Intel Icelake Server delayed

[Taf the Ghost]

This one is a little messy, as it's information you glean from what isn't mentioned.

 

At a JP Morgan conference ( full transcript ) , Intel's Murthy Renduchintala said,

Quote

And so, therefore we’re comfortable with the 14-nanometer roadmap that will give us leadership products in the next 12 to 18 months, as we seek to optimize the cost structure and yields of our 10-nanometer portfolio.

What this means is that Intel's release in the Server products this year (called Cascade Lake-SP) will be their only Server release until Icelake-S in either very late 2019 or early 2020. And that's at the current roadmap, which has already slipped by years.

 

https://twitter.com/TMFChipFool/status/996480113901539349

Quote

7nm Rome vs 14nm Cascade Lake. RIP, Intel DCG share & margins. Wow.

 

What this means is, beyond 10nm from Intel being a disaster, is that AMD has the opportunity to be massively ahead of Intel in the Server space, if AMD can launch the Zen2-based "Rome" parts in 2019. That's a pretty big deal, as it would mean AMD would actually have a node advantage for the first time on Intel, along with a much larger number of cores on their socket. At this time, however, we don't know what this means for Icelake-S, the Desktop version, of the Icelake on 10nm+ node. H2 2019 is looking far more likely right now. Roughly in the same window Coffee Lake-S launched, along with the rumored release window for the 8c Coffee Lake parts. (Coffee Lake-R is the current rumored name.)

[samcool55]

I guess this shows how quickly stuff can change.

less than 2 years ago amd looked basically dead, right now it's already much better and it seems like both companies will have to battle it out. Good for us i guess!

[Tellos]

Just means like they have been saying for consumer 10nm delayed AGAIN. Intel is suffering baddly for this because long term they may have to just literily SKIP it and eat the losses.

[Princess Luna]

5 minutes ago, Taf the Ghost said:

if AMD can launch the Zen2-based "Rome" parts in 2019. That's a pretty big deal, as it would mean AMD would actually have a node advantage for the first time on Intel,

Well I don't know, kinda pointless AMD get on 7nm if such node shrink only allowed them to reach Intel's numbers on 14nm which's the double of the lithography... so eh... don't really agree much here.

[xentropa]

AMD 7nm = Intel 10nm

[samcool55]

2 minutes ago, Princess Cadence said:

Well I don't know, kinda pointless AMD get on 7nm if such node shrink only allowed them to reach Intel's numbers on 14nm which's the double of the lithography... so eh... don't really agree much here.

Afaik the 7nm node that AMD will use is basically the same as 10nm Intel will use.

I don't know why exactly that is but that's what some websites said...

[Princess Luna]

9 minutes ago, samcool55 said:

.

Not really, Intel silicon is usually way more polished than AMD's the reason why 10nm is taking so long because Intel really is trying to make it the best node possible, I doubt AMD can match it's potential when done even being on 7nm, reason behind Intel's current 14nm still being so ahead of AMD's 14nm and now 12nm... lithography is just part of a bigger picture:

as you can see Intel's density is considerable ahead and this has a straight impact on IPC.

 

I really think the media is making a much bigger fuzzle over this than needed... better delay but end up with the best product again than rush just to make a "lithography racing"

[Tellos]

@Princess Cadence I'd agree normally, cept this is the third delay I beleive. Also IPC gains are not whats got intel ahead or better silcon, it's clockspeed. But thats beginning to close up.

[leadeater]

23 minutes ago, samcool55 said:

Afaik the 7nm node that AMD will use is basically the same as 10nm Intel will use.

I don't know why exactly that is but that's what some websites said...

There's lots of different measurements and gap sizes, pitches etc in a silicon die. Saying 7nm or 10nm is an extremely basic way of representing it and is the smallest size present in a die but there are also other factors around it to. Another part of is is not everything in the die is the same actual node size, cache can be bigger. It's really complicated and hard to understand fully.

 

Quote

One simple metric is gate pitch (gate width plus spacing between transistor gates) multiplied by minimum metal pitch (interconnect line width plus spacing between lines), but this doesn’t incorporate logic cell design, which affects the true transistor density. Another metric, gate pitch multiplied by logic cell height, is a step in the right direction with regard to this deficiency. But neither of these takes into account some second order design rules. And both are not a true measure of actual achieved density because they make no attempt to account for the different types of logic cells in a designer’s library. 

https://newsroom.intel.com/editorials/lets-clear-up-node-naming-mess/

 

Quote

Every chip maker, when referring to a process node, should disclose its logic transistor density in units of MTr/mm2 (millions of transistors per square millimeter) as measured by this simple formula. Reverse engineering firms can readily verify the data.

