[Exist50]

The actual curve. As is the curve is completely arbitrary and easy to tweak to make look nice.

It's not arbitrary without the voltages. We see the frequency numbers. That forces a scale.

The only way the curve could be manipulated would be compressing or expanding the x axis, which changes nothing about either iso-voltage comparisons or Vmax to Vmax.

And, again, we see these same results in real silicon...

 

[Thala]

Good points, but too black and white i think. TSMC clocked A72 that was designed for 2Ghz operation in phone, moving from 16nm to 7nm at 4ghz.

To 4.2GHz even So there is no doubt the ARMs Cortex A line of cores could clock north of 4GHz if one is willing to play the power penalty. I am not saying they will necessarily achieve 5GHz without adding pipeline stages, but the limit apparently is somewhere between 4-5GHz.

 

[dullard]

How much more voltage does Sunny Cove take to reach 3.9GHz compared to Willow Cove? What happens if you continue to allow for more voltage on Sunny Cove to hit higher frequencies like they allowed on Willow Cove?

Lets try extrapolating that graph to get rough answers.

1) From the extrapolation, Sunny Cove takes ~56% more voltage than Willow Cove at 3.9 GHz. I don't know what voltage Sunny Cove requires for 3.9 GHz. A quick search came up with something like 1.1 V, although that could be wrong. So, roughly 0.4 V more? Please correct me if the 1.1 V value is wrong, I'd be happy to edit this post.

2) From the extrapolation, it looks like 4.2 GHz (conservative) to 4.25 GHz (generous) if Sunny Cove could handle more voltage vs 5.0 GHz for Willow Cove.

 

[Hitman928]

Lets try extrapolating that graph to get rough answers.

1) From the extrapolation, Willow Cove takes ~56% more voltage than Sunny Cove at 3.9 GHz. I don't know what voltage Sunny Cove requires for 3.9 GHz. A quick search came up with something like 1.1 V, although that could be wrong. So, roughly 0.4 V more? Please correct me if the 1.1 V value is wrong, I'd be happy to edit this post.

2) From the extrapolation, it looks like 4.2 GHz (conservative) to 4.25 GHz (generous) if Sunny Cove could handle more voltage vs 5.0 GHz for Willow Cove.
View attachment 44049

I'm assuming you meant to type Sunny Cove takes ~56% more voltage at 3.9 GHz. That looks like a pretty good 'eye' estimate from your chart. Here's a fun estimate to try given your above numbers, what voltage would Willow Cove be using at 2.5 GHz?

Edit: or for some real fun, 2 GHz.

 

[IntelUser2000]

I mean something's got to give. If Gracemont is indeed around Skylake IPC and can reach 4Ghz, Intel is a bunch of imbecilles for not designing a SKU with 32 Gracemont cores and 2 Golden Cove cores (it would take up similar area and fit on the ring due to gracemont 4 core "clusters"). It would be similar in size to Alder Lake and It would absolutely beat 5950X in throughput oriented tasks (e.g. rendering) and even give 3960X a run of it's money when bandwidth or I/O is not the limit.

No way it'll beat 5950X. Zen 3 is what, 20-25% faster per clock compared to Skylake? Plus it has SMT, and clocks higher. And it'll be behind in every program that uses more than 2 cores. Maybe in the most optimistic scenario it'll come close but that's about it.

Games and consumer applications aren't perfectly threaded but it's definitely using more than 4 cores now. People aren't going to buy something that's slower in >75% of usage scenarios, because we're not in 2018 where Skylake was still a decent chip.

That's the tradeoff. If they'd absolutely want to beat 5950X, it would be a 8+16 part or even 8+32.

I might be wrong, but IMO even getting 3.5 Ghz out of cores so "tight" would be an unprecedented achievement.

Really? Icelake had to lower frequencies by 20%, but Tremont clocks 5-7% higher than Goldmont Plus Refresh despite a bigger design change using the same garbage 10+ process. But you think a much improved 10nm+++(ESF) will huff and puff to reach the same 6% higher with a relatively unlimited power budget?

Whether they are willing to is also different from what's possible.

Nevermind that the engineering sample clocks 100MHz lower than the "unprecedented achievement".

2) From the extrapolation, it looks like 4.2 GHz (conservative) to 4.25 GHz (generous) if Sunny Cove could handle more voltage vs 5.0 GHz for Willow Cove.

