News Intel unwraps Lunar Lake architecture: Up to 68% IPC gain for E-cores, 16% IPC gain for P-Cores

[bit_user]

intel tends to conflate IPC with general performance gain from clock speed increases in most of their marketing. so you'll often see things like 68% ipc gain

I scrutinized all of the slides. Those slides do indeed state iso-frequency. The end-notes confirm this.

I haven't personally seen them abuse the term "IPC" in the past, but I won't contest you on that.

what they're comparing it to is a 2.4ghz part vs a 3.2ghz part, meaning almost all the 68% "ipc gain" is from a clock speed improvement.

They give separate figures for single-threaded (and multi-threaded) speedup.

The way they rigged both IPC and the speedup figures is by comparing against LP E-cores, which are disadvantaged on memory access, clock speed, and lack L3 cache. In the multithreaded speedup case (as the article points out), they're literally comparing 4 cores against 2!

if they do get 68% IPC gain that's great, but i doubt it.

I expect it'll be more like 60% (fp) when compared to a Crestmont core on the CPU die. We'll have to wait and see.

 

[snizzledizzle]

Can people please stop using the word Architect as a verb please ?
"Redesigned" would have been perfect for this article.
Architected or Rearchitected do not exist.
Cheers.

You may want to consult a dictionary on that one.

Even googling the definition of the word provides this as the second definition:

verb
COMPUTING
verb: architect; 3rd person present: architects; past tense: architected; past participle: architected; gerund or present participle: architecting

1. design and configure (a program or system).
   "few software packages were architected with Ethernet access in mind"

 

[oofdragon]

Lmao Intel just going to take a beating from AMD once again

 

[bit_user]

Hey, does anyone want to speculate on what the C-shaped strip of grey material is, to the left of the CPU? Is it to balance the mechanical loading of the heatsink?

​

 

[DavidC1]

The way they rigged both IPC and the speedup figures is by comparing against LP E-cores, which are disadvantaged on memory access, clock speed, and lack L3 cache. In the multithreaded speedup case (as the article points out), they're literally comparing 4 cores against 2.

Are you people serious here?

It's faster than Raptor Cove in both Integer and FP. It was previously "Skylake-level" if you took into account that in FP, it was hugely behind.

Golden Cove is 23% faster than Raptormont in Integer and 66% faster in FP. Now the gap is in slight favor for Skymont.

 

[thestryker]

My question is, what does this mean for Intel's 5N4Y plan?

Sitting in mid 2024, and the expected 5th node to come before the end of 2025, when in actuality they've only achieved one of those nodes. Being on TSMC for this round, means 3 other nodes in 18 months. I just don't see it.

5N4Y if I'm remembering correctly included Intel 7 so we've got 7 (ADL/RPL and SPR/EMR), 4 (MTL), 3 (GNR/SRF), 20A later this year (ARL), and 18A next year (PNL, and probably others I can't think of).

>LNL shares core architecture with ARL and that's about it. It doesn't share the same tile configuration, core configuration or graphics technology.

LNL is what Intel has to offer new for this year, and that's what will be benchmarked against the competition. That's where all the hype is. It will go up against Ryzen 370 with 12 cores w/ 24T, and the result will likely be a massacre. That's not to mention all the other existing parts, including MTL.

You're arbitrarily comparing different market segments because they're "new" which doesn't make any sense.

BTW, "tile configuration" and "core configuration" aren't exactly "marketing knowledge," but I appreciate your confusion. It must be tough to tell them apart.

Market and marketing mean two different things. Your attempt at being pedantic or whatever you're trying doesn't actually work. You're contending two different products are in fact the same thing when they're not.

 

[Silas Sanchez]

Be ready for another 300+ Watt CPU for about 16% IPC uplift, same as AMD, no surprise in the industry.

