[Anandtech] Intel's Architecture Day 2018

[PeterScott]

Intel's Architecture Day 2018: The Future of Core, Intel GPUs, 10nm, and Hybrid x86

I know some bits of this were getting some mention in other threads, but it seems like it was big enough to warrant it's own discussion.

Highlights for me were a new architecture (finally) Sunny cove, that looks to have potential (4-5 wide allocation) to deliver Intels first real IPC improvement in MANY years, and also what should be a nice improvement in the the IGP, and a look at Intels multichip solutions(Foveros) and Intel Big-little.

The Q&A was also somewhat interesting, like where they admit being node-locked on their designs really messed them up, and they will be more node agnostic in the future.

Overall, more actual news than we have seen from Intel in some time.

 

[jpiniero]

Exactly! So accusing others of believing the earth is flat because there are "facts" that prove them wrong when there are none is just amusingly ridiculous

The changes are extensive enough though that there should be some decent amount of IPC gain.

 

[Nothingness]

The changes are extensive enough though that there should be some decent amount of IPC gain.

Indeed. I would be disappointed by anything less than 15% higher IPC on a wide range of benchmarks.

 

[PotatoWithEarsOnSide]

The Earth is flat.

 

[Arachnotronic]

What technical counterbalance did we see exactly? I only read messages after messages of people talking about an unreleased x86 chip. It's hard to swallow but currently Apple micro arch is better than Intel for a huge part of the market.

Apple micro-architecture is better than Intel for a huge part of the market, there's no denying that and it's not really hard to swallow either. Apple has been doing great things in silicon for a long time and that strong, focused execution coupled with Intel's stumbles was a perfect recipe for Intel to lose its lead in micro-architecture.

 

[mikk]

The reduced max clocks isn't likely to be an issue since Icelake is mobile U/Y only.

Problem with a theoretical Sunny Cove on 14 nm is of course that the core would be much bigger and draw more power; and throwing in 2 core cores would be too much. 28W Comet Lake U by the way is I believe supposed to be released before anything Icelake.

Sunny Cove cores might come with some efficiency improvements and in case there is a healthy IPC increase over the Skylake generation they could clock 14nm based Sunny Cove cores much lower which means lower voltage requirements. For mobile products the efficiency gains from 10nm are even more useful than for desktop products, 14nm Sunny Cove makes more sense for desktop to me. Do we know if Lakefield is on 10nm?

 

[Arachnotronic]

Sunny Cove cores might come with some efficiency improvements and in case there is a healthy IPC increase over the Skylake generation they could clock 14nm based Sunny Cove cores much lower which means lower voltage requirements. For mobile products the efficiency gains from 10nm are even more useful than for desktop products, 14nm Sunny Cove makes more sense for desktop to me. Do we know if Lakefield is on 10nm?

Lakefield is 10nm.

 

[mikk]

I think the only 14nm based Sunny Cove options are mainstream desktop and entry Xeon E3.

 

[DrMrLordX]

Indeed. I would be disappointed by anything less than 15% higher IPC on a wide range of benchmarks.

15% compared to what? Skylake/Kabylake/CoffeeLake?

The Earth is flat.

Shh don't encourage them.

I think the only 14nm based Sunny Cove options are mainstream desktop and entry Xeon E3.

If they actually launch 14nm processors based on Sunny Cove . . . ugh. Will this be after 10c/20t CometLake?

 

[moinmoin]

Personally I don't consider the increase in IPC or the design going wider as a direct unquestionable positive. Rather I'd be more interested in a IPW (instructions per watt), and any wider CPU design actually beating the efficiency of modern GPUs at the same work.

That aside the one area becoming more and more important is power efficiency of the uncore (especially since it doesn't scale down as well to smaller nodes).

One revolutionary improvement I'd like to see realized sometime is a hardware level implementation of Using Multiple Threads to Accelerate Single Thread Performance without the need for specific compiler optimizations.

 

[mikk]

15% compared to what? Skylake/Kabylake/CoffeeLake?

Skylake/Kabylake/CoffeeLake is the same architecture, don't you know this? It is more than obvious he is comparing this to the Skylake architecture.

 

[DrMrLordX]

Skylake/Kabylake/CoffeeLake is the same architecture, don't you know this? It is more than obvious he is comparing this to the Skylake architecture.

Duh. Why did you think I included the slashes? They're all the same thing, by different names. Not sure how many times I have to repeatedly demonstrate that fact in one thread. It's ridiculous.

