Discussion Intel current and future Lakes & Rapids thread

[PeterScott]

High volume manufacturing of 10nm is officially delayed into 2019.

I wouldn't want to be the CEO that squandered both Intel's CPU IPC lead, and it's Fab process lead, at the same point in time...

Though it is bit early to start writing Intel's obituary, as some seem ready to do.

Intel could easily have clear cut mainstream socket leadership just by releasing 8C Coffee Lake.

That will hold them in good position until AMD releases the next generation 7nm Ryzen parts, which I expect will ship early 2019, at which point it seems unlikely Intel 10nm will be ready for high performance parts. Then things get VERY interesting.

Fireworks start in 2019.

 

[mikk]

That will hold them in good position until AMD releases the next generation 7nm Ryzen parts, which I expect will ship early 2019

You will be disappointed if you expect early 2019 for Ryzen 2.

 

[witeken]

You will be disappointed if you expect early 2019 for Ryzen 2.

What do you think as release date?

 

[Thunder 57]

What do you think as release date?

I think they'll have it out in time for "back to school"; Hopefully well before though.

 

[PeterScott]

You will be disappointed if you expect early 2019 for Ryzen 2.

I think it depends on which Fab they are using. If AMD's contracts force them to use GF for Ryzen 2, then I don't think it will be early 2019.

If they use TSMC, it shouldn't be much of an issue. TSMC has already started Volume 7nm production.

 

[jpiniero]

That will hold them in good position until AMD releases the next generation 7nm Ryzen parts, which I expect will ship early 2019, at which point it seems unlikely Intel 10nm will be ready for high performance parts. Then things get VERY interesting.

At this point you kind of have to assume you won't see anything HVM with 10nm unless it's super tiny (eg: EMIB) or something radically changes at 10++ to improve yield.

 

[french toast]

I don't know if Intel will be screwed (any more than they already are). They must have their engineers pulling alot of overtime making the necessary changes to the process and possibly the silicon implementation to get Ice Lake out in HVM.

It does look like the biggest change in years for the core architecture: they actually increased caches that have been the same size since Nehalem, only Skylake with AVX 512 dared rebalance them going for more L2 at the cost of L3 area.
By the same logic there must be more on the execution side in Icelake that requires some fundamental growth in memory to keep it feed: it's a balance of latency and size that bottlenecks a core and this time size is winning. I don't think Skylake is so close to that bottleneck to require 50% more L1 in order to achieve a couple% more IPC, we'll see what they have managed...

The thing is, Intel does have about 3 years of engineering IP and innovation to put into icelake (probably more as it is so delayed)...on the negative side, the design was likely completed years ago (2016) and is just sat on ice ( missed the pun) being held back by foundry troubles, or they could be updating it as they go along..expensive but the best bet and they certainly have the time/resources/engineering talent to do just that.

For their sake I hope this is not a 2016 product put on ice and launched 3 years later on dodgy process, something tells me the must have kept updating the design as they go along... probably alot wider.
How do we know the size of the L1 cache?.

You will be disappointed if you expect early 2019 for Ryzen 2.

I think Q2 2019 looks a good bet...on a mature platform.
Of course yields will not be optimum so if they get a sniff that icelake is going to be late..they might delay to Q3.
I just hope both products have gone wider and innovative and that both processes deliver..what a battle this is going to be.

 

[SAAA]

The thing is, Intel does have about 3 years of engineering IP and innovation to put into icelake (probably more as it is so delayed)...on the negative side, the design was likely completed years ago (2016) and is just sat on ice ( missed the pun) being held back by foundry troubles, or they could be updating it as they go along..expensive but the best bet and they certainly have the time/resources/engineering talent to do just that.

For their sake I hope this is not a 2016 product put on ice and launched 3 years later on dodgy process, something tells me the must have kept updating the design as they go along... probably alot wider.
How do we know the size of the L1 cache?.
I think Q2 2019 looks a good bet...on a mature platform.
Of course yields will not be optimum so if they get a sniff that icelake is going to be late..they might delay to Q3.
I just hope both products have gone wider and innovative and that both processes deliver..what a battle this is going to be.