4.2GHz is exactly what Sunny Cove can clock without limits. As evidenced by the 1068G7 part. We know from the server Icelake that it's severely limited in it's ability to clock. 5GHz on TGL-H is 19% higher.

 

[Zucker2k]

No way it'll beat 5950X. Zen 3 is what, 20-25% faster per clock compared to Skylake? Plus it has SMT, and clocks higher. And it'll be behind in every program that uses more than 2 cores. Maybe in the most optimistic scenario it'll come close but that's about it.

Games and consumer applications aren't perfectly threaded but it's definitely using more than 4 cores now. People aren't going to buy something that's slower in >75% of usage scenarios.

That's the tradeoff. If they'd absolutely want to beat 5950X, it would be a 8+16 part or even 8+32.

Whoever stole @IntelUser2000's account, please give it back!

 

[IntelUser2000]

I'm assuming you meant to type Sunny Cove takes ~56% more voltage at 3.9 GHz. That looks like a pretty good 'eye' estimate from your chart. Here's a fun estimate to try given your above numbers, what voltage would Willow Cove be using at 2.5 GHz?

Dullard, the reason that graph can be misleading is because the bottom is 0.5GHz, not 0. Yes I see you tried to compensate for that, but it takes much more than 0.1V or so to sustain 400MHz. Actually it's something close to 0.4V or more.

I don't think we can get exact numbers due to that. Overall you got it though.

Whoever stole @IntelUser2000's account, please give it back!

Long time ago a friend asked for assistance in buying a computer. I let my bias seep in and recommended a Pentium D system. I realized my mistake a bit after. I won't do that again.

Admitting you made a mistake is the first step in fixing things. And I feel Intel isn't fully there yet. Unfortunately that process will hurt. So while I'm still more excited with their chip than anyone else's, I'm much more open now. Hope AMD dominates for a few more years and becomes more of a duopoly. Whatever is the best, or whatever offers good price/performance I will use and recommend.

Zen 3 will be >50% faster than the 32 core Gracemont @ 4GHz in low number of threads and it won't look pretty nor will be received well.

Flock of chickens still don't work.

 

[dmens]

To 4.2GHz even So there is no doubt the ARMs Cortex A line of cores could clock north of 4GHz if one is willing to play the power penalty. I am not saying they will necessarily achieve 5GHz without adding pipeline stages, but the limit apparently is somewhere between 4-5GHz.

Nothing to it. This is a test/demo chip. Resynthesize the RTL with massively oversized gates and no floorplan limitation then shove a lot of power into the part, and you can crank up an ARM core to high frequencies. The issue is whether it is a viable product to sell, considering its low power characteristics just went out the window.

 

[Thunder 57]

Dullard, the reason that graph can be misleading is because the bottom is 0.5GHz, not 0. Yes I see you tried to compensate for that, but it takes much more than 0.1V or so to sustain 400MHz. Actually it's something close to 0.4V or more.

I don't think we can get exact numbers due to that. Overall you got it though.



Long time ago a friend asked for assistance in buying a computer. I let my bias seep in and recommended a Pentium D system. I realized my mistake a bit after. I won't do that again.

Admitting you made a mistake is the first step in fixing things. And I feel Intel isn't fully there yet. Unfortunately that process will hurt. So while I'm still more excited with their chip than anyone else's, I'm much more open now. Hope AMD dominates for a few more years and becomes more of a duopoly. Whatever is the best, or whatever offers good price/performance I will use and recommend.

Zen 3 will be >50% faster than the 32 core Gracemont @ 4GHz and it won't look pretty nor will be received well.

Flock of chickens still don't work.

Friends don't let friends but Pentium D's. Glad you learned from your mistake.

 

[Hans de Vries]

These socalled VF curves are just "illustrations". To quote my own post from last year:

If only these graphs weren't so crappy inconsistent. The same curve is very different on two consecutive pages.

View attachment 28009

 

[Gideon]

No way it'll beat 5950X. Zen 3 is what, 20-25% faster per clock compared to Skylake? Plus it has SMT, and clocks higher. And it'll be behind in every program that uses more than 2 cores. Maybe in the most optimistic scenario it'll come close but that's about it.

