[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.

 

[coercitiv]

Sunny Cove, even if it achieved it's perf. / clock goals, will use inherently more power to run at iso clocks (efficiency may hopefully go up, but absolute power will definitely go up). Couple that with 9900K being already very power constrained, and it should be immediately obvious that both base clocks and sustainable all-core boost clocks are bound to come down. So even by staying on 14nm the new design will lose clocks. (except for the single-core marketing figure, that can stay nice and pretty - and mostly useless)

Delivering desktop Sunny Cove on 14nm is just as big of a compromise as it would be on their unrefined 10nm, it will affect high core count throughput just as badly, if not worse.

Maybe now, in the wake of AMDs CES presentation of Zen 2, people will finally understand Sunny Cove on 14nm would do little to keep the competition in check.

The 8C/16C Zen 2 sample matched 9900K @ 4.7Ghz performance in CB15 while using considerably lower power and running bellow final clocks. As it stands now, with the first 7nm performance numbers at hand, I find it even harder to believe a functional 10nm would bring about a big drop in frequency. Even if it's a best case scenario, Zen 2 peformance in CB15 can't be all IPC driven, there has to be a decent jump in frequency too, which means TSMC is likely delivering 4.5Ghz+ on 7nm.

 

[epsilon84]

Maybe now, in the wake of AMDs CES presentation of Zen 2, people will finally understand Sunny Cove on 14nm would do little to keep the competition in check.

The 8C/16C Zen 2 sample matched 9900K @ 4.7Ghz performance in CB15 while using considerably lower power and running bellow final clocks. As it stands now, with the first 7nm performance numbers at hand, I find it even harder to believe a functional 10nm would bring about a big drop in frequency. Even if it's a best case scenario, Zen 2 peformance in CB15 can't be all IPC driven, there has to be a decent jump in frequency too, which means TSMC is likely delivering 4.5Ghz+ on 7nm.

Even with Zen+ AMD already scores higher than Intel at the same clocks in CB15 (see link below) I think its fair to say it's a bit of an outlier in the grand scheme, how many real world apps perform better clock for clock on Zen+ versus CFL?

https://www.guru3d.com/articles-pages/amd-ryzen-5-2600x-review,9.html

I can see why AMD used Cinebench though. It heavily favours their SMT implementation over Intel HT and assuming a 10% uplift in IPC relative to Zen+ it probably wouldn't take much more than a 4GHz Zen 2 chip to match a 9900K in CB15. Hence the great power numbers for the Zen2 sample.

It's probably not the best way to estimate future performance based on these Cinebench numbers, IMHO...

 

[coercitiv]

It's probably not the best way to estimate future performance based on these Cinebench numbers, IMHO...

I wasn't estimating future Zen 2 performance, I was talking about power usage and clocks as it relates to this thread.

 

[epsilon84]

I wasn't estimating future Zen 2 performance, I was talking about power usage and clocks as it relates to this thread.

And the reason for the low power in the Zen 2 sample is probably because it's running at much lower clocks. You'd see the same benefit from Sunny Cove, even a 14nm version of it.

A hypothetical 14nm Sunny Cove with +10% IPC could see the same results here because you could probably clock it at 4.2 - 4.3GHz and have it match a 9900K.

So even if clocks regress slightly, it's probably still worth it as a last resort to at least get better power consumption - a 4.5GHz 14nm Sunny Cove would likely beat a 4.7GHz 9900K but with lower power.

 

[mikk]

Yes exactly. It depends on the clock speeds they need to reach a specific target which depends on the IPC increase we don't know yet. We have Core based 6C CPUs running at 45W or 4C CPUs running at 15W....with much lower clocks of course. The power efficieny over 4 Ghz drops significantly.

 

[Spartak]

Exactly. Some people here are too much focussed on all core turbo speeds. This demo shows exactly why Intel needs a 10core sunny cove, at lower all core clockspeeds, on 14nm.

Single / dual core turbospeeds which they do market on the box can remain the same; there's ample thermal room there.

 

[DrMrLordX]

I care about all-core turbo speeds because that's where I'm going to try to run my chip 24/7. I'm crazy like that. Lots of people are running their 9900ks @ 5.0 GHz too so don't say that I'm alone.

 

[Spartak]

You can't use your particular 'need to overclock all cores to N GHz' obsession as an argument to why sunny cove or more cores doesnt make sense on 14nm. More transistors enables better performance within the same power budget.

It's also more expensive. But at this point it would be suicide for Intel to be slaughtered on the desktop for almost one and a half year if 20Q4 is still the launch date for sunny cove on 10nm+. The Zen2 ES humiliated the 9900K and anyone just looking at the benchmark result didnt get the memo.

I can see comet lake still using skylake if they can get 10nm+ HVM late '19 / out the door early '20. But seeing they called their coffee lake refresh '9th gen' instead of 8x50 to me points further to their next gen chips featuring a new naming scheme, possibly even a rebranding. You dont start a new naming scheme on the architecture you have been reusing for 4 years. They could call comet lake 10xxx but that's a horrible naming scheme when the 9th gen didnt even have a new codename for the same architecture, and they'll need the 'fancy next gen' stamp for marketing purpose more than ever this year.

 

[DrMrLordX]

You can't use your particular 'need to overclock all cores to N GHz' obsession as an argument to why sunny cove or more cores doesnt make sense on 14nm. More transistors enables better performance within the same power budget.

If Intel wants a halo chip to replace the 9900k, then they'd better bring the clocks. Furthermore, we have no evidence to conclude that Sunny Cove on 14nm++ would have desirable voltage scaling at clocks well below 5 GHz. If it's a wider core than Skylake then it's going to need even more die area per core than Skylake/Kabylake/CoffeeLake and it's going to consume more power clock-per-clock. Dropping clocks by 10% to "cancel out" the gains from the uarch may not be as desirable as you might think.