The only source there's left for L1 size is this geekbench entry: https://browser.geekbench.com/v4/cpu/2400363
So 50% more L1 data and it looks like L3 tripled too.
I too hope it's not just a 2016 design released late 2019, it would still be better than nth time a Skylake refresh, but they likely have included fixes for meltdown/spectre and some minor changes in these years.

 

[french toast]

Any thoughts on what turbo is? Single core is insane!.

 

[mikk]

I think Q2 2019 looks a good bet...on a mature platform.

What do you think as release date?

H1 2019 is doubtful imho, sometime in H2 2019 I guess. But with the typical delays in mind everything can happen. Everyone expected Ryzen 1 in 2016 for example, some even in H1 2016!

 

[ksec]

if Icelake is a new uArch, then what is Sapphire Rapid? uArch for Server?

Because I am still skeptic of Icelake given it has a "Lake" within its name. i.e Not much of an improvement.

It kind of make sense to have a uArch specially toiled for DC, now that it has reached 50% of total revenue, and are still growing, there are more DC built in China and India, as well as many upgrade cycle to Servers.

 

[ksec]

The only source there's left for L1 size is this geekbench entry: https://browser.geekbench.com/v4/cpu/2400363
So 50% more L1 data and it looks like L3 tripled too.
I too hope it's not just a 2016 design released late 2019, it would still be better than nth time a Skylake refresh, but they likely have included fixes for meltdown/spectre and some minor changes in these years.

This result has been flagged as inaccurate

 

[SAAA]

This result has been flagged as inaccurate

Yes the result is inaccurate (notice the impossibly high memory score) but the sku readings are valid, things like name, code and caches should be correct. Frequency I'm not too sure because the individual subscores are too low for a 2GHz cpu, maybe it was sitting at idle speed of 600 MHz or such or there were software bugs.

Regardless of this particular result I'm personally expecting at least 10% more IPC from Icelake over Skylake, less it would feel too little too late in 4-5 years of waiting...
Sapphire is the "tock" after Icelake, might be a bigger change as the name implies, but it depends if they really shake things up in the core or just keep going with incremental gains.

 

[jpiniero]

if Icelake is a new uArch, then what is Sapphire Rapid? uArch for Server?

There is (or at least was going to be) Icelake Server. Sapphire Rapid is intended to be the next server product after that. Client is getting various more Lakes.

 

[Eug]

Regardless of this particular result I'm personally expecting at least 10% more IPC from Icelake over Skylake, less it would feel too little too late in 4-5 years of waiting...
Sapphire is the "tock" after Icelake, might be a bigger change as the name implies, but it depends if they really shake things up in the core or just keep going with incremental gains.

Where is this 10% IPC increase estimate coming from?

Is it just a random guess or do you have a reason to think it is 10%?

What about clockspeeds then?

 

[Excessi0n]

Whenever I see a company going on with SJW nonsense, I know they are shooting themselves in the foot, and the next few years might show that BK has shot off both of Intel's feet.

That "SJW nonsense," as you put it, leads directly to a superior workforce.

 

[CHADBOGA]

That "SJW nonsense," as you put it, leads directly to a superior workforce.

LOL Thanks for the laugh.

 

[Gideon]

You will be disappointed if you expect early 2019 for Ryzen 2.

Yeah, I wouldn't be holding my breath for Zen 2 in 1H19, let alone 1Q19.

The data-points we have right now are, that 7nm vega is in the labs (wildly rumored to be the GF 7nm pipe-cleaner) and that AMD promises to sample 7nm CPUs in the very end of 2018.

So naturally we should compare that info to zen 1, which started sampling in 2Q16, and was expected to be released in very late 4Q16 and was subsequently delayed a few months.

Based on that, it seems like very-late Q2 at the earliest, with the usual quarterly delay quite possible.