That's the tradeoff. If they'd absolutely want to beat 5950X, it would be a 8+16 part or even 8+32.

First I agree that some of my claims were dubious and my wording was bad (especially the "unprecedented" and "imbecile" part). I also agree with most of your other statements. I was the first to bring the same arguments up when somebody claimed that Alder Lake can outperform a R5950X and it's "simple math" (ignoring the Clocks, SMT and IPC deficit). And yes, The "flock of chickens" approach is fundamentally stupid for a general purpose CPU.

The main point to drive home was that a 2 + 32 chip would have similar die-area as a 8 + 8 chip. Something that perhaps wasn't that obvious to most) and if all people expect were true, Intel would probably add more small cores to he chip than it has.

Games and consumer applications aren't perfectly threaded but it's definitely using more than 4 cores now. People aren't going to buy something that's slower in >75% of usage scenarios, because we're not in 2018 where Skylake was still a decent chip.

It was a stupid example, but to my defence, I never meant it to be the only Alder Lake SKU, rather a possible SKU on top of the current one targeting the small niche that needs to run near embarrassingly parallel stuff (rendering, some scientific code) with some light desktop workloads on top without breaking the bank. Yes it wouldln't actually make any financial sense. You'd probably still need at least 6 big cores for good all-around performance.

. And it'll be behind in every program that uses more than 2 cores.

But here is where I disagree. My original claim was quite specific:

In throughput limited programs, meaning embarrassingly parallel stuff like rendering, such SKU (32 x Gracemont running at 3.8 - 4Ghz with Skyale IPC + 2/4 Golde Cove on top) would be faster, and I absolutely stand by it.

Let's do the napkin math:

1. Zen 3 indeed has about 20-25% more IPC than Skylake. Let's pick 25% here.
2. Zen 3 has SMT which gives about 30-40% more performance (In this techpowerup article it's 30% in cinebench, 36% in corona and 46% in Blender for 3900X), let's pick 40%.
3. Zen 3 doesn't actually have higher clocks vs these hypothetical Gracemont cores as the 16 Core turbo, of a 5950X fully loaded, is actually about 3.85 GHz. So it would be a wash.

1.25 * 1.40 = 1.75

So all in all a Zen 3 core in 5950x would end up having roughly ~175% of the performance of this magical Gracemont core in Cinebench. But the proposed SKU has 2x the gracemont cores and 2 Golden Cove cores (with 4 threads) on top. It would surely beat a reasonable 145W ~4Ghz 16 core Zen 3 CPU at Intel's ridiculous (250W) PL2 numbers.

Obviously I don't think such an SKU would actually work quite as well in practice (which was the whole point of this endevour).

I'm hesitant to believe that Gracemont actually has exactly Skylake's performance in all (including heavy FP) workloads or that it can sustain 4Ghz indefinitely in such loads without any thermal issues (due to thermal density). If anything I was trying to give a good counter-example.

Really? Icelake had to lower frequencies by 20%, but Tremont clocks 5-7% higher than Goldmont Plus Refresh despite a bigger design change using the same garbage 10+ process. But you think a much improved 10nm+++(ESF) will huff and puff to reach the same 6% higher with a relatively unlimited power budget?

Whether they are willing to is also different from what's possible.

Nevermind that the engineering sample clocks 100MHz lower than the "unprecedented achievement".

Fair enough. I already agreed that the wording was bad. 3.5 Ghz is a reasonable target, 3.3 -> 3.8 is a 15% gain.

I don't doubt Intel could lay out Gracemont in a way it could actually do 4+ Ghz given enough voltage if that was their main thing they set out to do.I do seriously doubt they designed these small cores in such a way in this very specific 8 + 8 Alder lake SKU, that was also meant to be used in mobile.

I would presume that such a design (essentially having 4 Skylake cores in the area of one Tiger Lake core) would have serious thermal density issues even if the voltages were fine for 4 Ghz. I just don't see Intel wasting design effort and die-are on that

 

[IntelUser2000]

I don't doubt Intel could lay out Gracemont in a way it could actually do 4+ Ghz given enough voltage if that was their main thing they set out to do.I do seriously doubt they designed these small cores in such a way in this very specific 8 + 8 Alder lake SKU, that was also meant to be used in mobile.

