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.





Intel Core Ultra 100 - Meteor Lake



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.)

 

[SiliconFly]

If Intel debuts their raytracing acceleration algorithm with the MTL tGPU, it could be a defining moment for them. Suddenly, people enjoying raytracing at minimum 30fps on their thin and light laptops! Nvidia is gonna get kneecapped, to speak nothing of AMD.

Even assuming that the tGPU has ray-tracing accelerators built-in (which i seriously doubt), it'll be no match to a dGPU with dedicated GDDR. MTL's tGPU is a nice upgrade for many common tasks & light gaming, but I wouldn't expect too much from it. Honestly, I don't think it'll even come close to a measly RTX 3050 mobile.

 

[mikk]

Even assuming that the tGPU has ray-tracing accelerators built-in (which i seriously doubt), it'll be no match to a dGPU with dedicated GDDR. MTL's tGPU is a nice upgrade for many common tasks & light gaming, but I wouldn't expect too much from it. Honestly, I don't think it'll even come close to a measly RTX 3050 mobile.

Meteor Lake patches suggest otherwise.

Intel Meteor Lake CPUs With TGPU Reportedly Features Ray Tracing Support, Sports FP64 Compute But Lacks XMX Units

In a new commit published on GitHub, Intel might have confirmed ray tracing support for their Meteor Lake CPUs with Xe-HPG TGPU.

wccftech.com

 

[SiliconFly]

Meteor Lake patches suggest otherwise.

Intel Meteor Lake CPUs With TGPU Reportedly Features Ray Tracing Support, Sports FP64 Compute But Lacks XMX Units

In a new commit published on GitHub, Intel might have confirmed ray tracing support for their Meteor Lake CPUs with Xe-HPG TGPU.

wccftech.com

Well. It looks like it has RT cores.

As reported by many (link & link), MTL's tGPU is a derivative of Arc A380M with 8 Xe cores (128 EUs) operating at just around 4 tflops, but lacks dedicated memory & XMX. This puts it as a big disadvantage when compared to a dedicated A380M.

The original A380M is more comparable to an abysmal GTX 1650 in raw performance without ray-tracing. And with ray-tracing enabled, it won't even match a GTX 1050 Ti (in terms of pure FPS only). Now thats a dedicated onboard A380M dGPU I'm talking about!

Now coming back to MTL's tGPU... (without dedicated GDDR & XMX), MTL's tGPU with ray-tracing enabled will bring the game to a grinding halt unless we reduce the native resolution to a poor level & reduce the render resolution to a laughable level & use XeSS to upscale.

In short, without RT, MTL's tGPU is expected to perform 20% less than a GTX 1650. And with RT on, it'll perform way worse than a GTX 1050 Ti (in pure frame rates only).

Like I mentioned before, MTL's tGPU is something we shouldn't get excited about. It's a welcome addition to MTL as it performs better than the previous Intel iGPUs. Thats all.

 

[Glo.]

Well. It looks like it has RT cores.

As reported by many (link & link), MTL's tGPU is a derivative of Arc A380M with 8 Xe cores (128 EUs) operating at just around 4 tflops, but lacks dedicated memory & XMX. This puts it as a big disadvantage when compared to a dedicated A380M.

The original A380M is more comparable to an abysmal GTX 1650 in raw performance without ray-tracing. And with ray-tracing enabled, it won't even match a GTX 1050 Ti (in terms of pure FPS only). Now thats a dedicated onboard A380M dGPU I'm talking about!

Now coming back to MTL's tGPU... (without dedicated GDDR & XMX), MTL's tGPU with ray-tracing enabled will bring the game to a grinding halt unless we reduce the native resolution to a poor level & reduce the render resolution to a laughable level & use XeSS to upscale.

In short, without RT, MTL's tGPU is expected to perform 20% less than a GTX 1650. And with RT on, it'll perform way worse than a GTX 1050 Ti (in pure frame rates only).

Like I mentioned before, MTL's tGPU is something we shouldn't get excited about. It's a welcome addition to MTL as it performs better than the previous Intel iGPUs. Thats all.

There is a lot of just assumptions.

