[Ashraf] 10nm "Lakefield" SoC with Intel big + little cores

[darkswordsman17]

I am puzzled by the fact you think Lakefield being on 10nm is strange, when they'll get regular Icelake chips on 10nm. Lakefield also sounds like a product coming later, so it'll have had longer time for the process to mature.

This is unlike shipping 100 million modems to Apple, because they need to guarantee that number has to work.

Also, Digitimes isn't always reliable.



Again, you are saying assumptions as they are facts.



HoloLens doesn't have enough of a market to make a custom chip for it. You are talking tens of thousands at best. If a company did ask, it makes much more sense its bigger ones like HP. They even made a custom package chip with on-package XMM 7560 WiFi to use it in the Spectre Folio.

Also it starts making sense to make such hybrid chips. If done well it can offer an optimal configuration between high single thread and multi-thread with low TDP.

Why does it need a large market? This chip actually seems perfect for a lower market due to the complexities of production at this time. It seems like someone asked Intel if they could make a chip like this and Intel did it as almost an R&D project, and found out they could. Seems like a pretty limited product capacity would have been an ideal scenario for it. And considering the high cost of the HoloLens, it'd likely be able to absorb a high per unit cost (compared to much cheaper tablets and phones) since I doubt this is cheap. Finding out they can make it and so are going to offer this to others if they want it (which means economies of scale can improve and this chip could be used for other markets/devices). Plus you're basing that on the current HoloLens. If they're able to make it cheaper, adoption will improve significantly.

Oh and there's this:
https://techcrunch.com/2018/11/28/m...act-to-outfit-soldiers-with-hololens-ar-tech/

Seems that rumors are saying the next HoloLens will use Snapdragon 850 though, but considering that Microsoft said there's been interest in it but people were turned off by the older hardware (the old Atom chip they were using), I feel like this chip would be a lot more interesting for developers (easier to port Windows apps, this will present a huge improvement in performance, and should offer low enough power on standby and has cellular capability). Plus it would make for direct continuation of what has already been developed for HoloLens.

How would a 64EU Intel GPU compare to modern ARM based SoC GPUs? (Non-Apple ones, like say the 850.) Maybe I'm wrong, but that seems like a pretty strong GPU, and it makes me think its for a product where visuals will be of extra emphasis. Suppose it could be for some other AR/VR headset.

Considering the other company Intel put so much work into making a similar product for...could this have been for Apple? As a way to try and get them to hold off on moving to their own stuff. Something for the Macbook non-Pro? And could be similar to the AMD GPU chip, where it was made for Apple, but maybe didn't end up using it so they're letting other OEMs use it if they want.

 

[Dayman1225]

Just a reminder that the Sunny Cove demo heat-sink was named ICL - U...

 

[DrMrLordX]

Wrong. There is lots of activity for the Open Source and Linux support on pages such as freedesktop which is not the case for Lakefield.

Anything particular to the IceLake/SunnyCove uarch also pertains to (part of) Lakefield. But I see your point.

I have to disagree, this is not the case. Intel didn't even confirm the architecture for Lakefield, they just confirmed 1 big core + 4 small cores. We know basically nothing about Tremont, assuming these are the smaller cores.

No, we only know (or think we know) what Ashraf has made public and what Intel chose to show us the other day. But has anyone confirmed core counts on an ICL-U/Y for a future shipping product?

You clearly don't follow the page here, there are ICL-U ES entries on various pages like Furmark, 3dmark, Sisoft, Geekbench every few weeks. The last Geekbench entry from ICL-U was leaked in October. The first entries are very old, more than a year. Lakefield on the other side is dead at this point, we did get a broken screen from some twitter user but that's it. First driver entries as well, Lakefield has been added much later.

Fair enough. We saw a lot of entries like that for Cannonlake though, and look at how that turned out. Hopefully ICL-U won't meet the same fate.

64. The actual product could be partially disabled or they offer multiple options.

I was actually thinking that, if Intel was still having problems with bad EUs due to yields, that they could build in 72 or 80 of them (like 2xGT2 from Cannonlake) and disable down to 64 for market segmentation purposes. That way, as long as 80% of EUs were good, they still had a shippable product.

