Intel's Meteor Lake CPU Breaks Ground with On-Package LPDDR5X Memory Integration

[AleksandarK]

During a recent demonstration, Intel showcased its cutting-edge packaging technologies, EMIB (embedded multi-die interconnect bridge) and Foveros, unveiling the highly-anticipated Meteor Lake processor with integrated LPDDR5X memory. This move appears to align with Apple's successful integration of LPDDR memory into its M1 and M2 chip packages. At the heart of Intel's presentation was the quad-tile Meteor Lake CPU, leveraging Foveros packaging for its chiplets and boasting 16 GB of Samsung's LPDDR5X-7500 memory. Although the specific CPU configuration remains undisclosed, the 16 GB of integrated memory delivers a remarkable peak bandwidth of 120 GB/s, outperforming traditional memory subsystems using DDR5-5200 or LPDDR5-6400.

Nevertheless, this approach comes with trade-offs, such as the potential for system-wide failure if a memory chip malfunctions, limited upgradeability in soldered-down configurations, and the need for more advanced cooling solutions to manage CPU and memory heat. While Apple pioneered on-package LPDDR memory integration in client CPUs, Intel has a history of using package-on-package DRAM with its Atom-branded CPUs for tablets and ultrathin laptops. While this approach simplifies manufacturing, enabling slimmer notebook designs, it curtails configuration flexibility. We are yet to see if big laptop makers such as Dell, HP, and Asus, take on this design in the coming months.



View at TechPowerUp Main Site | Source

 

[chrcoluk]

No thanks, GPU's with no VRAM upgradability is bad enough, dont need same problem on system memory.

 

[londiste]

While as a consumer this is a bit worrying it is not like many/most mobile uses have memory modules these days. As we go for thinner and lighter the memory tends to be soldered anyway.
On-package memory has a couple big benefits primarily due to shorter traces from CPU/IMC to RAM - lower power consumption, easier to reach higher clock speeds. Physically smaller package is a nice to have for mobile applications as well.

 

[tabascosauz]

I suppose the bigger issue here is thermal management - M1 and M2 draw far less power than most configurations of Alder Lake and Raptor Lake. It might not be a longevity problem having DRAM in Apple, but here?

No thanks, GPU's with no VRAM upgradability is bad enough, dont need same problem on system memory.

In case you haven't noticed, you can't upgrade soldered down DDR4/DDR5 and LPDDR3/LPDDR4/LPDDR5(x) either.

SO-DIMMs haven't been a thing on most non-gaming consumer laptops for a long time, and neither is DDR5 desirable from a performance standpoint for larger iGPUs regardless (680M and 780M). Laptops that are big enough for RAM expandability shouldn't be changing anytime soon just because Intel is coming up with different packaging for thinner laptops.

 

[JustBenching]

I suppose the bigger issue here is thermal management - M1 and M2 draw far less power than most configurations of Alder Lake and Raptor Lake. It might not be a longevity problem having DRAM in Apple, but here?

Do they though? Raptor lake (and zen 4) are more efficient than m1 and m2. They are just pushed to draw higher wattage cause they have the room. Nothing prevents you from locking them down to 60w and still smack the m's around. The CPU portion of the apple CPU's are mediocre. The GPU is where they excel at, probably because of that huge memory bandwidth they enjoy.

 

[randomUser]

Do they though? Raptor lake (and zen 4) are more efficient than m1 and m2. They are just pushed to draw higher wattage cause they have the room. Nothing prevents you from locking them down to 60w and still smack the m's around. The CPU portion of the apple CPU's are mediocre. The GPU is where they excel at, probably because of that huge memory bandwidth they enjoy.

I don't have newest and greatest intel+win laptop. But i have 3 year old ultrabook with 8250U.
Usual scenario. I open the lid and it keeps on doing something drawing 30-45W of power for like 10-15 mins. Then, it sits at around 8W with high count of spikes up to 20W.

Now i also have M1 macbook pro 13" (which is 1,5 years old now), which is same formfactor as the win ultrabook. When i open lid, it draws 3W. Some spikes up to 10 watts, but rarely. It's pretty much always staying in the range of 3-5W power consumption.

The intel laptop is considerably weaker. Doing same tasks longer than macbook.