The picture im getting is we are hitting a 6ghz wall with no way in the short term to get past it without high power draw, the wall is a barrier to easy gains by simply raising clock speeds. The engineers at AMD and Intel must know this, they know its hard to raise clock speeds without all that extra heat and power draw coming on in an exponential way, advances in architecture can only do so much for gains.
This is the problem with shrinking transistors, particularly the thickness of the gate oxide insulating layer, as the switching speed increases, the capacitive nature of the transitor comes more and more into play and leads to more current leaking out in various ways, that leaked current must be paid for and it gets turned into heat. A combo of smaller transistors & faster switching is what bad.
From Intels recent desktop cpus, stability is an issue so im kinda notsurprised AMD went soft and didn't raise the clock frequency, that shows me they care about stability and dont wont headaches.
So i think we wont see big gain and efficiency jumps unless there are some new transistor designs and better coolers, particuly they need to do somethimg to improve the bottleneck at the cpu-cooler interface. Exotic materials with higher thermal conducivity-more so at the actual base of the air cooler, high perf thermal paste, heatsink and cooling on the back of mobo? larger area cpus? With the base of air coolers, the thicker the narrow cross section is the worse the bottle neck is. But no matter the cooling (within reason) these cpus nowadays just crank out more perf while sitting at their temp limits when you increase the cooling.

 

[Giroro]

Does anybody know why Intel is claiming HEVC has 1.4x the file size compared to AV1?
The testing I've seen before, when AV1 was new/unsupported, showed the quality per bitrate/file size of AV1 to be much closer HEVC, and sometimes behind - with AV1's advantage primarily being in its cost to license.

When/why did people start saying AV1 was so much smaller?

 

[Roland Of Gilead]

5N4Y if I'm remembering correctly included Intel 7 so we've got 7 (ADL/RPL and SPR/EMR), 4 (MTL), 3 (GNR/SRF), 20A later this year (ARL), and 18A next year (PNL, and probably others I can't think of).

Well, IIRC, Intel 4 was the second node of the 5, with Lunar Lake to follow. They've only dropped that as of yet.

I guess that if they are currently running with TMSC, then how will the transition work for the other 3 nodes expected by end of 2025? It's seems a bit of a stretch.

 

[Eximo]

Hey, does anyone want to speculate on what the C-shaped strip of grey material is, to the left of the CPU? Is it to balance the mechanical loading of the heatsink?

​

​

I suspect that matches the height of the memory chips and covers up some surface mount capacitors that you would normally see around a CPU. If you didn't have it, I can easily imagine the mounting pressure to be uneven and that would put a lot of pressure on the corners of the silicon there. ANd since it isn't monolithic, that would easily cause localized damage.

 

[JamesJones44]

LNL's on-package memory will be a significant limitation for adoption. It works for Apple, because Apple owns its walled garden. I understand the rationale, which is to lower power use by any means to compete against ARM, but OEMs will want to have more diverse memory configurations. Along with low core-count, LNL will only fit into the premium ultramobile niche, while Arrow Lake mobile will fill out the rest of the segments

If we are being honest, a huge majority of laptop buyers will never upgrade their RAM or storage. It might turn a few people off, but I don't think it will be enough to really bother Intel from a sales perspective.

I'm a little surprised it's limited to 32 GB however. Even Apple allows 64+ GB on the Pro/Max lines.

 

[thestryker]

Well, IIRC, Intel 4 was the second node of the 5, with Lunar Lake to follow. They've only dropped that as of yet.

I guess that if they are currently running with TMSC, then how will the transition work for the other 3 nodes expected by end of 2025? It's seems a bit of a stretch.

Intel 3 is already shipping as Sierra Forest officially launched and Intel 20A is Arrow Lake so it's just 1 node that needs to appear next year.

 

[Notton]

Hey, does anyone want to speculate on what the C-shaped strip of grey material is, to the left of the CPU? Is it to balance the mechanical loading of the heatsink?

​

​

The C-shaped thing is a "Stiffener" on Intel's slides
It might also serves as a load balancer, although I haven't seen those on any recent laptop, no matter what they are using.

 

[Roland Of Gilead]

Intel 3 is already shipping as Sierra Forest officially launched and Intel 20A is Arrow Lake so it's just 1 node that needs to appear next year.