 

[jpiniero]

Considering the other company Intel put so much work into making a similar product for...could this have been for Apple?

Intel isn't going to be able to provide a high enough volume for Apple's liking. Have to agree that it is either HoloLens or the Surface Phone.

 

[IntelUser2000]

Why does it need a large market? This chip actually seems perfect for a lower market due to the complexities of production at this time.

Because Intel doesn't work that way.

They don't even provide a custom chip for the HEDT market, they recycle server chips. They sell few millions per year at $1000 or more.

Lakefield isn't a simple change. You have to integrate two very different cores, and make them work in synergy. If you want to make it in the Windows world where one little software or firmware mishap can make it not functional, then they better do it well. Either the customer is pretty big or the customer asking is just an incentive.

Look what happened when they developed Iris parts. Apple may have asked them, but for a company like Intel that has $30-plus billion in revenue just in PCs, they are quite small. They would have asked in addition to Apple if Iris allowed them to expand their market.

Also, Lakefield-type chips might be the future of computing. The hybrid configuration isn't merely to save idle power, that logic is almost nonsensical. These types of chips can have few oversized cores plus sea of small cores for maximum single and multi thread performance in a constrained TDP. Rather than having 8 Icelake cores, you can have 2 Icelake+(in other words, much larger than Icelake) and 12-16 Tremont cores.

Their Platform 2015 presentation many years ago may be a reminder that they might have had these chips much earlier if 14/10nm didn't get delayed.

It may be just 7W mobile now, but we might see it expanded all the way to the desktop.

But has anyone confirmed core counts on an ICL-U/Y for a future shipping product?

The Icelake-Y chip was leaked by chrisdar as having 4 cores, 5.2W TDP and using LPDDR4-3733.

I was actually thinking that, if Intel was still having problems with bad EUs due to yields, that they could build in 72 or 80 of them (like 2xGT2 from Cannonlake) and disable down to 64 for market segmentation purposes

Or, or, maybe they are actually good to go. There's always the possibility that they are still fooling themselves and 10nm is garbage, but that's the prevailing opinion outside Intel, so let's look at alternatives. Maybe after 4 years they are at a point where a 64EU GPU can be shipped.

Look at history. They are full of surprises because things the majority did not expect happened.

 

[DrMrLordX]

Well I mean, I did sort of expect 10nm to work eventually. Maybe not 2019/2020 eventually but eventually.

Still it was really weird that Intel could ship an i3-8121U that worked, but that they had to disable the entire iGPU to do so. That's not a small thing. AND they went well out of their way to explain that the next gen of Intel iGPUs would be shippable with some flawed EUs in there. It makes sense to assume that they might still have some problems with getting all the EUs working perfectly, all the time. Otherwise there would have been no reason to make such a change to their iGPU lineup.

 

[jpiniero]

There's always the possibility that they are still fooling themselves and 10nm is garbage, but that's the prevailing opinion outside Intel, so let's look at alternatives. Maybe after 4 years they are at a point where a 64EU GPU can be shipped.

Not fooling themselves, but fooling Wall Street. You can still get a ton of chips even with horrifically bad yield, but you are talking in the thousands to hundreds of thousands of good processors compared to the 67M PCs that were sold in the Q3 which were mostly Intel. They will come clean eventually, but only then.

 

[dark zero]

Makes me wonder... Intel would make a return to the phone market or that is deader than dead with this new configuration?

 

[IntelUser2000]

Otherwise there would have been no reason to make such a change to their iGPU lineup.

We have to assume if Icelake comes as they say it will be, then whatever problems they were having with Cannonlake got resolved with Icelake. Cannonlake is 10nm, while Icelake is 10nm+.

You can still get a ton of chips even with horrifically bad yield,

No, just... no.

Makes me wonder... Intel would make a return to the phone market or that is deader than dead with this new configuration?

Entering the Android phone market has bigger problems than simply having a good chip. It's going against an entrenched ecosystem.

