Discussion Intel current and future Lakes & Rapids thread

[Dayman1225]

Intel finally launches Icelake Xeons. 20%+ IPC, upto 40c/80t and 270w.

Intel 3rd Gen Xeon Scalable (Ice Lake SP) Review: Generationally Big, Competitively Small - AnandTech

Intel Xeon Ice Lake Edition Marks the Start and End of an Era - ServeTheHome

Intel Launches 3rd Gen Ice Lake Xeon Scalable - Wikichip

As part of the early ship program for Icelake they have already shipped 200k+ units.

 

[Gideon]

Intel finally launches Icelake Xeons. 20%+ IPC, upto 40c/80t and 270w.

Intel 3rd Gen Xeon Scalable (Ice Lake SP) Review: Generationally Big, Competitively Small - AnandTech

Intel Xeon Ice Lake Edition Marks the Start and End of an Era - ServeTheHome

Intel Launches 3rd Gen Ice Lake Xeon Scalable - Wikichip

As part of the early ship program for Icelake they have already shipped 200k+ units.

Can't wait till we have Alder Lake and Sapphire Rapid for really cutthroat competition in (most likely) 2022H1

 

[eek2121]

Can't wait till we have Alder Lake and Sapphire Rapid for really cutthroat competition in (most likely) 2022H1

Definitely a solid step up for Intel. Still a bit behind AMD with regards to efficiency.

 

[jpiniero]

It's definitely better than Skylake Server so that's a plus. The single core frequency regression is not a big deal in servers but does make it obvious that any HEDT/W processors wouldn't really make sense.

 

[moonbogg]

That's not how I remember this happening. I recall that the local GFA (Geekbench Fetish Association) was pretty adamant that RKL would be this gaming champion riding on the back of some absurd sounding IPC gains. So I'd say that the surprise was more like negative, and that came with the first real i7 "leaks".

When I said i7 I meant Nehalem.

 

[firewolfsm]

Being fabbed on 10SF, which should be comparable to TSMC 7nm in density and efficiency, I think the Icelake Xeon review makes it pretty clear that Intel's newest cores are simply less efficient than current Zen. Tigerlake cores wouldn't change the equation much there.

 

[jpiniero]

Being fabbed on 10SF, which should be comparable to TSMC 7nm in density and efficiency, I think the Icelake Xeon review makes it pretty clear that Intel's newest cores are simply less efficient than current Zen. Tigerlake cores wouldn't change the equation much there.

It's not on 10SF. They might respin it to 10SF/10ESF if the rumor that SPR is high end only, but that isn't the case right now.

 

[uzzi38]

It's not on 10SF. They might respin it to 10SF/10ESF if the rumor that SPR is high end only, but that isn't the case right now.

Speaking of which:

https://twitter.com/i/web/status/1379467917738184707

Just from the looks of this there are clearly lower end dies (24c 225W here) but given it's listed as the "lowest volume" I'd say that there's a good chance SPR's main focus will be on SKUs that sit above ICL-SP.

 

[jpiniero]

Speaking of which:

https://twitter.com/i/web/status/1379467917738184707

Just from the looks of this there are clearly lower end dies (24c 225W here) but given it's listed as the "lowest volume" I'd say that there's a good chance SPR's main focus will be on SKUs that sit above ICL-SP.

I was thinking that SPR would replace most or all of Plat and this ICL respin would be Gold and below. That wouldn't stop them from doing some lowerish core count models in Plat.

You're looking at what, 800+ mm2 of silicon, for SPR?

 

[DrMrLordX]

As part of the early ship program for Icelake they have already shipped 200k+ units.

Intel has a funny definition of "early".

Definitely a solid step up for Intel. Still a bit behind AMD with regards to efficiency.

. . . and raw MT performance, which makes a big difference in the server room. It isn't beating out the EPYC 7742 much less EPYC 7763. Intel's low core count per socket is biting them in the posterior, and yet they're burning a lot of power to achieve those results anyway.

 

[lobz]

When I said i7 I meant Nehalem.

Oh in that case I take it back! Those were fun times

 

[uzzi38]

I was thinking that SPR would replace most or all of Plat and this ICL respin would be Gold and below. That wouldn't stop them from doing some lowerish core count models in Plat.

You're looking at what, 800+ mm2 of silicon, for SPR?

I wouldn't be surprised if it's significantly more than that. Assuming that you have all of the dies on package ofc.

 

[jpiniero]

Intel has a funny definition of "early".

Haven't they been sampling/selling Ice Lake Server since like Q2 2019?

 

[Dayman1225]

Haven't they been sampling/selling Ice Lake Server since like Q2 2019?