Nice that you quoted an Intel sponsored piece 😉

 

[usertests]

CAMM2 LPDDR5 maxes out at 7500
Lunar Lake gets 8500

CAMM2: Revolution in the memory market: Up to 128 GB capacity, 8.5 Gbps speed and space savings | igor´sLAB

CAMM (Compression Attached Memory Module) is a memory standard specially developed for laptops and portable platforms to offer maximum performance with minimum space requirements.

www.igorslab.de

https://www.notebookcheck.net/DDR6-...-speeds-between-8-8-and-21-Gbps.840140.0.html

There are already plans to offer CAMM2 in 8533, same as Lunar Lake, and 9600 soon after. DDR6 CAMM could end up peaking at ~21000 MT/s.

The benefits of Lunar Lake's memory-on-package are primarily lower costs, power consumption, and smaller board space. It's also predictable, you know that you're getting reasonably fast memory.

External and on-package memory could co-exist in the future, but Lunar Lake isn't the showcase of that.

Does anybody know why Intel is claiming HEVC has 1.4x the file size compared to AV1?
The testing I've seen before, when AV1 was new/unsupported, showed the quality per bitrate/file size of AV1 to be much closer HEVC, and sometimes behind - with AV1's advantage primarily being in its cost to license.

When/why did people start saying AV1 was so much smaller?

I remember seeing that HEVC and AV1 were much closer years ago. What probably happened is that the AV1 encoders improved. Encoding efficiency isn't locked on release, it can get better with further development. I was surprised to see that they are claiming that VVC is only +10% better than AV1, looks good for AV1 and gives them the time they need to get AV2 out the door.

One % doesn't tell the whole story since there are scenarios where one codec can be better/worse.

 

[thestryker]

Nice that you quoted an Intel sponsored piece 😉

Haha, yeah, you gotta go with what you have though! It's the only piece from a reputable source that actually had any real testing that I'm aware of.

 

[TheSecondPower]

That's neither here or there though.
If foveros doesn't have the lag, again I don't know if it does or not, then the intel quote was and still is justified.

As I recall, the Intel quote about "glued chips" specifically targeted AMD's first-generation Epyc CPUs, which had 4 chips each with its own memory controller and could operate in non-unified memory mode or unified memory mode (as a setting in the BIOS), with unified memory mode not being ideal because there was a huge amount of latency when accessing something on another chip's memory. I don't recall Intel ever attacking Zen 2 with this claim, since Zen 2-based Epyc and Ryzen processors have a single memory controller for all chips and usually perform well in workloads that were designed for a unified memory architecture. And Intel's chips should theoretically do even better since they use a silicon interposer and Lunar Lake has the memory controller on the same chip as the cores.

One chip at a time. They are announcing what they have nearly ready.

It has been somewhat typical for Intel to start with ultrabooks "U", then move on to laptop "H", and finally do desktop "S"

Though I want to say Alder lake desktop came first which was atypical. Q4 2021 Alderlake S vs Q1 2022 for Alderlake U/H

When Alder Lake came out Tiger Lake (Willow Cove, which Sunny Cove with improved cache on 10SF) was considered pretty successful on mobile but Rocket Lake (Cypress Cove, which was Sunny Cove on 14nm+++++) was considered a failure on desktop. Intel really needed a win on desktop while mobile was going strong.

 

[TheSecondPower]

It'll be interesting to see the results of multi-threaded benchmarks on Lunar Lake. In theory it's at a disadvantage with 4LPE + 4P compared to the 2LPE + 2P + 8E of Meteor Lake U-series. But the E cores have 50% higher IPC and the P cores are more numerous and have 14% higher IPC. In theory on the new node the clock speeds should be the same or higher. I think LNL should perform better in heavily threaded workloads. However in MTL-U the 2P + 8E cores are on the ring bus whereas the 4LPE cores in LNL are not on the same bus as the 4P cores, so in many threaded workloads—espcially ones sensitive to core-to-core latency or cache—LNL should perform worse. But all the cores can access the new 8 MB side/SLC cache which might offset that disadvantage.

 

[baboma]

>>but OEMs will want to have more diverse memory configurations.