Intel's former CTO Pat Gelsinger explained it well. There are only two scenarios where a newcomer can penetrate an entrenched market. One, if the players in the entrenched markets are doing absolutely terribly. Two, if the newcomer is out-of-this-world better. I'm not talking 20-30% better. I'm talking 5-7 years ahead. Think Broxton in 2010.

It's in practice impossible for Intel to enter this market. It won't prevent them from trying though. And whatever miniscule goodwill they had got burnt when they cancelled Broxton.

Paul Otellini = Burned Intel's chances of taking the majority market in mobile
Brian Kraznich = Burned whatever little was left

 

[jpiniero]

No, just... no.

At 40 mm2 (lets say that's Lakefield's 10 nm die size) you still get roughly 300 pristine chips per wafer even at 4 defects/sqcm (horrific). That's a mill and half product at 5k wafers.

 

[Shivansps]

Makes me wonder... Intel would make a return to the phone market or that is deader than dead with this new configuration?

Doubt it, this is most likely a niche product of a basicly a tech demo.

 

[NTMBK]

I wonder if this chip is for Google? It would make perfect sense for a Chrome OS tablet, given that the current gen uses Intel.

 

[Yotsugi]

I wonder if this chip is for Google? It would make perfect sense for a Chrome OS tablet, given that the current gen uses Intel.

Apple/MS.

 

[NTMBK]

Apple/MS.

Microsoft maybe. Apple I would expect to use one of their own designs, instead of helping create a chip that their competitors can use.

 

[Dayman1225]

Microsoft maybe. Apple I would expect to use one of their own designs, instead of helping create a chip that their competitors can use.

Crystal Well comes to mind

 

[NTMBK]

Crystal Well comes to mind

Sure, but 2013 was a different era for Apple. They had only just released the A7 CPU, and there was a huge gulf between what they could achieve and what Intel could achieve. I don't think they're in the same place now that they have a chip as powerful as Intel's best Ultrabook parts.

 

[jpiniero]

Now Intel wouldn't be able to deliver enough volume for Apple's liking. MS yes.

 

[DrMrLordX]

We have to assume if Icelake comes as they say it will be, then whatever problems they were having with Cannonlake got resolved with Icelake. Cannonlake is 10nm, while Icelake is 10nm+.

I hope so, but until we actually see 10nm+ parts shipping in volume, it's all up in the air.

No, just... no.

Entering the Android phone market has bigger problems than simply having a good chip. It's going against an entrenched ecosystem.

Just wanted to say that the last two lines you responded to were not quotes from me . . . no big deal. Just wanted to clarify.

 

[IntelUser2000]

Just wanted to say that the last two lines you responded to were not quotes from me . . . no big deal. Just wanted to clarify.

Yea, I know that. Sorry for the confusion. I use reply for the first response but I just quote them by itself. I read past responses before responding, I figure most people do the same so they'll figure out the sentences were from them.

 

[insertcarehere]

Whatever the original request and usecase is for the hybrid Intel SoC being shown off, the resulting configuration puts it right in the firing line of Qualcomm 8cx, it'd be interesting to see how both products stack up to each other.

 

[beginner99]

Makes me wonder... Intel would make a return to the phone market or that is deader than dead with this new configuration?

Intels problem was and is also on the software side which is optimized for ARM. it's the opposite ARM faces in servers were most software is x86.

 

[Charlie22911]

Portable XBox or a PSP/PS Vita successor, with the ability to play console titles on the go with minimal work on the developers part; that would be cool. A bit unrealistic though I’m sure.

 

[Arachnotronic]

Portable XBox or a PSP/PS Vita successor, with the ability to play console titles on the go with minimal work on the developers part; that would be cool. A bit unrealistic though I’m sure.

The possibilities for this kind of chip are pretty wild actually. We'll see what OEMs actually do with it.

 

[jpiniero]

https://pbs.twimg.com/media/DwWcR8rUwAEKuqX.jpg

That is what the Lakefield board looks like.

 

[NTMBK]

And Lakefield name is confirmed: https://www.anandtech.com/show/1377...m-hybrid-x86-foveros-chip-is-called-lakefield