Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads

[Tigerick]

As Hot Chips 34 starting this week, Intel will unveil technical information of upcoming Meteor Lake (MTL) and Arrow Lake (ARL), new generation platform after Raptor Lake. Both MTL and ARL represent new direction which Intel will move to multiple chiplets and combine as one SoC platform.

MTL also represents new compute tile that based on Intel 4 process which is based on EUV lithography, a first from Intel. Intel expects to ship MTL mobile SoC in 2023.

ARL will come after MTL so Intel should be shipping it in 2024, that is what Intel roadmap is telling us. ARL compute tile will be manufactured by Intel 20A process, a first from Intel to use GAA transistors called RibbonFET.





As mentioned by Tomshardware, TSMC will manufacture the I/O, SoC, and GPU tiles. That means Intel will manufacture only the CPU and Foveros tiles. (Notably, Intel calls the I/O tile an 'I/O Expander,' hence the IOE moniker.)

 

[Kepler_L2]

The people that believe it will be higher are taking it from a Turin Specint benchmark showing a 30% performance gain (as you know). I don’t think that means that Zen 5 client has a 30% IPC improvement. Milan showed a 30% improvement in Specint too and Zen 3 client launched with a 19% IPC gain.

I’ve got no data on Zen 5 client so I can’t say anything definitively but I don’t think it’s getting a 30% ST bump in a single generation. Even if it did, it’ll have 30-40% less memory bandwidth than ARL-S that it’ll be held back in certain workloads anyway.

If you're referring to Jim Keller's presentation that's not where Zen5 numbers come from.

 

[adroc_thurston]

The people that believe it will be higher are taking it from a Turin Specint benchmark showing a 30% performance gain (as you know). I don’t think that means that Zen 5 client has a 30% IPC improvement. Milan showed a 30% improvement in Specint too and Zen 3 client launched with a 19% IPC gain.

I'm sorry to spoil your party but Zen5 is very much a fat IPC bump with a few bins of speed loss.
It is what it is; the thing's sampled far and wide even in the Strix form by now.

but I don’t think it’s getting a 30% ST bump in a single generation

Literally every single numbered AMD bump was at least 30% 1t. lol.

 

[H433x0n]

I'm sorry to spoil your party but Zen5 is very much a fat IPC bump with a few bins of speed loss.
It is what it is; the thing's sampled far and wide even in the Strix form by now.

I consider 20% IPC gain a fat bump. They’d need a ~25% IPC increase to break even with a 300mhz clock regression, a ~22-23% IPC increase to break even with a 200mhz clock regression.

Literally every single numbered AMD bump was at least 30% 1t. lol.

Not sure what you meant by this but I’m assuming you meant previous generations. They’ve only gotten a 30% 1T improvement in a single generation once with Zen 4 and it came with a 700mhz higher clocks, a new node and DDR5. Zen 5 isn’t working with the same conditions that gave Zen 4 that big boost. It’s got to all be done with raw IPC.

 

[H433x0n]

If you're referring to Jim Keller's presentation that's not where Zen5 numbers come from.

I’m not.

 

[adroc_thurston]

I consider 20% IPC gain a fat bump

That's normal.
Fat is twice that.

They’d need a ~25% IPC increase to break even with a 300mhz clock regression, a ~22-23% IPC increase to break even with a 200mhz clock regression.

They're not aiming to "break even".

They’ve only gotten a 30% 1T improvement in a single generation once with Zen 4

No.
Zen2 was 30% 1t, and Zen3 was 30% 1t or therein.
It's a combination of IPC and clock bumps, but in case of Zen5 there's no clock bump!
just IPC.

 

[H433x0n]

That's normal.
Fat is twice that.

They're not aiming to "break even".

No.
Zen2 was 30% 1t, and Zen3 was 30% 1t or therein.
It's a combination of IPC and clock bumps, but in case of Zen5 there's no clock bump!
just IPC.

So just to be clear- you’re saying that Zen 5 has a 40% IPC gain and is only losing a few hundred mhz of clocks?

I wish this forum had a !RemindMe bot for situations like these.

 

[yuri69]

ARL will be surely fine even vs that magical +40% IPC Zen 5. Intel firmly holds the volume of major markets and AMD slowly goes away from mobile/OEM desktop.