AFAIK the data for 200k is from end of year 2020/ Jan 2021 -> Now

 

[Hulk]

As expected, supply shortage is everywhere, even Intel is not lucky one

Exclusive: Demand To Exceed Supply For Intel Rocket Lake Due To Substrate Shortage

It would appear that gamers can't catch a break. Products that rely on TSMC, such as AMD parts, have had tight supply for a while now (with a few exceptions). Intel CPUs, considering they are built in their own foundry, have had a free run of the proverbial house for a while now and CPU […]

wccftech.com

Help me out here. The article reads that Intel will allocating more i9 parts than i7 and i5 so there will be increased shortages for the lower end parts. If this is true then Intel's yields must be fab-ulous, right? If a company is selling every part it manufactures and each part is binned then sales are a direct correlation to yields I assume? But then again they've had 6 years to get 14nm working properly so I guess I shouldn't be surprised.

 

[Bouowmx]

The lower-end stack of Ice Lake Xeon looks kinda ok for HEDT? Though, it's not happening.
4310 -> 'i9-11900X': 12 cores, $501
4314 -> 'i9-11920X': 16 cores, $694
4316 -> 'i9-11940X': 20 cores, $1002
5318Y -> 'i9-11960X': 24 cores, $1273
6330 -> 'i9-11980XE': 28 cores, $1894 (this core count is where the HCC die tops out)

 

[DrMrLordX]

Haven't they been sampling/selling Ice Lake Server since like Q2 2019?

Sampling? Sure. There have been Ice Lake-SP results in GB5 for some time, in core counts all the way down to 6 (if I recall correctly). Selling . . . ? Not sure. I'm sure Facebook might have gotten some early silicon, but who knows?

 

[JoeRambo]

Not sure why would anyone want an early sample. Not after ZEN2 based EPYC for sure. If you are hyper converger/consolidator/cloud monster without in-house ARM cpu program, you are probably already rocking some high core AMD server. Or early sampling next get ARM stuff?

Now if you just need certain capacity server, things not that interesting either. Performance increase is so tiny, that early adopter problems associated with "sampling" make whole thing moot. Why would someone trade 10-15% of clock for 10-15 IPC increase?

Pricing? Say you were buying 6248R, for 2700$ ( with whatever discount, but lets assume they are const for IceLake-SP). What does your dollars get you? ~6348 for $3072. 4 more cores and clock regressions across the board. 53XX Golds are no-go due to loosing clock in Ghz range, even if tasty priced.

 

[Dayman1225]

Not sure why would anyone want an early sample. Not after ZEN2 based EPYC for sure. If you are hyper converger/consolidator/cloud monster without in-house ARM cpu program, you are probably already rocking some high core AMD server. Or early sampling next get ARM stuff?

As much as I agree the market simply does not work like that. Intel is the incumbent - people will buy intel because they are Intel. Inertia in DC is next level and upgrade cycles are much longer than what me and you are used to. I’d be willing to bet that ICL-SP out ships Rome and Milan combined.

 

[JoeRambo]

As much as I agree the market simply does not work like that.

I was commenting about wanting early samples, as Skylake-SP was early sampled to hyperscaled for almost a year before public release. No one except big server vendors "want" ICL-SP samples this time + i am sure they have been getting some since 2019?
And ICL-SP will outship anything AMD has to offer, even without incumbent status, unlike people around this forum believe, not every CPU InteL or AMD ship are 64 or 40 core halo products. Just like on ther desktop, if 64C TR or 16C 5950X exists, people are still buying 6-8C CPUs in epic volumes in all segments.

 

[eek2121]

Intel has a funny definition of "early".



. . . and raw MT performance, which makes a big difference in the server room. It isn't beating out the EPYC 7742 much less EPYC 7763. Intel's low core count per socket is biting them in the posterior, and yet they're burning a lot of power to achieve those results anyway.

Sapphire Rapids will close that gap considerably.

EDIT: According to rumors, SR will feature up to 56 golden cove cores.

 

[Edrick]

The lower-end stack of Ice Lake Xeon looks kinda ok for HEDT? Though, it's not happening.
4310 -> 'i9-11900X': 12 cores, $501
4314 -> 'i9-11920X': 16 cores, $694
4316 -> 'i9-11940X': 20 cores, $1002
5318Y -> 'i9-11960X': 24 cores, $1273
6330 -> 'i9-11980XE': 28 cores, $1894 (this core count is where the HCC die tops out)

I think the next HEDT we will see from Intel will be SPR. And honestly, I am quite excited about it.

 

[Ajay]

Sapphire Rapids will close that gap considerably.

EDIT: According to rumors, SR will feature up to 56 golden cove cores.

Only 56?! The package footprint looks huge. I also thought that SPR will have two CPU dice per package. Seems a bit sad that Intel still won't get to 64 cores with that setup. I suppose that I/O and the mem controller will be pretty large on those dice.

 

[jpiniero]

Only 56?! The package footprint looks huge. I also thought that SPR will have two CPU dice per package. Seems a bit sad that Intel still won't get to 64 cores with that setup.

It's four. The client Golden Cove is probally going to be very large as it is, plus you have to add the second AVX-512 unit plus AMX and all the server fun stuff.

 

[Ajay]

It's four. The client Golden Cove is probally going to be very large as it is, plus you have to add the second AVX-512 unit plus AMX and all the server fun stuff.

Thanks. Forgot about the second AVX unit and AMX - going senile I guess. Intel is really doubling down on ML/DL.