The facts are:

There is no such GPU as A380M, only A370M, and that has lower clock speed than MTL GPU will have: over 2 GHz vs 1.55 GHz, and has less GPU bandwidth than MTL tGPU will have: 112 vs 120 GB/s bandwidth, based on leaked specs.

In essence, on paper, MTL tGPU should be slightly faster than A370M, and A370M scores 3800 pts in 3D Mark TS.

 

[SiliconFly]

There is a lot of just assumptions.

The facts are:

There is no such GPU as A380M, only A370M, and that has lower clock speed than MTL GPU will have: over 2 GHz vs 1.55 GHz, and has less GPU bandwidth than MTL tGPU will have: 112 vs 120 GB/s bandwidth, based on leaked specs.

In essence, on paper, MTL tGPU should be slightly faster than A370M, and A370M scores 3800 pts in 3D Mark TS.

Then it's probably a derivative of A370M with faster clocks and no dedicated GDDR and 8 Xe & RT cores. Still puts it on par with a GTX 1650 in raw performance (+/- 20% max). And this is without RT. If we turn on RT at native resolution, it'll bring it to a crawl.

Honestly, I'm excited that MTL's tGPU is quite a significant upgrade over previous gen. But we shouldn't be comparing it with the likes of RTX mobile dGPUs. For example, even the entry level RTX 4050 will be at least 3 to 4 times faster than MTL's tGPU in raw performance. And with DLSS 3.5, it's not even a comparison anymore.

I like Intel and they're doing good. But we shouldn't get our expectations too high.

 

[Glo.]

Then it's probably a derivative of A370M with faster clocks and no dedicated GDDR and 8 Xe & RT cores. Still puts it on par with a GTX 1650 in raw performance (+/- 20% max). And this is without RT. If we turn on RT at native resolution, it'll bring it to a crawl.

Honestly, I'm excited that MTL's tGPU is quite a significant upgrade over previous gen. But we shouldn't be comparing it with the likes of RTX mobile dGPUs. For example, even the entry level RTX 4050 will be at least 3 to 4 times faster than MTL's dGPU in raw performance. And with DLSS 3.5, it's not even a comparison anymore.

I like Intel and they're doing good. But we shouldn't get our expectations too high.

4050 will be AT BEST 2-2.5 times faster than MTL GPU. Not 3-4 times faster. In raster perf, obviously.

370M is the same GPU core as is in A380 but with 32 bit bus(64 bit total, hence the 112 GB/s bandwdith). MTL-P tGPU is completely different design with completely different memory controller.

 

[mikk]

Well. It looks like it has RT cores.

As reported by many (link & link), MTL's tGPU is a derivative of Arc A380M with 8 Xe cores (128 EUs) operating at just around 4 tflops, but lacks dedicated memory & XMX. This puts it as a big disadvantage when compared to a dedicated A380M.

I'm not so sure about that. Xe HPG and Xe LPG differ - to what extent is unknown. Yes Xe LPG is a derivate from Xe HPG but it doesn't necessarily mean they are identical.

Also keep in mind Arc A380 dGPU runs via the PCI Express bus which could introduce latency penalties over a tGPU version. DG1 was often slower than the iGPU version for this reason even with a bandwidth advantage.: https://forums.anandtech.com/thread...x11-performance.2526376/page-30#post-40335377

The biggest dGPU advantage is the power budget. Not all MTL versions will run 45+W constrained, and this is CPU+GPU combined. A380 requires 45-55W for 2450 Mhz depending on the game (GPU only).

 

[A///]

Recently, most of the "high-end" laptops are actually powered by AMD cpus. It's a tectonic shift considering Intel dominated this space for more than a decade.

A $5,000 gaming laptop by Asus is not what I consider high end. It is high end for a different market.

 

[SiliconFly]

4050 will be AT BEST 2-2.5 times faster than MTL GPU. Not 3-4 times faster. In raster perf, obviously.

370M is the same GPU core as is in A380 but with 32 bit bus(64 bit total, hence the 112 GB/s bandwdith). MTL-P tGPU is completely different design with completely different memory controller.

Even if MTL tGPU has a completely different memory controller, it still has to "share" the low-bandwidth DDR with MTL's tCPU. This is a big blow to it's performance (due to both low-bandwidth issue & "sharing" issue).

