Discussion Intel current and future Lakes & Rapids thread

[Geddagod]

It is what it is, AmpereOne is also getting roadkilled in scale-out favela wars.

It and EMR actually get really fat unified LLCs (albeit slow); that's interesting.
Well, more interesting than 144 meh atoms on a stick.
It's not leadership core/thread density and power isn't anything too impressive either soooo...

The product originally targeted N3 so obviously they will.
Tall cell spam ends circa i3, really.

EMR looks to be interesting because of that, but GNR only seems to be interesting because of the new node. Other large improvements don't seem to be present (other than perf metrics, but I'm referring to structure and parts of the CPU). Based on Pat's comments about the new core in GNR, I thought the core (LNC) might be very interesting, but all I'm hearing is RWC+ now a days. Not that GNR as a whole isn't interesting, just that I think SRF is just as interesting.

While SRF won't take any 'overall' crowns- it might be able to take it in one scenario- lowest power draw per thread at a certain (almost certainly low) performance threshold.
Read elsewhere what cloud customers are mainly after is efficiency at a certain level of performance they can license to their customers on a core-by-core basis. While GNR might be more efficient under heavy load at higher clocks, the narrower SRG cores should be able to take the lead at lower power levels, like GRC does over GLC.
Core density is great as well, but certain cloud providers are having second thoughts on high core density Bergamo because of rising costs related to higher core counts - AWS

First time I'm hearing SRF targeted N3. Thought LNC and SKM were the cores originally targeting N3, with RWC/Crestmont/SRG were originally targeting Intel 7nm/Intel 4.
I always thought that LNC was targeting Intel 3 + TSMC 3nm initially, as a 'tock' to RWC in MTL, so I think it's very believable tall cell spam ended there- with LNC+SKM.

 

[H433x0n]

Well, more interesting than 144 meh atoms on a stick.

By this logic AMD is the most boring company in the industry. If I were to be extremely uncharitable and bad faith - It’s just playing legos and re-using the same ccx across everything and outsourcing the rest to TSMC.

 

[adroc_thurston]

just that I think SRF is just as interesting.

It's meh.

but certain cloud providers are having second thoughts on high core density Bergamo because of rising costs related to higher core counts - AWS

Amazon just says stuff to still justify running an internal server program.

By this logic AMD is the most boring company in the industry

Yea except that part where they make leadership products, duh.

 

[DrMrLordX]

Not every fix has regressions

That's ducky then. And it was a software fix so no need to issue hardware recalls.

 

[uzzi38]

SRF looks like it's going to be pretty small all things considered
144 cores, 144MB L2, and 108MB L3

Using Intel 4, that's gonna give us:
(144/4) x 6.2mm^2 (cores+L2) =~220
(144/4) x 1.15mm^2 (L3) = ~40

I'm guessing each 'block' of the SRF layout is going to be ~7.5mm^2, accounting for the mesh, tweaks in Crestmont vs SRG, but also potential slight density uplift in Intel 3. Using 36 core cluster active, 4 clusters for MC, and 2 for dead core clusters for yields, we could be getting a 6x7 square layout for ~315 mm^2. Accounting for EMIB too, I don't think SRF compute tile is going to be too much larger than 350 mm^2 right?

Wouldn't mind if anyone checked my calculations either 6.2mm^2 for Crestmont was shown in the Zen 4C vs E-cores thread though, I'll link my exact comment later.

Yeah that's the point of it. Not the most performant product, but should be very cheap for the level of performance it does bring.

 

[uzzi38]

WCCFtech lmao
(right column is threads, left is cores)
View attachment 82766

Slightly related, I think this architecture release cadence would make sense:
ARL - Lion Cove
PTL - Cougar Cove (minor improvement)
BSTL? - Panther Cove, or what ever the original arch in panther lake was.
???- Hopefully the next 'big' architecture shift (royal core), or a slight improvement over panther cove (an effective tick)

I think the rumors that PTL switched what arch it was using would be a negative, as in instead of using a 'new' core arch compared to LNC, it would be shifting over to a 'tick' generation core. Knowing Intel, it's best to assume a downgrade lol.
Don't think we will be seeing 'royal core' until 2027-2028. And even that might be a optimistic.

When is this chart from? Want to know how hard I should be laughing.

 

[poke01]

Yea except that part where they make leadership products, duh

Everything they have leadership but GPUs

 

[moinmoin]

By this logic AMD is the most boring company in the industry.

You know that's on purpose? AMD is adamant about convincing everybody of its spotless execution (which actually isn't spotless, but any comparison atm gives them plenty leeway). Excellent and as such predictable execution is boring. They manage to create competitive building blocks and then scale those up and down their whole product range. Intel manages to fail at both currently.

 

[itsmydamnation]

You know that's on purpose? AMD is adamant about convincing everybody of its spotless execution (which actually isn't spotless, but any comparison atm gives them plenty leeway). Excellent and as such predictable execution is boring. They manage to create competitive building blocks and then scale those up and down their whole product range. Intel manages to fail at both currently.

yep and right now the "glue" is still heavily based on K10 era.........