Zen2 was 30% 1t, and Zen3 was 30% 1t or therein.
It's a combination of IPC and clock bumps, but in case of Zen5 there's no clock bump!
just IPC.

3900X vs 2700X often barely reached 20% T1 gain unless AVX...

 

[adroc_thurston]

you’re saying that Zen 5 has a 40% IPC gain and is only losing a few hundred mhz of clocks?

I've only ever said 32%.

3900X vs 2700X often barely reached 20% T1 gain unless AVX...

not in specINT.

ARL will be surely fine even vs that magical +40% IPC Zen 5.

That's cope.

Intel firmly holds the volume of major markets and AMD slowly goes away from mobile/OEM desktop.

By being very-very cheap with RPLs, yes.
282 in particular.

 

[H433x0n]

I've only ever said 32%.

That doesn’t achieve what you’re claiming. A 32% IPC increase (which I doubt btw) coupled with a 300mhz clock regression gets Zen 5 a ST performance lead similar to what 13900K has right now. Thats best case scenario assuming it hits every bit of that 32% gain.

So 10% better ST (best case), while having worse efficiency at low power scenarios, less bandwidth and much less MT performance lower down the stack. That’s what all this hype is about?

Edit: To be clear, I don’t think it will have a 32% IPC improvement. I think Turin will hit those types of numbers in Specint 1T but Turin is getting a bunch of advantages that client won’t have (more memory bandwidth, more power budget, etc).

 

[adroc_thurston]

to what 13900K has right now

what does it have in specint? (please no cinememe)

while having worse efficiency at low power scenarios

what

less bandwidth

How is this a metric?
It's not power or perf or area.

 

[aigomorla]

I'm sorry to spoil your party but Zen5 is very much a fat IPC bump with a few bins of speed loss.
It is what it is; the thing's sampled far and wide even in the Strix form by now.

Literally every single numbered AMD bump was at least 30% 1t. lol.

That's normal.
Fat is twice that.

They're not aiming to "break even".

No.
Zen2 was 30% 1t, and Zen3 was 30% 1t or therein.
It's a combination of IPC and clock bumps, but in case of Zen5 there's no clock bump!
just IPC.

what does it have in specint? (please no cinememe)

what

How is this a metric?
It's not power or perf or area.

And what point do these post have in this thread except to incite or flame bait?
Is Zen5 even out? Has anyone even posses a the CPU to provide valid numbers for you justifying posting about Zen5 in an INTEL thread?

Haven't i told all of you guys, we are tired of hearing about Zen 5 in an Intel thread?

Please keep talking about AMD with no point or reason in a Intel thread.
Pretty soon you won't be allowed to post in ANY intel posts if you keep ignoring us moderators.

We are done giving out freebies.

I will say this again.
You are OK to bring AMD into this discussion providing you have a valid reason like a basis of comparison but you better have a reason and the words INTEL must also be included in that post to provide validations.
But bringing Zen5 which is not even out yet, and then making claims which can not be proven yet, (even if its ends up being true), is TROLLING / INCITING, people that just want to talk about Intel.

No more Freebie's will be given to anyone who violates this.
If you want to talk about Zen5 goto the Zen5 thread, don't pollute the intel thread.


Moderator Aigo

 

[yuri69]

Intel promises a yearly cadence with the Core line. So Zen 5 has to somehow compete vs Intel lineup in 2024, 2025, and partially in 2026.

Like I said it looks pretty solid for Intel.

 

[H433x0n]

what

There’s a reason why Phoenix is monolithic and STX Halo is so different from Zen 5 client desktop. You know why.

How is this a metric?
It's not power or perf or area.

Bandwidth isn’t performance, good to know. Feeding your big powerful cores through a straw will have no effect.

 

[Joe NYC]

By being very-very cheap with RPLs, yes.
282 in particular.

I think it is quite telling that AMD can't outsell in any category. Regardless of price, and quality of the product.

 

[aigomorla]

This is why i asked all of you guys to stop talking about Zen5, as it always starts a wall of of ZEN, which does not belong in a Intel Thread.

I won't repeat this again.

