Discussion Intel current and future Lakes & Rapids thread

[uzzi38]

It's a fair point if you are correct but you can see the breakdown of the clock speeds. That's why median clocks are reported.

Look at the breakdown of the test I linked. All 3.8GHz.

Geekbench's "breakdowns" aren't done in heavy or even multi-core loads.

You can see the same thing if you check general MTS chips as well btw.

Matisse certainly can't hold 4.35GHz all-core at stock on a 3700X

https://browser.geekbench.com/v4/cpu/15534916.gb4

And this is normal for MTS results.

 

[Gideon]

Geekbench's "breakdowns" aren't done in heavy or even multi-core loads.

That's certainly true. Every time I've run the bench, the clocks listed are always near Single Core Turbo, never what is the average clock-speed on an all-core load. Overall Geekbench is a very bursty load as well, almost not taxing desktop processors compared to some other workloads.

 

[IntelUser2000]

Look at the breakdown of the test I linked. All 3.8GHz.

That's very interesting, thanks.

That just says bad benchmark is bad.

Overall Geekbench is a very bursty load as well, almost not taxing desktop processors compared to some other workloads.

Yup, it takes 3-4 seconds to run each test. Plus, Geekbench pretty much scales linearly with clocks meaning it fits in the core caches and doesn't really hit the memory system or even L3, similar to Dhrystone tests.

 

[lobz]

That's very interesting, thanks.

That just says bad benchmark is bad.



Yup, it takes 3-4 seconds to run each test. Plus, Geekbench pretty much scales linearly with clocks meaning it fits in the core caches and doesn't really hit the memory system or even L3, similar to Dhrystone tests.

that would make the 1.25 MB L2 a major advantage in Geekbench.

 

[Gideon]

BTW, somebody on twitter shared a slide that's apparently from Intel Russia's roadmap (though it seemed a bit dated).

Now if these year numbers are accurate .... EDIT: it seems they are not, Sapphire Rapid still stated for 2021Q4)

EDIT, and here is the complete slidedeck:

https://imgur.com/a/fq7bTFe

So Sapphire Rapid is still coming Q4 2021 ,supporst PCIe4 and DDR5 but will replace Ice-Lake only partiallly (its surprising that Ice Lake should totally succeed Cascade Lake, if this is to be believed):

 

[Gideon]

Anyway, if these slides are accurate, Intel plans to run with Ice-Lake for most mainstream servers at least up to 2023 (e.g. those that fit to Cascade Lake specs, e.g. probably <=32 cores) and Sapphire Rapid will only be an ultra-high-end offering (replaces Cooper Lake and reaches a little bit down from there).

This means that Sapphire Rapid might be competitive in the ultra high-end but In mainstream servers It's almost a full year of Genoa vs Ice-Lake (after a year of Milan vs Ice-lake)

 

[DrMrLordX]

Why is Sapphire Rapids being relegated to a reduced role?

 

[Gideon]

Why is Sapphire Rapids being relegated to a reduced role?

Maybe it uses exotic packaging? The other guess would be 7nm but that should be near-impossible in 2021.

 

[piokos]

Anyway, if these slides are accurate, Intel plans to run with Ice-Lake for most mainstream servers at least up to 2023 (e.g. those that fit to Cascade Lake specs, e.g. probably <=32 cores) and Sapphire Rapid will only be an ultra-high-end offering (replaces Cooper Lake and reaches a little bit down from there).

This means that Sapphire Rapid might be competitive in the ultra high-end but In mainstream servers It's almost a full year of Genoa vs Ice-Lake (after a year of Milan vs Ice-lake)

Of course they plan to run Ice Lake SP for few years. Why would they not want to do that? It has to make money.
Which absolutely doesn't mean they won't launch a new generation every year.

Ice Lake U is perfectly capable and SP will be as well.
The only real issue: high core count for monolithic Ice Lake SP. They'll probably want to match Zen2's 64-core as well. We'll see how it goes.

 

[Gideon]

Of course they plan to run Ice Lake SP for few years. Why would they not want to do that? It has to make money.
Which absolutely doesn't mean they won't launch a new generation every year.

They specifically mention Cascade Lake Refresh (which is a refresh of a Skylake-SP refresh ,Cascade Lake) I doubt they would forget to mention an Ice-Lake update

Ice Lake U is perfectly capable and SP will be as well.
The only real issue: high core count for monolithic Ice Lake SP. They'll probably want to match Zen2's 64-core as well. We'll see how it goes.