On top of that, like @mikk mentioned, it has to share the constrained power budget with MTL's tCPU, which again is another serious blow to it's performance.

4050 is around 2 to 2.5 times faster than discreet A370M graphics in raw performance.

4050 will be around 3 to 4 times faster than MTL's tGPU in raw performance. Since, discreet A370M graphics is naturally faster than MTL's tGPU (since MTL tGPU "shares" memory with tCPU, uses slower DDR & has low power-budget than A370M, in spite of higher clocks).

And we're not even talking about ray-tracing performance. Life has never been easy for an iGPU compared to a dGPU. We'll have to wait and see. Only 10 more days to go.

 

[eek2121]

While I don’t think it will quite match discrete variants, I do suspect the MTL GPU will be “within the ballpark”.

Where it lands in terms of the competition? Who knows.

 

[SiliconFly]

Where it lands in terms of the competition? Who knows.

True.

Also, MTL is all about power-efficiency, a brand new tGPU & VPU and chiplet design in particular. Overall, it's gonna be a great package & the competition won't have an answer to MTL for at least 5 to 6 months i think.

But I believe we shouldn't focus too much on specifics like individual performance of each tile comparing them to previous gen or discrete parts. As far ad MTL is concerned, it's the entire SoC that counts I think. Might be worth it if they price it right & there's decent availability.

 

[AMDK11]

Ok what? Zen 3 vs Zen 2 was a massive increase in IPC, RWC should be nowhere near as close.

The average Zen3 IPC gain is 19%.I'm not saying RedwoodCove will get a similar increase, but the Zen4 level, which is an average IPC increase of 13%, should be achievable.I would bet on a range of +10-15% average IPC increase. Zen4 is also a new core , contrary to what they claim.

 

[AMDK11]

Wish those RWC improvements bring as much performance to the table as I think it would. But like I said, Intel being Intel, I would rather see it before believing it.

GoldenCove, although much wider decoding from 4 in SunnyCove/CypressCove to 6, is still not an ideal core.There is a lot to improve and I think RedwoodCove is about optimization, reconstruction and expansion to remove bottlenecks and increase IPC.I am not saying that the increase will be at the level of 20%, but around 10-15% should be achievable.Let me remind you that Zen4 means an average IPC increase of 13%.

RedwoodCove is a new x86 core, not just GoldenCove+.

1. SandyBridge new x86
2. IvyBridge(SandyBridge+)

1. Haswell new x86
2. Broadwell(Haswell+)

1. SunnyCove new x86
2. WillowCove(SunnyCove+)

1. GoldenCove new x86
2. RaptorCove(GoldenCove+)

Normally RaptorCove should be a transition to Intel4 with larger L2 but it still remains in Intel 7. The new x86 core is RedwoodCove and I suspect, although I may be wrong, that ArrowLake may include RedwoodCove+ in Intel 20A with 3MB L2.

 

[eek2121]

True.

Also, MTL is all about power-efficiency, a brand new tGPU & VPU and chiplet design in particular. Overall, it's gonna be a great package & the competition won't have an answer to MTL for at least 5 to 6 months i think.

But I believe we shouldn't focus too much on specifics like individual performance of each tile comparing them to previous gen or discrete parts. As far ad MTL is concerned, it's the entire SoC that counts I think. Might be worth it if they price it right & there's decent availability.

Yeah, I am honestly waiting for reviews to drop. My current laptop has Cezanne in it. Battery life is…okay.

If lightweight 6+8 based MTL-P laptops beat current AMD offerings in terms of perf/watt, I might have to pick one up. Ideally wanting something capable of light gaming that is less than 3 lbs and chargeable over USB-C.

 

[SiliconFly]

although I may be wrong, that ArrowLake may include RedwoodCove+ in Intel 20A with 3MB L2.

Well, the jury is still out in this matter. It may have RWC+ (likely) or the 1st iteration of LNC (more likely). But most "believable" rumors suggest more in favor of LNC rather than RWC+

 

[FangBLade]

True.

Also, MTL is all about power-efficiency, a brand new tGPU & VPU and chiplet design in particular. Overall, it's gonna be a great package & the competition won't have an answer to MTL for at least 5 to 6 months i think.