If it is in fact 15% better than Skylake, then fooey on that. Better hope they can keep up their clockspeeds.

One revolutionary improvement I'd like to see realized sometime is a hardware level implementation of Using Multiple Threads to Accelerate Single Thread Performance without the need for specific compiler optimizations.

Ah, the infamous "reverse hyperthreading". If only.

 

[jpiniero]

If they actually launch 14nm processors based on Sunny Cove . . . ugh. Will this be after 10c/20t CometLake?

Yeah. Intel does need something for the 2020 Desktop; and assuming 10 nm is not going to be suitable, doing a newer core on 14 nm makes sense although I would think they would rather use Willow Cove if it's possible.

 

[DrMrLordX]

Yeah. Intel does need something for the 2020 Desktop; and assuming 10 nm is not going to be suitable, doing a newer core on 14 nm makes sense although I would think they would rather use Willow Cove if it's possible.

If they do use Sunny Cove on 14nm, they may have to start producing SKUs without iGPUs to save space. Or else those dice could start to get large.

 

[coercitiv]

For mobile products the efficiency gains from 10nm are even more useful than for desktop products

I think it's high time we put this old rule to rest. With mainstream core count exploding in the last 2 years, sustained performance on desktop can use efficiency gains just as much as mobile does.

 

[PotatoWithEarsOnSide]

There's more scope for efficiency gains on desktop, though efficiency hasn't been the primary priority. In the event that it does increase in importance, you can bet your bottom dollar that it gets used effectively.
The argument, if it even exists, is that its harder to gain performance after efficiency than it is to gain efficiency after performance.

 

[NTMBK]

There's more scope for efficiency gains on desktop, though efficiency hasn't been the primary priority. In the event that it does increase in importance, you can bet your bottom dollar that it gets used effectively.
The argument, if it even exists, is that its harder to gain performance after efficiency than it is to gain efficiency after performance.

That's the kind of logic that gets you the Pentium 4.

 

[mikk]

I think it's high time we put this old rule to rest. With mainstream core count exploding in the last 2 years, sustained performance on desktop can use efficiency gains just as much as mobile does.

This is not the point. On desktop you don't have a battery, for mobile you have and the sustained clock speeds on a 4C 15W is really low. You can't put it to rest when even Intel agrees.

 

[coercitiv]

On desktop you don't have a battery, for mobile you have and the sustained clock speeds on a 4C 15W is really low. You can't put it to rest when even Intel agrees.

Are we talking about the same Intel and the same desktop that saw 9900K use 160W rather than 95W to make an impression? Intel's latest 8C/16T chip was begging for 10nm to keep power and heat in check.

14nm desktop Sunny Cove only makes sense as a fallback plan when yields can't sustain a high core count 10nm chip, sure it's better than what we have atm but not what we really need. The power savings of 10nm for 8C+ desktop products are just as badly needed as on mobile, and Intel agreeing or disagreeing doesn't change the laws of physics.

 

[LTC8K6]

Is it possible to have a 14nm 2/4/6 core die and a 10nm 8/10/12 core die?

 

[jpiniero]

14nm desktop Sunny Cove only makes sense as a fallback plan when yields can't sustain a high core count 10nm chip, sure it's better than what we have atm but not what we really need. The power savings of 10nm for 8C+ desktop products are just as badly needed as on mobile, and Intel agreeing or disagreeing doesn't change the laws of physics.

Yields are obviously the biggest issue; but clock speeds are a factor too. No point in moving the desktop to 10 nm if it's going to end up being slower in games due to the lower clocks, even with higher IPC.

 

[coercitiv]

Yields are obviously the biggest issue; but clock speeds are a factor too. No point in moving the desktop to 10 nm if it's going to end up being slower in games due to the lower clocks, even with higher IPC.

The frequency that matters for gaming today is not the same as the max frequency the node allows for. (all-core boost vs. single-core boost)

Let's assume a pessimistic 10% drop in fmax and an even more pessimistic 5% IPC jump as far as gaming is concerned. Assuming core count parity, this pessimistic edition Sunny Cove could still run at 4.5Ghz all-core boost and match a 4.7Ghz CFL in performance. And that would probably be the 65W TDP chip.

You're welcome to make the case for 10nm not even being able to run at 4.5GHz, though. Where do you see it, back to 4.2Ghz as the original Skylake?