Well the slide mentions it in no uncertain terms. Cooper-Lake till Sapphire Rapid launches. Ice Lake goes up to 38 cores max

 

[Ajay]

Maybe it uses exotic packaging? The other guess would be 7nm but that should be near-impossible in 2021.

Well, I suppose it could be 7nm - that would explain the low volume. On 10nm, one would think the volume would be much higher.
But this slide introduces many uncertainties that, without explanation, just seems to confuse our vision of Intel's roadmap.

 

[uzzi38]

Maybe it uses exotic packaging? The other guess would be 7nm but that should be near-impossible in 2021.

Definitely 10nm. It's got a lot of cores though, and some other interesting things about it, so I'm not surprised it's going to be a kind of premium option.

 

[vstar]

Definitely 10nm. It's got a lot of cores though, and some other interesting things about it, so I'm not surprised it's going to be a kind of premium option.

Interesting... do we have any idea about the core counts of the SPR chips?

 

[Ajay]

Definitely 10nm. It's got a lot of cores though, and some other interesting things about it, so I'm not surprised it's going to be a kind of premium option.

Oh, okay - lots of core on 10nm == even lower yield. Yeesh.

 

[Adonisds]

Ate those rumors that Ocean Cove will have 1.8x the IPC is Skylake and 7nm will clock even higher plausible?

 

[Markfw]

Ate those rumors that Ocean Cove will have 1.8x the IPC is Skylake and 7nm will clock even higher plausible?

When you don't have benchmarks, anything is possible, since no one will know how fast it is or what the IPC is !

Marketing=(1.8x IPC + 5 ghz),

benchmark that no one will ever see = (1.1x IPC + 4 ghz)

 

[jpiniero]

Oh, okay - lots of core on 10nm == even lower yield. Yeesh.

In theory Saphhire Rapids uses chiplets, although probably not to the extent that AMD does. Might be similar to the now cancelled Cooper AP which was 56.

 

[DrMrLordX]

Maybe it uses exotic packaging? The other guess would be 7nm but that should be near-impossible in 2021.

Sapphire Rapids has been 10nm on every roadmap/slide I've ever seen.

Ate those rumors that Ocean Cove will have 1.8x the IPC is Skylake and 7nm will clock even higher plausible?

I thought Ocean Cove had been significantly altered and/or cancelled.

 

[trexfromouterspace]

I thought Ocean Cove had been significantly altered and/or cancelled.

No, you're thinking of Ocean Cove. Ocean Cove was axed, but Ocean Cove is proceeding according to plan.

 

[Gideon]

No, you're thinking of Ocean Cove. Ocean Cove was axed, but Ocean Cove is proceeding according to plan.

Yeah, original Ocean Cove was rumord to have been axed during Krzanich era, whatever is in progress now shouldn't be that.

The same leaker about Core counts and TDP. Sapphire Rapid had better bring massive IPC gains if this is correct:

https://twitter.com/x/status/1268176240097161217

• CPX : 24C , 165-200W , 4S
• ICX : 24C , >200W , 2S
• ICX : 32C , 150-200W 2S
• SPR : 48C , <200W , 1-2S
• SPR : 32C , >200W , 2S/4-8S

*not Max cores.

 

[Ajay]

No, you're thinking of Ocean Cove. Ocean Cove was axed, but Ocean Cove is proceeding according to plan.

????

 

[Ajay]

In theory Saphhire Rapids uses chiplets, although probably not to the extent that AMD does. Might be similar to the now cancelled Cooper AP which was 56.

Well, I hope they do. Otherwise, large core count 10nm CPUs will be unobtanium. They've been talking about EMIB and Foveros for long enough.

 

[Exist50]

I thought Ocean Cove had been significantly altered and/or cancelled.

Ocean Cove is dead, period. Both the name and the architecture behind it.

 

[lobz]

Definitely 10nm. It's got a lot of cores though, and some other interesting things about it, so I'm not surprised it's going to be a kind of premium option.

7nm is Granite with Golden Cove, right?

 

[uzzi38]

Interesting... do we have any idea about the core counts of the SPR chips?

Yes. Enough that it would have been very, very competitive against Milan.

...were it not going to hit the market a year or perhaps more later and around the same timeframe as Genoa.