I backtracked to see why the topic started in the first place since it started to feel meaningless.

IgorLabs sample has Gracemont clocking at 3.4GHz, but only from 1-4 cores. 5-8 cores it clocks at 3GHz.

The 6W Goldmont Plus part is about 2GHz under MT. Based on the performance and the ~30% perf/clock gain, 6W Tremont is probably around 2.3-2.4GHz, and 10W Tremont at about 2.7-2.8GHz.

Unlike the U-series, or even H-series Core parts, Goldmont Plus doesn't throttle clocks at all under such conditions, meaning it's easy to extrapolate.

It's possible we'll see 3.8GHz Gracemont, but MT clocking at 3.4GHz. That's still quite a gap from Zen 3. That doesn't mean Intel will do it since if it requires voltage increase, the 13% increase in frequency will require 50% extra power.

You are likely correct on Intel not doing physical redesign of Gracemont just for Alderlake.

I'm hesitant to believe that Gracemont actually has exactly Skylake's performance in all (including heavy FP) workloads or that it can sustain 4Ghz indefinitely in such loads without any thermal issues (due to thermal density). If anything I was trying to give a good counter-example.

There's nothing special about Skylake. It's a 5 year old design. In a few months it'll celebrate it's 6th birthday. What makes it sound better than it actually is that they are still selling the damn thing.

The year is 2015, and Intel launches Skylake. We still have Airmont, which is practically Silvermont in performance.

All the while Skylake is sold for another 6 years, we go from Airmont to Goldmont, a 30% increase in performance at the same clock, and then Goldmont to Goldmont Plus, which is another 30-40%, and finally Tremont, which is yet another 30% increase in perf/clock, plus some clocks every transition.

 

[JoeRambo]

There's nothing special about Skylake. It's a 5 year old design. In a few months it'll celebrate it's 6th birthday. What makes it sound better than it actually is that they are still selling the damn thing.

As design it older by few more years, It was released in 2015, but taped out in 2014 and probably has ideas from ~2012 timeframe. The fact that it took AMD almost 8 years to beat it with ZEN3 is just mind boggling. And it highlights Intel's failures even more.

 

[dullard]

I'm assuming you meant to type Sunny Cove takes ~56% more voltage at 3.9 GHz. That looks like a pretty good 'eye' estimate from your chart. Here's a fun estimate to try given your above numbers, what voltage would Willow Cove be using at 2.5 GHz?

Edit: or for some real fun, 2 GHz.

D'oh. You are correct, I swapped the two names as I was typing the post. I have corrected it.

• 43% less for Willow Cove at 2.5 GHz. That works out to be 74% more for Sunny Cove than Willow Cove.
• 45% less for Willow Cove at 2.0 GHz. That works out to be 82% more for Sunny Cove than Willow Cove.

Again, these are quite rough estimates from an extrapolation.

 

[dullard]

These socalled VF curves are just "illustrations". To quote my own post from last year:

That is how CPUs work though. Depending on exact designs, both graphs can be 100% correct. You can design a CPU to be more power efficient (your top graph) which will have a bit lower voltage for the same frequency but a lower max frequency. Or you can design a CPU to be a higher max frequency (your bottom graph) which will have a lower frequency at a given voltage but a higher max frequency.

That is how optimizations go: there are tradeoffs. You might make one tradeoff for a mobile CPU (top graph) and a different tradeoff for a desktop CPU (bottom graph). There isn't just one Willow Cove.

 

[Mopetar]

As design it older by few more years, It was released in 2015, but taped out in 2014 and probably has ideas from ~2012 timeframe. The fact that it took AMD almost 8 years to beat it with ZEN3 is just mind boggling. And it highlights Intel's failures even more.

I think a good bit of Intel's problems came from issues on the process side of things which left the CPU teams in a lurch. The 10nm node has been such a debacle that people tend to forget that 14nm had some early issues as well. I wonder how many times the design team was told that they had to throw everything away because 10nm wasn't happening or that some part of the design wouldn't work on the process and start over again or that they needed to fall back to using 14nm that had some improvements to it. Of course they were also told not to worry about making any seriously radical changes, because we'll have 10nm working soon. This time we promise!