 

[leadeater]

21 minutes ago, Princess Cadence said:

Intel silicon is usually way more polished than AMD's the reason why 10nm is taking so long because Intel really is trying to make it the best node possible, I doubt AMD can match it's potential when done even being on 7nm, reason behind Intel's current 14nm still being so ahead of AMD's 14nm and now 12nm... lithography is just part of a bigger picture:

I believe in the CPU core areas of the die Zen is actually more dense than Intel was at launch, there were a few articles covering it I'll see if I can find them and edit this post. 

 

Edit:

 

https://www.anandtech.com/show/11170/the-amd-zen-and-ryzen-7-review-a-deep-dive-on-1800x-1700x-and-1700

 

Intel crushes on cache die area efficiency though.

[Taf the Ghost]

10 minutes ago, Princess Cadence said:

Not really, Intel silicon is usually way more polished than AMD's the reason why 10nm is taking so long because Intel really is trying to make it the best node possible, I doubt AMD can match it's potential when done even being on 7nm, reason behind Intel's current 14nm still being so ahead of AMD's 14nm and now 12nm... lithography is just part of a bigger picture:

as you can see Intel's density is considerable ahead and this has a straight impact on IPC.

 

I really think the media is making a much bigger fuzzle over this than needed... better delay but end up with the best product again than rush just to make a "lithography racing"

For CPU nodes, it's really the SRAM density that's going to matter. That takes up the majority of the die space, especially with x86 being in such a place where you're really just tweaking things.

[Princess Luna]

3 minutes ago, leadeater said:

Intel crushes on cache die area efficiency though.

Well Ring Bus destroys Infinity Fabric alright hehe, all in all I still stand for a simple fact that AMD is doing no miracle, the technology is there and they are doing what they can to catch up, literally all the cards are being played with this 7nm node.

 

It doesn't automatically turn Intel into a failure, only time can tell, it's all too soon we might end up with AMD delaying just as much at some point when Intel manages to move on and the story turns all around... the issue of speculation.

[porina]

7 minutes ago, leadeater said:

I believe in the CPU core areas of the die Zen is actually more dense than Intel was at launch, there were a few articles covering it I'll see if I can find them and edit this post. 

 

Without looking for a reference, isn't part of the improvement going from Intel's original 14nm to 14+ or 14++ was an increase in size to improve on power characteristics? Smaller might be better for density and maybe reducing costs, but it can be more complicated than that as an overall product balance.

[leadeater]

8 minutes ago, Princess Cadence said:

Well Ring Bus destroys Infinity Fabric alright hehe, all in all I still stand for a simple fact that AMD is doing no miracle, the technology is there and they are doing what they can to catch up, literally all the cards are being played with this 7nm node.

 

It doesn't automatically turn Intel into a failure, only time can tell, it's all too soon we might end up with AMD delaying just as much at some point when Intel manages to move on and the story turns all around... the issue of speculation.

Well considering the CPU cores only make up like half the die taking a small win there doesn't do much if you're 50% (random number) worse in the other half of the die area.

 

Advantage will always go Intel on the fab side, 100% control and ownership means a lot when you're fighting for 1% gains.

[laminutederire]

16 minutes ago, Princess Cadence said:

Not really, Intel silicon is usually way more polished than AMD's the reason why 10nm is taking so long because Intel really is trying to make it the best node possible, I doubt AMD can match it's potential when done even being on 7nm, reason behind Intel's current 14nm still being so ahead of AMD's 14nm and now 12nm...

It may be the case, but the most probable explanation is that AMD can jump onto the 7/10nm train with processes having poorer yields, solely because of the way they modularly build their cpus. Since Intel is still doing monolithic dies, it gives them more performance, but it also makes them unable to mass produce their 10nm Cpus, because their suffer a lot from poor yields. That's also part of why threadripper and epyc are soooooo much cheaper than Intel's counterpart. Yes the Intel sticker adds a premium, but it is sometimes a factor of two between the prices, and that's because Intel probablu cannot really push their prices down for rentability reasons.

[leadeater]

9 minutes ago, porina said:

Without looking for a reference, isn't part of the improvement going from Intel's original 14nm to 14+ or 14++ was an increase in size to improve on power characteristics? Smaller might be better for density and maybe reducing costs, but it can be more complicated than that as an overall product balance.

No idea, didn't really follow any of the specifics for 14nm/14nm+/14nm++, I just assume they made each one more better in some way otherwise why bother .