But I believe we shouldn't focus too much on specifics like individual performance of each tile comparing them to previous gen or discrete parts. As far ad MTL is concerned, it's the entire SoC that counts I think. Might be worth it if they price it right & there's decent availability.

I wouldn't be so sure; MTL is nothing special, and it will only have 6P cores, which is a bottleneck for gaming. AMD already has chiplet design in laptops, albeit for the high-end segment, but they can easily adapt if the competition pressures them. Plus, there's the 3D cache on laptops, etc. They can respond in various ways, and gaming performance will be the best advertisement, with the 3D cache being the king both in terms of performance and efficiency. Don't forget the much more stable drivers; Arc is developing intensively, and no matter how many resources they allocate to the GPU division, they can't compensate for the decades of optimization that AMD and Nvidia have behind them.

 

[AMDK11]

Well, the jury is still out in this matter. It may have RWC+ (likely) or the 1st iteration of LNC (more likely). But most "believable" rumors suggest more in favor of LNC rather than RWC+

I have changed my approach and try to be careful with forecasts. It would be nice if ArrowLake had LionCove with 8-way decoding and ROB 700+, but recent leaks from Intel themselves make me think ArrowLake is on RedwoodCove+ cores and LionCove is only on LunarLake. I'd like to be wrong.

RaptorCove was supposed to be a transition to Intel 4, i.e. GoldenCove+ with 2MB L2, but at some point they decided to stay with Intel 7. RedwoodCove as a new core with 3MB L2 on Intel 20A. This would mean that we will only see LionCove in 2025 on Intel 18A.

EDIT:
I remember Keller saying that the changes from generation to generation will be greater. It's quite possible he meant that we have 4-way decoding from Conroe (Core 2) to Sunny/CypressCove, which is a long time. GoldenCove introduced 6-way decoding, so from GoldenCove to RedwoodCove, because LionCove, according to leaks, already has 8-way decoding. So it turns out that 6-way decoding will only be available for a few years compared to 4-way decoding. Perhaps this is the context of Jim's statement.

 

[Geddagod]

GoldenCove, although much wider decoding from 4 in SunnyCove/CypressCove to 6, is still not an ideal core.There is a lot to improve and I think RedwoodCove is about optimization, reconstruction and expansion to remove bottlenecks and increase IPC.I am not saying that the increase will be at the level of 20%, but around 10-15% should be achievable.Let me remind you that Zen4 means an average IPC increase of 13%.

RedwoodCove is a new x86 core, not just GoldenCove+.

1. SandyBridge new x86
2. IvyBridge(SandyBridge+)

1. Haswell new x86
2. Broadwell(Haswell+)

1. SunnyCove new x86
2. WillowCove(SunnyCove+)

1. GoldenCove new x86
2. RaptorCove(GoldenCove+)

Normally RaptorCove should be a transition to Intel4 with larger L2 but it still remains in Intel 7. The new x86 core is RedwoodCove and I suspect, although I may be wrong, that ArrowLake may include RedwoodCove+ in Intel 20A with 3MB L2.

That order and grouping is wrong and the general idea there is wrong as well. New "big" cores came the architecture after the node shrink.
Raptor Cove was not a "separate" architecture like redwood cove will be, or like what palm cove, ivy bridge was. Raptor Cove was not a transition to Intel 4, that was always redwood cove. Intel themselves claimed Raptor Lake was never meant to exist, and they only released it when it was apparent MTL would be delayed. RPC was always designed for Intel 7.
Also you could remind me that Zen 4 was 13% IPC all you want, but AMD is a different company. Intel has always gotten ~5-10% IPC jumps from their node shrink architectures, unlike AMD who targets higher. There is literally no reason to expect RWC to deviate from this trend.
The planned order, barring Intel 4 delays, would have been:
PLMC (old arch ported)
SNC (new arch)
WLC (old arch)
GLC (new arch)
RWC (old arch ported)
LNC (new arch)

 

[Glo.]