Zen5 needs to go ->

Discussion - Zen 5 Discussion (EPYC Turin and Strix Point/Granite Ridge - Ryzen 8000)

Well, since many folks already got their hands (or at least going to get) on Zen 4 CPUs , time to discuss about Zen 5 (Zen 4 already old news :D) We already got roadmaps and key technologies like AIE Some things we already knew Dr. Lisa Su and Forrest Norrod already mentioned at FAD...

forums.anandtech.com

This thread is for discussion regarding Meteor / Arrow / Lunar.... not Turin , Strix, and Granite.

Moderator Aigo

 

[DrMrLordX]

That it's an announcement that isn't delaying something for another year. Intel just being able to deliver something on time is hopefully a return to form for the company.

Seems like we're setting a low bar here. And they didn't even demo the product!

If the efficiency claims hold up and it delivers better battery life than Raptor lake mobile

We should have at least had a controlled demo for them to showcase that efficiency.

 

[S'renne]

So I suppose the NPU part of Meteor Lake isn't that significant when the AI technologies was the focus of this year's innovative event? (just trying to spark another relevant topic)

 

[SiliconFly]

So I suppose the NPU part of Meteor Lake isn't that significant when the AI technologies was the focus of this year's innovative event? (just trying to spark another relevant topic)

Ya. It's a bit intriguing considering the NPU takes only a smaller amount of die space.

And Intel seems to be over emphasizing it. They've even started calling MTL Laptops as AI PCs! Sounds kinda silly.

 

[S'renne]

Ya. It's a bit intriguing considering the NPU takes only a smaller amount of die space.

And Intel seems to be over emphasizing it. They've even started calling MTL Laptops as AI PCs! Sounds kinda silly.

I mean they are running AI tasks for all sorts of client side purposes through local hardware as opposed of cloud AI with open sourced projects..and the main topic of the event is AI focused. I do like how the chiplets works onwards from Meteor Lake.

Also, did I misremember or misunderstand something, but how does the whole SoC has 128 GB/s bandwidth across all tiles?

 

[KompuKare]

That was all a huge media campaign from Intel yesterday.
Time will tell whether it was all more than words, certainly in terms of process execution Intel have a lot to prove.
Perf/watt is the killer IMO. The E cores are fine but the P cores can be crazy. Guess Meteor Lake is mobile so no crazy 13900K super-clock modes.
Last times they went mobile only was all a bit... suspect... in terms of nodes etc.
That they've gone max 6 P cores is strange and potentially hard to sell.
Their packaging is pretty ambitious - hope they have all the capacity.
Guess this that leaves Nvidia as the last of the big 3 to not get innovative with packaging.

 

[SiliconFly]

Intel promises a yearly cadence with the Core line. So Zen 5 has to somehow compete vs Intel lineup in 2024, 2025, and partially in 2026.

Like I said it looks pretty solid for Intel.

3 generations of CPUs in less than 1.5 years, thats pretty solid execution!

 

[S'renne]

3 generations of CPUs in less than 1.5 years, thats pretty solid execution!

ignoring the further delayed Meteor Lake yeah

 

[Timorous]

3 generations of CPUs in less than 1.5 years, thats pretty solid execution!

Given their track record I for one will be waiting for them to actually deliver before talking about solid execution.

 

[SiliconFly]

Also, did I misremember or misunderstand something, but how does the whole SoC has 128 GB/s bandwidth across all tiles?

MTL's NOC Scalable Fabric connects all the tiles at 128 GB/s which is perfectly fine for MTL. For example, a 4-tile AMD EPYC SoC has it's 4 compute tiles interlinked by Infinity Fabric at only 42.6 GB/s per link, bringing the 4 link aggregate to a total of 170.6 GB/s. This bandwidth is more than sufficient for EPYC.

In MTL's case, it has only one compute tile. So, I think, Intel might have decided 128 GB/s is more than enough. Intel NOC Scalable Fabric is more advanced than AMD's Infinity Fabric. The NOC Scalable Fabric can directly link multiple dies and modules seamlessly with each other using fast links and acts more like a network switch rather than independent die-to-die interconnects.

(More info: AT, wikichip & link)

 

[inf64]

3 generations of CPUs in less than 1.5 years, thats pretty solid execution!

That is the "projected" roadmap. Knowing intel's past, it can be 3 years until we see that last part from the slide hit the store shelf. Just saying.