[porina]

1 minute ago, leadeater said:

No idea, didn't really follow any of the specifics for 14nm/14nm+/14nm++, I just assume they made each one more better in some way otherwise why bother .

Yup, but in context of this thread, was pointing out more better doesn't necessarily have to mean smaller.

[Taf the Ghost]

2 minutes ago, laminutederire said:

It may be the case, but the most probable explanation is that AMD can jump onto the 7/10nm train with processes having poorer yields, solely because of the way they modularly build their cpus. Since Intel is still doing monolithic dies, it gives them more performance, but it also makes them unable to mass produce their 10nm Cpus, because their suffer a lot from poor yields. That's also part of why threadripper and epyc are soooooo much cheaper than Intel's counterpart. Yes the Intel sticker adds a premium, but it is sometimes a factor of two between the prices, and that's because Intel probablu cannot really push their prices down for rentability reasons.

@Princess Cadence

 

The issue isn't that AMD will have an all-encompassing Node lead. They won't, though all of the future nodes are looking to be fairly similar when compared. The issue is at Intel's current roadmap, AMD might be on the new node for between 3-12 months over Intel. And they will have a very competitive product on a per-core basis. Then, AMD is going to have more cores per Socket than Intel. 

 

And all of this is predicated on Intel actually getting 10nm+ out at Volume by late 2019. We're still not sure that's going to happen.

[Brooksie359]

44 minutes ago, Princess Cadence said:

Well Ring Bus destroys Infinity Fabric alright hehe, all in all I still stand for a simple fact that AMD is doing no miracle, the technology is there and they are doing what they can to catch up, literally all the cards are being played with this 7nm node.

 

It doesn't automatically turn Intel into a failure, only time can tell, it's all too soon we might end up with AMD delaying just as much at some point when Intel manages to move on and the story turns all around... the issue of speculation.

If they can close the gap with 7nm it will be huge. I mean ryzen is already a good product with one of the main drawbacks being that it's still a fair bit behind Intel in gaming. If they can get a decent increase in clock speed as well as some architectural improvements it possible that gap could close. I think it will be exciting to see how it all plays out and it will be good if AMD closes the gap. That would cause Intel to lower prices if they no longer have the performance advantage. 

[Taf the Ghost]

3 minutes ago, Brooksie359 said:

If they can close the gap with 7nm it will be huge. I mean ryzen is already a good product with one of the main drawbacks being that it's still a fair bit behind Intel in gaming. If they can get a decent increase in clock speed as well as some architectural improvements it possible that gap could close. I think it will be exciting to see how it all plays out and it will be good if AMD closes the gap. That would cause Intel to lower prices if they no longer have the performance advantage. 

Even the gaming has closed quite a bit. Main difference now is simply clocks & optimizations. There's no IPC difference in gaming between the companies. It's all about certain I/O latencies. See Steve's video for a lot of benchmarks to prove the point. https://www.youtube.com/watch?v=AUyF--fJaaM 

 

The IPC difference between AMD & Intel is in the FP calculations. If you leverage things in the correct way, there's up to 30% IPC advantage for Intel. We'll see what AMD does with Zen2, though most are expecting AVX2 units. But it won't necessarily matter that much if AMD has 64 core parts vs Intel's 32 core parts for upwards of a year before Intel brings any IPC/IP improvements.

 

[Trixanity]

1 hour ago, Princess Cadence said:

Well Ring Bus destroys Infinity Fabric alright hehe, all in all I still stand for a simple fact that AMD is doing no miracle, the technology is there and they are doing what they can to catch up, literally all the cards are being played with this 7nm node.

 

It doesn't automatically turn Intel into a failure, only time can tell, it's all too soon we might end up with AMD delaying just as much at some point when Intel manages to move on and the story turns all around... the issue of speculation.

That's a half-truth if I ever saw one. It's not like a ring bus is anything new and AMD could have gone down that route but ring buses don't scale very well. Infinity Fabric allows scaling at the cost of latency. The economics alone are fantastic for a company like AMD. Would AMD have gone with a ring bus if they had five times the size they have now? Perhaps. However Intel has recognized that a ring bus isn't the future for massive core count chips hence the mesh topology on newer high core count chips. 

 

And "this 7nm node" isn't really singular. AMD has three potential foundry partners to play with: Globalfoundries, TSMC and Samsung. And that goes for every single chip they wish to tapeout. So I don't really see how you could claim AMD is banking on something that may not pan out. They're fabless unlike Intel and therefore have a somewhat flexible playbook (should be noted that AMD has a wafer agreement with Globalfoundries that requires a minimum of orders with the option to pay a 'fine' for not ordering enough).