That's just speculation of course. I would very much like to see Zen 2 earlier, just not getting my hopes up yet, without any further info. After all GloFo might screw up it's 7nm as well, and the architecture might need respins (which adds a couple months of delays). I mean we don't even have Engineering Samples yet ...

 

[IntelUser2000]

Where is this 10% IPC increase estimate coming from?

Is it just a random guess or do you have a reason to think it is 10%?

It's probably just a guess, but a quite reasonable one. Other than the Netburst to Core uarch transition, we didn't have huge perf/clock increases. Netburst was driven by factors other than just technical, so it wasn't Intel's best work.

Pentium III with 256KB cache to Pentium M Banias: 30%
Pentium M Banias to Core 2: 30%
Core 2 Penryn to Nehalem: 5-10%(Single thread)
Nehalem to Sandy Bridge: 15%
Ivy Bridge to Haswell: 10%
Broadwell to Skylake: 10%

We also saw a big gain with Ryzen over Bulldozer uarch chips because it was driven by similar reasons as Netburst. Easy pickings in improvements have come and gone long ago, so we shouldn't expect something big. I guess 15% over Skylake cores are possible if we say Cannonlake was supposed to bring 3-5% and Icelake does 10% over that. 10% number is thus based on historical data.

 

[ksec]

Well, they said we should expect 10% in Skylake, see what happen to that.

The data-points we have right now are, that 7nm vega is in the labs (wildly rumored to be the GF 7nm pipe-cleaner) and that AMD promises to sample 7nm CPUs in the very end of 2018.

I remember reading somewhere 7nm Vega is TSMC, not GF.

 

[Gideon]

Well, they said we should expect 10% in Skylake, see what happen to that.I remember reading somewhere 7nm Vega is TSMC, not GF.

Might be. Considering the schedules, I would really like Zen 2 to bee TSMC and Vega GloFo, but I doubt it would fly with the infamous agreement, as Vega generates way too few wafers.

EDIT:
Yes, Vega 20 is TSMC. Too bad, as that means that Global Foundries 7nm is still a big unknown till the end of this year

 

[plopke]

What most people here are forgetting , even if AMD has higher core count , IPC ,SMT , frequency , OC ability , stability , power ,price,.... The amount of people buying Intel will still be a lot higher , the few times AMD had better CPUs(price/performance), it was almost impossible to convince people to buy AMD. It was less of a problem for ATi/AMD, but for example when the RX470 went for 130 dollars , it was still hard to convince people to get it over a 1050(not the Ti !!!!). For mainstream people to catch on or replace 10 years of Intel CPUs/NVIDIA experience you have to be on top for a while until it catches on and even then (even without shady business practices).

So if Intel executes it well , I see no problem with a delay until 2019 , 7nm of the competition could surprise me but not sure if the jump from 12/14/16 -> 7/10/11/12 will be major until like 2019 itself.

 

[IntelUser2000]

plopke: Because decisions are based largely on trust. It's something that's gained over many years. This isn't just about the consumers that buy the system as a whole. It has to do with system integrators and OEMs too.

There are more companies that make Nvidia cards than AMD ones. As a company, you aim for the low-hanging fruit, which is making a product that's easiest to sell and has the highest total available market. That's Nvidia. Now this is just for video cards. For Intel vs AMD think about system integrators, motherboard companies, OEMs.

 

[Eug]

I'm still trying to get my head around the idea that Cannon Lake is nearly totally dead. If the jump is from Kaby Lake Y direct to Ice Lake Y, and Ice Lake might not come out until 2019 H2, that would mean 3 years between chip generations. Really?

What most people here are forgetting , even if AMD has higher core count , IPC ,SMT , frequency , OC ability , stability , power ,price,.... The amount of people buying Intel will still be a lot higher , the few times AMD had better CPUs(price/performance), it was almost impossible to convince people to buy AMD

AMD's Opteron became pretty damn popular after it came out. And here I'm not talking about overclockers in their mom's basements, but big corporations buying server farms.

 

[french toast]

4 years ^^