Intel doesn't really have a choice about what's going to happen to them over the next two years until they sort through their 10nm troubles. Sunny Cove may well be the last uarch change that can't be back-ported (Intel has never explained exactly what they meant when they said they weren't going to tie uarch to process anymore; that could be Willow Cove) which means 14nm Sunny Cove may not even be an option.

Unlike AMD, Intel is aiming at mobile first this year. We get IceLake U/Y and that's about it. Don't expect them to do much on the desktop.

 

[TheGiant]

Well SunnyCove on 14+++ may have a point only if the 10nm rollout is in high volume in late 2020.
If Intel manages the IPC uplift and downlock an 8C sunnycove to 4,5GHz, where the voltage and power isn't such a big problem and with some tweaks, it can compete with the new ryzen
and if its overclockable to 5GHz with even bigger power than current 9900K, they will have the performance crown and can make a marketing around it
currently there is too much fog around intel release date of anything
but the sunny cove looks finally as worthy skylake succesor

 

[Spartak]

If Intel wants a halo chip to replace the 9900k, then they'd better bring the clocks. Furthermore, we have no evidence to conclude that Sunny Cove on 14nm++ would have desirable voltage scaling at clocks well below 5 GHz. If it's a wider core than Skylake then it's going to need even more die area per core than Skylake/Kabylake/CoffeeLake and it's going to consume more power clock-per-clock. Dropping clocks by 10% to "cancel out" the gains from the uarch may not be as desirable as you might think.

Intel doesn't really have a choice about what's going to happen to them over the next two years until they sort through their 10nm troubles. Sunny Cove may well be the last uarch change that can't be back-ported (Intel has never explained exactly what they meant when they said they weren't going to tie uarch to process anymore; that could be Willow Cove) which means 14nm Sunny Cove may not even be an option.

Unlike AMD, Intel is aiming at mobile first this year. We get IceLake U/Y and that's about it. Don't expect them to do much on the desktop.

Intel very explicitly explained the problem with backporting wasnt related to process or sunny cove but with the tools lacking for that. They now have the tools and the means to backport sunny cove to 14nm. If they will do it or not remains the question but they would commit corporate suicide if they didnt.

Also, sunny cove will never see 10nm for the desktop. Here as well Intel mentioned first gen 10nm doesnt have the clocks for it. The desktop will never see 10nm.

 

[DrMrLordX]

I dunno man, Comet Lake makes it look like Intel has already cast the dice on how to move forward. Guess we'll see how things really work out.

 

[coercitiv]

The desktop will never see 10nm.

From Ian Cutress' interview with Raja, Keller, Murthy.

Q: Will we ever see a 10nm monolithic desktop CPU at the high end?

A: Yes

 

[Adawy]

Yeah, i'm pretty excited to see Intel's UHP 10nm process, but it's still unclear whether it's gonna be for Comet Lake or Ice Lake. If Comet Lake is even real.

 

[jpiniero]

Yeah, i'm pretty excited to see Intel's UHP 10nm process, but it's still unclear whether it's gonna be for Comet Lake or Ice Lake. If Comet Lake is even real.

Comet is very real but also 14 nm based.

From Ian Cutress' interview with Raja, Keller, Murthy.

I have serious doubts about that. I think you're more likely to see them try to rush out 7 nm instead once (if?) possible. Or using a 14 nm Willow Cove based CPU cores with a product using Forveros (and not monolithic)

 

[DrMrLordX]

If Comet Lake is going to be real, it's probably a Skylake-based stopgap before Intel can launch desktop IceLake. At least, that seems the most-probable scenario.

 

[jpiniero]

If Comet Lake is going to be real, it's probably a Skylake-based stopgap before Intel can launch desktop IceLake. At least, that seems the most-probable scenario.

Comet isn't a stopgap, it's a replacement really.

 

[DrMrLordX]

Comet isn't a stopgap, it's a replacement really.

A replacement for what? Desktop IceLake?

The way I see it, Intel's most-probable scenario is June/July 2019 for Comet Lake and who-knows-when in 2020 for desktop IceLake. Maybe Feburary/March if they are lucky.

 

[Spartak]

From Ian Cutress' interview with Raja, Keller, Murthy.

Key operative word was "first gen". 10nm+ is when we will see desktop chips.

 

[exquisitechar]

I think Sunny Cove won't even come to desktop, Willow Cove will.

 

[Spartak]

Comet isn't a stopgap, it's a replacement really.

if comet lake is coffee lake based it can't get much closer to the definition of stopgap.

 

[Spartak]

I think most people here agree we won't see a 10nm desktop Sunny Cove. My assumption is this is what got 'killed'.

This means we either get 14nm skylake IIII and 10nm+ Sunny Cove / Willow cove next year, or 14nm Sunny Cove this year, and 10nm+ Willow cove next year.

Skipping a gen on the desktop seems unlkiely to me, but trailing the mobile processor by more than a year is even less likely. So Sunny Cove next year is unlikely for all scenarios but a 20Q1 launch.

Another thing not so much discussed was the rumored 12nm half node replacing the 10nm. I guess that would be a best case scenario for Intel if they can get a 12nm Sunny Cove out for the desktop this year.

 

[jpiniero]

A replacement for what? Desktop IceLake?

Yeah.

if comet lake is coffee lake based it can't get much closer to the definition of stopgap.

But that implies they would release a Desktop Icelake at some point, which they won't.

 

[Spartak]

desktop Ice Lake cancelled != Sunny Cove cancelled. We can either see it on 14nm, a half node, or 10nm+. Even 10nm is a remote possibility, only later.

 

[LTC8K6]

I think we will see 10nm on the desktop sooner than most people think.