I have become so sceptical of intel and windows based laptops, i would be afraid to buy one now seeing how intel stagnated for decades doing same chips over and over again.
macbook and ultrabook cost the same at that time i bought them.

 

[tabascosauz]

Do they though? Raptor lake (and zen 4) are more efficient than m1 and m2. They are just pushed to draw higher wattage cause they have the room. Nothing prevents you from locking them down to 60w and still smack the m's around. The CPU portion of the apple CPU's are mediocre. The GPU is where they excel at, probably because of that huge memory bandwidth they enjoy.

Any hardware gets more efficient if you power limit, which isn't even guaranteed to be possible on either Ryzen or Intel laptops. That's not how it behaves as it comes as a product to you. So yeah, there is something that prevents you from locking an -H down to 60W, namely that it's not representative of how the product is shipped, and it's not something relevant to the average user or reasonable laptop reviewer. As the products are, Zen 4 is the only one that can currently give Apple a run for their money in efficiency in the relevant power segment.

In any case, it looks like Meteor Lake (at least the leaks we've seen so far) is more aimed at the -U and -P segment, the latter of which is in line with larger M1/M2 mobile configurations in max power consumption (around 50-60W), so maybe it's still a non-issue for the designs that Intel is looking for with this specific packaging of Meteor Lake.

 

[JustBenching]

I don't have newest and greatest intel+win laptop. But i have 3 year old ultrabook with 8250U.
Usual scenario. I open the lid and it keeps on doing something drawing 30-45W of power for like 10-15 mins. Then, it sits at around 8W with high count of spikes up to 20W.

Now i also have M1 macbook pro 13" (which is 1,5 years old now), which is same formfactor as the win ultrabook. When i open lid, it draws 3W. Some spikes up to 10 watts, but rarely. It's pretty much always staying in the range of 3-5W power consumption.

The intel laptop is considerably weaker. Doing same tasks longer than macbook.

I have become so sceptical of intel and windows based laptops, i would be afraid to buy one now seeing how intel stagnated for decades doing same chips over and over again.
macbook and ultrabook cost the same at that time i bought them.

The 8250u is a midrange cpu from 2017....

Also if it is under load on booting windows it means you have lots of stuff starting with your windows os. I have a 6900hs, it just boots and sits there at idle 2 to 3 watts.

Any hardware gets more efficient if you power limit, which isn't even guaranteed to be possible on either Ryzen or Intel laptops. That's not how it behaves as it comes as a product to you. So yeah, there is something that prevents you from locking an -H down to 60W, namely that it's not representative of how the product is shipped, and it's not something relevant to the average user or reasonable laptop reviewer. As the products are, Zen 4 is the only one that can currently give Apple a run for their money in efficiency in the relevant power segment.

In any case, it looks like Meteor Lake (at least the leaks we've seen so far) is more aimed at the -U and -P segment, the latter of which is in line with larger M1/M2 mobile configurations in max power consumption (around 50-60W), so maybe it's a non-issue for the designs that Intel is looking for with this specific packaging of Meteor Lake.

I've not yet come across a laptop that doesn't allow you to at least select between silent performance and turbo modes. Silent modes usually power limit the cpu to 30w or less (it's 25 on the 3 laptops I have). You don't get into the bios for that, you do it right there from the desktop on a pre-installed app. There is even a keyboard button on some laptops dedicated for that.

 

[Assimilator]

Do they though? Raptor lake (and zen 4) are more efficient than m1 and m2.

Sorry mate, I'm the biggest Apple hater there is but that statement is just pure rubbish. The power efficiency of Arm has always been ahead of x86 - while AMD has made remarkable strides there with Zen 4 mobile, they still tend to overvolt their CPUs to allow them to boost to unnecessarily high frequencies. As for Intel, their ancient and creaking Skylake+++++ re-re-respins aren't even in the same league as either Arm/Apple or AMD.

 

[bug]

While as a consumer this is a bit worrying it is not like many/most mobile uses have memory modules these days. As we go for thinner and lighter the memory tends to be soldered anyway.
On-package memory has a couple big benefits primarily due to shorter traces from CPU/IMC to RAM - lower power consumption, easier to reach higher clock speeds. Physically smaller package is a nice to have for mobile applications as well.