 

[podspi]

How so?

 

[soresu]

yep and right now the "glue" is still heavily based on K10 era.........

"Based on" covers a whole range of possibilities.

Based on could mean like PCIe v5 is based on PCIe v1 which has significant legacy compatibility - same with USB.

Or it could mean something based on earlier work, but entirely deprecating support for earlier versions of the standard.

 

[A///]

Intel Granite Rapids-D CPUs Gets Spotted For the First Time, Will Replace Ice Lake-D

Intel's Granite Rapids-D Xeon CPUs officially surface for the first time, according to a support document that reveals the next-gen lineup.

wccftech.com

Intel Finally Addresses Sapphire Rapids MCC CPU Bug, Firmware Update On The Way

A defect with Sapphire Rapids Xeon CPUs caused a pause in shipments, and Intel has now released a statement clarifying the situation.

wccftech.com

 

[A///]

I was not expecting a patch so soon.

what the heck is a -d xeon?

 

[A///]

Amazon just says stuff to still justify running an internal server program.

that but also companies like oracle and vmware or until recently for the 2nd one were getting dicky with pricing per core. did amd not announce a partnership with vmware this year or last to address this issue? or was that the pensando stuff?

 

[moinmoin]

How so?

"Based on" covers a whole range of possibilities.

I interpreted his text to refer to the fact that till now AMD is still using organic substrate as interconnection between dies/chiplets. So really old tech, used in clever ways to give a still decisive competitive advantage. V-cache and MI300 are the first serious steps away from it.

Intel on the other hand never really tried anything based on organic substrate alone, and their other approaches to packaging so far showed that flexibility through reusability targeting high scalability hasn't really been a serious consideration yet either.

 

[adroc_thurston]

Everything they have leadership but GPUs

MI300 is very leadership and is a GPU.

what the heck is a -d xeon?

Low power BGA Xenos for edge/embedded/etc.

 

[Geddagod]

When is this chart from?

Don't know where WCCFTech got it from, but here's the article link

You know that's on purpose?

He literally said, in the next line "If I were to be extremely uncharitable and bad faith"

Amazon just says stuff to still justify running an internal server program

They were referring to HCC models as a whole, which their own CPUs are also.

 

[H433x0n]

You know that's on purpose? AMD is adamant about convincing everybody of its spotless execution (which actually isn't spotless, but any comparison atm gives them plenty leeway). Excellent and as such predictable execution is boring. They manage to create competitive building blocks and then scale those up and down their whole product range. Intel manages to fail at both currently.

I was purposely being bad faith / dismissive. I don’t actually think that way. I see the merits in their current strategy.

 

[adroc_thurston]

They were referring to HCC models as a whole, which their own CPUs are also.

Ehhh, there are merits to what they say but everyone wants 192c on a stick still.

 

[moinmoin]

I was purposely being bad faith / dismissive. I don’t actually think that way. I see the merits in their current strategy.

It's not wrong though. While AMD tries a lot of new tech and new approaches, it's still very conservative and playing safe a lot at the final step of execution.

Intel on the other hand keeps making very risky moves where I keep wondering why they keep doing it that way where other approaches seem much more likely to succeed.

 

[A///]

He literally said, in the next line "If I were to be extremely uncharitable and bad faith"

he'd change his name to jensen?

 

[adroc_thurston]

Intel on the other hand keeps making very risky moves

Well they have to; they're behind.
Just in some cases (like PVC) it was too much.

While AMD tries a lot of new tech and new approaches, it's still very conservative and playing safe a lot at the final step of execution.

Yea but the underlying IP (like the cores themselves) tends to be rather ambitious.

 

[H433x0n]

Intel on the other hand keeps making very risky moves where I keep wondering why they keep doing it that way where other approaches seem much more likely to succeed.

Hubris within the company culture is endemic. Sometimes it works out well (as seen with PowerVia), other times you end up with Sapphire Rapids that will ultimately cost them billions.

 

[poke01]

MI300 is very leadership and is a GPU.

In the consumer space, they don't period. Their software stack is also not great.

With RT and path tracing as well. Heck even Intel is ahead in this field. You can make the greatest software in the world but if the software sucks what's the point??

AMD's RX 7000s series can't use VR properly due to issues in software.

 

[adroc_thurston]

In the consumer space, they don't period

They just don't really care (well, don't care to make big enough dies anyway; might change).

Their software stack is also not great.

Client s/w works really well, ROCm is okay but not portable (just yet).

With RT and path tracing as well

They're mostly waiting for DXR to evolve into something not 1995.

Heck even Intel is ahead in this field

Intel has >15 years of expertise in ray tracing; literally the only thing LRB was good at (comparatively).

AMD's RX 7000s series can't use VR properly due to issues in software.

Pretty sure they've just fixed that, like days ago.