 

[PotatoWithEarsOnSide]

An Fmax drop of 10% combined with an IPC gain of 11% = standing still. That's the problem; you don't upgrade just to stand still. It's why Ryzen hasn't been widely accepted as an upgrade from say a 2500/2600K.
I suspect that the ne t gen Ryzen will become the pinnacle for some time, though it won't be much better than the 9900K itself.
Whilst it is great to have choice at the top end, it's just a shame that the performance gains from one generation to the next are so anaemic.

 

[jpiniero]

I am sort of assuming that Comet Lake's many core turbo will be higher than Coffee Lake Refresh.

You're welcome to make the case for 10nm not even being able to run at 4.5GHz, though. Where do you see it, back to 4.2Ghz as the original Skylake?

Low to mid 4's, yeah. Which is fine for U/Y mobile parts, since it seems that the 8565U laptops are having trouble actually hitting 4.6 long enough to complete Cinebench ST, most are closer to like 4.2.

 

[DrMrLordX]

Is it possible to have a 14nm 2/4/6 core die and a 10nm 8/10/12 core die?

Sure, but you're wasting a lot of time and money doing that. Optical shrinks are not free. And it would be a little weird if they did, let's say, 10c/20t Comet Lake followed by 6c/12t Not-Exactly-IceLake (14nm), followed up by I-don't-know-how-many-cores Real-IceLake (10nm) that had lower clocks but used a lot less power.

Yields are obviously the biggest issue; but clock speeds are a factor too. No point in moving the desktop to 10 nm if it's going to end up being slower in games due to the lower clocks, even with higher IPC.

At this point we have no idea what will be the frequency range or behavior of the 10nm process Intel finally decides to sell to the public. Remember that 7nm is supposed to be shipping late 2021 as well. 10nm is looking to have a very narrow window since the first products might not even ship for desktop until 2020.

The frequency that matters for gaming today is not the same as the max frequency the node allows for. (all-core boost vs. single-core boost)

Bear in mind that many of us overclock, and we'll be trying to hit that all-core turbo, all the time. Or higher. Trying, mind you.

Let's assume a pessimistic 10% drop in fmax and an even more pessimistic 5% IPC jump as far as gaming is concerned. Assuming core count parity, this pessimistic edition Sunny Cove could still run at 4.5Ghz all-core boost and match a 4.7Ghz CFL in performance. And that would probably be the 65W TDP chip.

I guess the real question is whether it could be overclocked to something higher than that? If it's anything like GF's 14nm then the answer is mostly "no" (it was stuck near the all-core turbo for the 1800x). If it's "yes" then we'll have to see how fast the power usage ramps up with the overclock.

The respected barriers for 14nm++ right now are in the range of 5.2-5.3 GHz for air/AiO depending on the chip (some are too hot for air). Intel doesn't actually ship a chip that runs those clocks on all cores all the time, but once we get our hands on them, the 9900k can do those speeds easily . . . with the right cooling. The 9700k even moreso.

You're welcome to make the case for 10nm not even being able to run at 4.5GHz, though. Where do you see it, back to 4.2Ghz as the original Skylake?

Er well, OCing factored in, Intel hasn't had a chip limited to those speeds since Broadwell (and some Broadwell could run faster, albeit at high voltages like 1.48v or higher).

Skylake was more like a 4.7GHz chip. Kaby was good for 4.8-5.0 GHz, and Coffeelake has gone as high as 5.2-5.3 GHz. Those are the speed limits we're really looking at right now.

I am sort of assuming that Comet Lake's many core turbo will be higher than Coffee Lake Refresh.

I would be really surprised by that. They already jacked up all-core turbo from Skylake->Kabylake->CoffeeLake->CoffeeLake Refresh. You think they can do it again on the same uarch with the same process? And in a 10c package? I doubt it. They're just gonna bolt on two more cores, TRY to maintain the same clocks they got with the 9900k, and hope for the best.

Low to mid 4's, yeah.

. . . so like Broadwell all over again? Let's hope not.

 

[jpiniero]

I would be really surprised by that. They already jacked up all-core turbo from Skylake->Kabylake->CoffeeLake->CoffeeLake Refresh. You think they can do it again on the same uarch with the same process? And in a 10c package? I doubt it. They're just gonna bolt on two more cores, TRY to maintain the same clocks they got with the 9900k, and hope for the best.

Many core turbo, not all.

. . . so like Broadwell all over again? Let's hope not.

Like I said, since it's an mobile only product it should be okay. I'll add that an unknown when it comes to Icelake is how the McIVR will affect the clocks.