This. Even for models where there is a RAM slot, since the other "stick" is soldered, you'd need another stick that's identical or at least uses about the same timings.

And if these will come in dual channel config, they will beat the crap out of cheap laptops that only come with one memory "stick".

 

[JustBenching]

Sorry mate, I'm the biggest Apple hater there is but that statement is just pure rubbish. The power efficiency of Arm has always been ahead of x86 - while AMD has made remarkable strides there with Zen 4 mobile, they still tend to overvolt their CPUs to allow them to boost to unnecessarily high frequencies. As for Intel, their ancient and creaking Skylake+++++ re-re-respins aren't even in the same league as either Arm/Apple or AMD.

Correct, if you allow intel or amd to draw an endless amount of power to turbo boost apple beats them handily in efficiency, but it also loses handily in performance. If you cap them so they don't draw an unlimited amount of power, then they are actually more efficient than the cpu part of the m cpus.

 

[AnarchoPrimitiv]

Intel probably HAS to do this for their new mobile iGPUs to compete with AMD's next gen. Just consider how good the current 780m is at 12CUs, and then think about how the next Zen5 APUs will have 16 CUs that are upgraded from the current RDNA3 ones ("RDNA3.5")....and then I've read rumors about an AMD APU that will have 40 CUs? 40 of the current RDNA3 CUs would be amazing as it is, but 40 of the improved ones are going to put up some serious performance against Intel's 128EUs (just to put that into perspective, the PS5 has 36 RDNA2 CUs, and the next-get AMD APUs could have 40 RDNA3.5 CUs!!)

 

[AusWolf]

What's the point of this? It's not like 120 GB/s is such a massive improvement over standard DDR 5. It's just an extra point of failure, imo.

 

[Nanochip]

I suppose the bigger issue here is thermal management - M1 and M2 draw far less power than most configurations of Alder Lake and Raptor Lake. It might not be a longevity problem having DRAM in Apple, but here?

you do realize the meteor lake cpu core chiplet is built on intel 4 EUV? So it should be more power efficient than alder lake and raptor lake.

 

[tabascosauz]

you do realize the meteor lake cpu core chiplet is built on intel 4 EUV? So it should be more power efficient than alder lake and raptor lake.

Neither 10ESF nor 4 years of "big" P-cores lineage from Ice Lake onwards solved Intel's PL1/PL2 conundrum in laptops, or change their fundamental approach to short/long term power management. Since 6 cores became a thing their mobile philosophy has always been to push to Tjmax and stay there until Tau expires.

Why should another new node and incremental P-core update be radically different? Any power or thermal headroom they'll just continue to exploit to bump up clocks, as they have done for years (and as is only natural to do).

Intel has been touting Meteor Lake as a unique "efficiency" milestone but I'll believe it when I see this drastic revamp of mobile power limits. In any case, already observed above that thermal issues probably not so likely in -U and -P.

 

[Nanochip]

Neither 10ESF nor 4 years of "big" P-cores lineage from Ice Lake onwards solved Intel's PL1/PL2 conundrum in laptops, or change their fundamental approach to short/long term power management. Since 6 cores became a thing their mobile philosophy has always been to push to Tjmax and stay there until Tau expires.

Why should another new node and incremental P-core update be radically different? Any power or thermal headroom they'll just continue to exploit to bump up clocks, as they have done for years.

Intel has been routing Meteor Lake as a unique "efficiency" milestone but I'll believe it when I see this drastic revamp of mobile power limits.

I think you will be pleasantly surprised by meteor lake. But there’s always zen if you’re not an Intel x86 fan. Phoenix mobile 7900 series is impressive. But according to leaked benchmarks, meteor lake is even faster at lower wattage. We’ll know the details soon enough.

 

[AusWolf]

you do realize the meteor lake cpu core chiplet is built on intel 4 EUV? So it should be more power efficient than alder lake and raptor lake.

Power and thermals are not the same thing. In fact, as chip density increases, cooling becomes more difficult.

 

[bug]

What's the point of this? It's not like 120 GB/s is such a massive improvement over standard DDR 5. It's just an extra point of failure, imo.

The point of LPDDR is just that: low power.
The point of embedding RAM is space saving. And probably reliability, since it makes all traces physically shorter.

 

[Nanochip]

Power and thermals are not the same thing. In fact, as chip density increases, cooling becomes more difficult.