Couple that with the specter and meltdown vulnerabilities necessitating further changes and it's pretty easy to see how Intel stagnated. By the time Skylake had launched AMD wasn't even making CPUs to compete against anything but Intel's low end so it's likely that they didn't feel a great deal of pressure to fix all of those problems as quickly as they needed to be fixed. When your only real competition is your previous product and you can't conceivably capture additional market share, the incentive to spend a lot of effort on improvements goes away as well. Instead Intel probably had a lot more focus on expanding in areas where they weren't as big (mobile devices) or new markets (cellular modems) where they didn't even have a presence, because from a business perspective that makes the most sense.

 

[Hitman928]

D'oh. You are correct, I swapped the two names as I was typing the post. I have corrected it.

• 43% less for Willow Cove at 2.5 GHz. That works out to be 74% more for Sunny Cove than Willow Cove.
• 45% less for Willow Cove at 2.0 GHz. That works out to be 82% more for Sunny Cove than Willow Cove.

Again, these are quite rough estimates from an extrapolation.

I suggest trying to figure out the voltages according to your graph at 2.5 GHz and 2 GHz. You'll see your rough estimates aren't even that. The origin and scaling of the x-axis is completely unknown.

 

[dullard]

I suggest trying to figure out the voltages according to your graph at 2.5 GHz and 2 GHz. You'll see your rough estimates aren't even that. The origin and scaling of the x-axis is completely unknown.

But why do you need to know the voltages of Willow Cove? Do you know the voltages of Sunny Cove? Did you need to know the voltages of Sunny Cove?

To me, what matters is mostly performance. Efficiency (related to power) is also important, but for my purposes it is secondary. Exact voltage values just don't really matter at all. Yes, power is related to voltage. But power is what matters to me, not voltage. You know the power of Sunny Cove and you can see from the graph how that would project to Willow Cove power usage. So, why do you need to know if something is 1.09 V or 1.08 V?

 

[Ajay]

Lets try extrapolating that graph to get rough answers.

1) From the extrapolation, Sunny Cove takes ~56% more voltage than Willow Cove at 3.9 GHz. I don't know what voltage Sunny Cove requires for 3.9 GHz. A quick search came up with something like 1.1 V, although that could be wrong. So, roughly 0.4 V more? Please correct me if the 1.1 V value is wrong, I'd be happy to edit this post.

2) From the extrapolation, it looks like 4.2 GHz (conservative) to 4.25 GHz (generous) if Sunny Cove could handle more voltage vs 5.0 GHz for Willow Cove.
View attachment 44049

This is a MARKETING slide. Having work with marketing folks before; I trust them as far as I can throw them.
Not liking my official report on a particular new component we were introducing, on of the guys met me in the lab to go over my testing results. I walked him through the whole thing, but he seemed particularly interested in a chart I had that was a bit different from the official chart. I'd circled a point on the graph as an 'outlier' because this component couldn't preform that well, and there must have been some error in my testing. Next thing I know we have a poster on the wall outside of marketing "Highest performing (gizmo) in the industry" - the value given was that of the outlier. Next thing I know I'm being called into a meeting with a higher ranking engineer, department head and the VP of engineering - I was still sore three days later.

 

[dullard]

This is a MARKETING slide. Having work with marketing folks before; I trust them as far as I can throw them.
Not liking my official report on a particular new component we were introducing, on of the guys met me in the lab to go over my testing results. I walked him through the whole thing, but he seemed particularly interested in a chart I had that was a bit different from the official chart. I'd circled a point on the graph as an 'outlier' because this component couldn't preform that well, and there must have been some error in my testing. Next thing I know we have a poster on the wall outside of marketing "Highest performing (gizmo) in the industry" - the value given was that of the outlier. Next thing I know I'm being called into a meeting with a higher ranking engineer, department head and the VP of engineering - I was still sore three days later.

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?

 

[Hitman928]

But why do you need to know the voltages of Willow Cove? Do you know the voltages of Sunny Cove? Did you need to know the voltages of Sunny Cove?

To me, what matters is mostly performance. Efficiency (related to power) is also important, but for my purposes it is secondary. Exact voltage values just don't really matter at all. Yes, power is related to voltage. But power is what matters to me, not voltage. You know the power of Sunny Cove and you can see from the graph how that would project to Willow Cove power usage. So, why do you need to know if something is 1.09 V or 1.08 V?

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.