I wouldn't be so sure; MTL is nothing special, and it will only have 6P cores, which is a bottleneck for gaming. AMD already has chiplet design in laptops, albeit for the high-end segment, but they can easily adapt if the competition pressures them. Plus, there's the 3D cache on laptops, etc. They can respond in various ways, and gaming performance will be the best advertisement, with the 3D cache being the king both in terms of performance and efficiency. Don't forget the much more stable drivers; Arc is developing intensively, and no matter how many resources they allocate to the GPU division, they can't compensate for the decades of optimization that AMD and Nvidia have behind them.

Considering that vast majority(like 95% of them) of computers that will use advanced APUs/SOCs from both Intel and AMD will not have dGPU I would not worry about gaming performance, which will be limited by the GPUs.

Until Arrow Lake-P and Strix Point Halo come out we have nothing to worry about CPU perf bottlenecking the GPUs.

 

[Geddagod]

I have changed my approach and try to be careful with forecasts. It would be nice if ArrowLake had LionCove with 8-way decoding and ROB 700+, but recent leaks from Intel themselves make me think ArrowLake is on RedwoodCove+ cores and LionCove is only on LunarLake. I'd like to be wrong.

RaptorCove was supposed to be a transition to Intel 4, i.e. GoldenCove+ with 2MB L2, but at some point they decided to stay with Intel 7. RedwoodCove as a new core with 3MB L2 on Intel 20A. This would mean that we will only see LionCove in 2025 on Intel 18A.

EDIT:
I remember Keller saying that the changes from generation to generation will be greater. It's quite possible he meant that we have 4-way decoding from Conroe (Core 2) to Sunny/CypressCove, which is a long time. GoldenCove introduced 6-way decoding, so from GoldenCove to RedwoodCove, because LionCove, according to leaks, already has 8-way decoding. So it turns out that 6-way decoding will only be available for a few years compared to 4-way decoding. Perhaps this is the context of Jim's statement.

Being more careful with your forecasts would mean following the general sentiment of the leakers, who all claim ARL has LNC. Not a single one think ARL uses RWC+. It's fine to have less popular and more speculative forecasts, but one should advertise it as such.
Your assumption that LNL is coming 2025 on Intel 18A is wrong. 18A vs N3 is all rumors (it's n3) but Intel themselves confirmed LNL will launch in 2024 numerous times this year.
I'm pretty sure the "ARL has RWC+" rumors started since ARL performance projections got leaked. Guys, LNC might just be a dud of an architecture. It happens. SNC wasn't all that great either, remember? I mean I hope those perf projections aren't true for the sake of competition in the market, but no leaker seems to be disputing it either...

 

[SiliconFly]

I wouldn't be so sure; MTL is nothing special, and it will only have 6P cores, which is a bottleneck for gaming. AMD already has chiplet design in laptops, albeit for the high-end segment, but they can easily adapt if the competition pressures them. Plus, there's the 3D cache on laptops, etc. They can respond in various ways, and gaming performance will be the best advertisement, with the 3D cache being the king both in terms of performance and efficiency. Don't forget the much more stable drivers; Arc is developing intensively, and no matter how many resources they allocate to the GPU division, they can't compensate for the decades of optimization that AMD and Nvidia have behind them.

Really? Tell me one existing Ryzen mobile chip that has a GPU that's even half as good as MTL's.

 

[SiliconFly]

make me think ArrowLake is on RedwoodCove+ cores and LionCove is only on LunarLake.

I too believe the same. Thats just how Intel usually operates. But most rumors point in the other direction. Need to wait for more clarity.

 

[Geddagod]

Thats just how Intel usually operates.

How so?

 

[Glo.]

Really? Tell me one existing Ryzen mobile chip that has a GPU that's even half as good as MTL's.

Both 680M and 780M Radeon iGPUs are much faster than 50% of MTLs supposed performance.

780M, for example scores 2800 pts in 3D Mark Time Spy Graphics. MTL-P 128 EU iGPU is supposed to score here around 3500 pts. 680 in TS Graphics got 2400 pts, average.

You are vastly underestimating the performance of iGPUs these days.

 

[Abwx]

Really? Tell me one existing Ryzen mobile chip that has a GPU that's even half as good as MTL's.

Half as good as something that doesnt exist at all...

Besides not only Intel s huge improvement numbers use as a reference a 32 EU iGP but also Intel s iGPUs have 3DMark numbers that are quite optimistic in comparison of real perfs in games, check the numbers here and there.