Fair point but cpu package temperature is correlated with power consumption is it not? Lower overall consumption, with an adequate cooling system, should mean temps will be under control.

 

[Denver]

"While this approach simplifies manufacturing, enabling slimmer notebook designs"

Unnecessary, it is better to have adequate cooling and stable performance.

 

[AusWolf]

Fair point but cpu package temperature is correlated with power consumption is it not? Lower overall consumption, with an adequate cooling system, should mean temps will be under control.

CPU temperature is correlated to power consumption, but not only to power consumption. There's a reason why modern CPUs (such as Intel mobile, or AMD chiplets) are so hot even when running on low power. More power means more heat. But higher density also means more heat, so there's a trade-off.

 

[bug]

Power and thermals are not the same thing. In fact, as chip density increases, cooling becomes more difficult.

Except there's no chip density increase here. The CPU is the same, the RAM is the same... They're just placed closer to each other.
And I'm pretty sure Intel's engineers aren't dumb to the point they design a CPU for space-saving, only to have that negated by the need of a bigger heatsink.

 

[dyonoctis]

Sorry mate, I'm the biggest Apple hater there is but that statement is just pure rubbish. The power efficiency of Arm has always been ahead of x86 - while AMD has made remarkable strides there with Zen 4 mobile, they still tend to overvolt their CPUs to allow them to boost to unnecessarily high frequencies. As for Intel, their ancient and creaking Skylake+++++ re-re-respins aren't even in the same league as either Arm/Apple or AMD.

While I agree that Alder laker and raptor lake are not efficient, they aren't Skylake +++++. Can we let that freakin arch die already ? It hasn't been relevant since 2021

What's the point of this? It's not like 120 GB/s is such a massive improvement over standard DDR 5. It's just an extra point of failure, imo.

DDR5 sodimms are limited to jedec speed and loose timings. 120GB/S is still a ~40% improvement over DDR5 5600 wich is the max available on so dimms atm. That packaging that might be good for the igpu too. Dell proposed a solution that they are pushing to be the new standard on laptop (camm) but people are weirdly upset about it and would rather stay on a from factor that reached its limits for ever.

 

[TheinsanegamerN]

I guess I dont see the big ground breaking thing.

AMD has used LP 7500x memory for a year now.

Apple did the on package memory thing 2 years ago.

What's the point of this?

Space savings, lower power use, better signaling allowing for faster speeds.

It's not like 120 GB/s is such a massive improvement over standard DDR 5. It's just an extra point of failure, imo.

Meteor lake is going to feature second gen iris cores, which are supposedly much faster then what they currently use, and a larger 128EU core over the current 96 EU. The current 96 EU benefits greatly from DDR5 VRAM in GPU tests. So a larger, faster GPU will want MOAR bandwidth.

We have already seen the effects form AMD, ryzen models with LP 7500 memory obliterate models with LP 6400 in benchmarks, to say nothing of those saddled with 5200 mhz sodimms.

Sorry mate, I'm the biggest Apple hater there is but that statement is just pure rubbish. The power efficiency of Arm has always been ahead of x86 - while AMD has made remarkable strides there with Zen 4 mobile, they still tend to overvolt their CPUs to allow them to boost to unnecessarily high frequencies. As for Intel, their ancient and creaking Skylake+++++ re-re-respins aren't even in the same league as either Arm/Apple or AMD.

There's lots to bash apple over, but efficiency is not one of them.

I'd LOVE to like the apple M series chips, but the total lack of backwards compatibility and vulkan support is a total deal breaker.

 

[Assimilator]

Meteor lake is going to feature second gen iris cores, which are supposedly much faster then what they currently use, and a larger 128EU core over the current 96 EU. The current 96 EU benefits greatly from DDR5 VRAM in GPU tests. So a larger, faster GPU will want MOAR bandwidth.

We have already seen the effects form AMD, ryzen models with LP 7500 memory obliterate models with LP 6400 in benchmarks, to say nothing of those saddled with 5200 mhz sodimms.

It's Intel's iGPU, though, so bandwidth is irrelevant because nobody sane is going to try use that POS for gaming. The only thing you use an Intel iGPU is for driving 2D displays not rendering 3D graphics, given that it's barely capable of the former at the best of times.