 

Both TSMC and Globalfoundries appear to be on time and with good numbers and AMD seems to have orders with both.

The reason people call Intel's 10nm a failure is it was supposed to be out in 2016. Then 2017, then 2018 and now it's 2019 unless you'd call their recent Cannon Lake chips in small quantities with broken GPUs onboard an actual launch. I mean sure it's a launch of some kind that they can prop up for their investors but in any useful sense it's not. And the initial 10nm process isn't meant to be better than 14nm - just smaller with lower power but no performance gains. Good for small chips like this tiny i3 but not useful for a server chip or a 8700k successor.

 

The biggest worry for AMD is their GPU roadmap being pretty bleak and whether Zen2 and beyond can live up to expectations. Nodes aren't their business anymore.

[Brooksie359]

51 minutes ago, Taf the Ghost said:

Even the gaming has closed quite a bit. Main difference now is simply clocks & optimizations. There's no IPC difference in gaming between the companies. It's all about certain I/O latencies. See Steve's video for a lot of benchmarks to prove the point. https://www.youtube.com/watch?v=AUyF--fJaaM 

 

The IPC difference between AMD & Intel is in the FP calculations. If you leverage things in the correct way, there's up to 30% IPC advantage for Intel. We'll see what AMD does with Zen2, though most are expecting AVX2 units. But it won't necessarily matter that much if AMD has 64 core parts vs Intel's 32 core parts for upwards of a year before Intel brings any IPC/IP improvements.

 

Yeah the main reason they are behind is latency and clock speeds. 4.2 vs 5ghz is a pretty big gap. 

[Taf the Ghost]

18 minutes ago, Trixanity said:

That's a half-truth if I ever saw one. It's not like a ring bus is anything new and AMD could have gone down that route but ring buses don't scale very well. Infinity Fabric allows scaling at the cost of latency. The economics alone are fantastic for a company like AMD. Would AMD have gone with a ring bus if they had five times the size they have now? Perhaps. However Intel has recognized that a ring bus isn't the future for massive core count chips hence the mesh topology on newer high core count chips. 

 

And "this 7nm node" isn't really singular. AMD has three potential foundry partners to play with: Globalfoundries, TSMC and Samsung. And that goes for every single chip they wish to tapeout. So I don't really see how you could claim AMD is banking on something that may not pan out. They're fabless unlike Intel and therefore have a somewhat flexible playbook (should be noted that AMD has a wafer agreement with Globalfoundries that requires a minimum of orders with the option to pay a 'fine' for not ordering enough).

 

Both TSMC and Globalfoundries appear to be on time and with good numbers and AMD seems to have orders with both.

The reason people call Intel's 10nm a failure is it was supposed to be out in 2016. Then 2017, then 2018 and now it's 2019 unless you'd call their recent Cannon Lake chips in small quantities with broken GPUs onboard an actual launch. I mean sure it's a launch of some kind that they can prop up for their investors but in any useful sense it's not. And the initial 10nm process isn't meant to be better than 14nm - just smaller with lower power but no performance gains. Good for small chips like this tiny i3 but not useful for a server chip or a 8700k successor.

 

The biggest worry for AMD is their GPU roadmap being pretty bleak and whether Zen2 and beyond can live up to expectations. Nodes aren't their business anymore.

AMD is mostly all-in on jumping to the next node in the GPU space. GCN just doesn't scale up as well as Nvidia's base CUDA approach, so they're looking to carve out in the middle until they can move to an entirely new Architecture.

[Taf the Ghost]

5 minutes ago, Brooksie359 said:

Yeah the main reason they are behind is latency and clock speeds. 4.2 vs 5ghz is a pretty big gap. 

Actually, as quite a number of 2017/2018 games have shown, the clockspeeds don't matter at even medium quality settings. It's back to being about I/O more than anything else. Somehow, we've add enough cores to circle back the problems of the Single Core days.

[Taf the Ghost]

3 minutes ago, VegetableStu said:

can anyone remember seeing a comparison table of lithography definitions between silicon fabs? I vaguely remember such a table existed but couldn't figure out exactly where ,_,

 

it spells out intel and goflo etc etc etc

Probably thinking of an Anandtech article, but I don't know which one. Wikichip has their list, but you have to go to 10nm to compare Intel's.

 

https://en.wikichip.org/wiki/7_nm_lithography_process