>If we are being honest, a huge majority of laptop buyers will never upgrade their RAM or storage. It might turn a few people off, but I don't think it will be enough to really bother Intel from a sales perspective.

As said, it's not about buyers, but OEMs. People are comparing current soldered LP RAM as being the same as on-package RAM, but the big difference is that RAM configuration is determined by the OEM at the point of manufacture, depending on whatever the need is at that point in time. With on-package memory, the OEM must forecast its RAM allocation needs when it places its purchase order, which is much longer in advance. The difference in flexibility is huge.

That said, LNL will be bottlenecked by its low core-count to fit into a fairly narrow niche (premium ultraportable), so the memory probably won't be a defining constraint, given the few SKUs it will occupy.

>I'm a little surprised it's limited to 32 GB however. Even Apple allows 64+ GB on the Pro/Max lines.

32GB is fine for LNL's niche. With 8C/8T, LNL isn't a powerhouse. Power users who want more RAM will likely opt for the Ryzen AI part, which should be much more powerful, and have higher demand. Ryzen AI is also configurable for any thermal envelope (IIRC 15-54W), so we should see it deployed across a wider range of premium segments.

I hold Intel stock, so I'd be the first in line to want it to do well with LNL. But the difference in core/thread count is so stark that you'd have to drink a lot of technobabble KoolAid to think LNL stands a chance against Ryzen.

 

[DS426]

Meteor Lake was quite unimpressive in terms of gen-on-gen gains but kind of a validation for Intel's tile strategy. Glad to see Intel really innovate and move the needle on Lunar Lake, although we can only wait and see how much this helps on Arrow Lake. I also didn't realize they were still using their own internal EDA tool, so kind of funny to me how much it helped them lunge forward... and also now have the late-to-the-party ability to adapt 99% of chip design to lithography nodes other than the intended target node.

 

[rluker5]

This is a lovely ad for Intel's foundry business.
/s


Uh, no. Granite Rapids uses Redwood Cove, not Lion Cove, and Sierra Forest uses Crestmont, not Skymont.





Read the slide carefully, @PaulAlcorn . In the lower left, it says this is iso-frequency!

​

Spoiler

​

On the next slide, Intel is clear that Raptor Cove still offers better peak performance.

​

Spoiler

​

That's crazy. Sure the Skymont cores are running at lower clocks, but Raptor Cove non-HT becoming the new baseline for the weakest threads on Arrow Lake with first at 24 then maybe 40 of them? Probably hundreds in a Xeon?

I've still got a 22nm Silvermont tablet that is basically just good for watching videos (and playing HL2 at 1080p) and it is SO far from Raptor cove performance. Also have an Airmont phone that has developed software compatibility issues from an abject neglect of updates for running emulated ARM. Those Atoms are rising so fast they're soon to be nearly the fastest thing out there. A couple of months soon.

​

​

TBH, I'm kinda disappointed it's only offering up to 1.2x perf/W, especially considering it's on TSMC N3B vs. Raptor Cove on Intel 7. Even if they were on the same node, I'd expect a bigger peak efficiency difference than that.

I wonder if Intel is comparing mainstream efficiency or their lab tuned efficiency? RPL can run way more efficiently at a bit lower clocks. I've messed with my 13900kf and 13600k and they run 4.8GHZ P-cores and 3.7 GHz E-cores at very low volts with HT disabled. Those clocks seem like the sweet spot for undervolting RPL.
If your laptop lets you disable HT you could try it with XTU. I've got my 13900kf at -.15v and my 13600k at -.18v with both running reduced volt curves to start with.

Raptor Cove can be much more efficient than the public perception, it just can also run very inefficiently if you let it clock high enough. The Intel volt curves are so far off that I don't know what to make of this latest one you posted. There seems to be so much fat to be cut.

Here's a couple screenshots of those AFAIK stable for everything clock/voltage combos:

Spoiler

I don't know how to undervolt as well on a Gigabyte mobo as Asus and for some reason the Gigabyte mobo isn't including system agent in the total watts used by the 13600k. It should really be ~7w higher for CPU package power.

I've had this bios set for a while. Also I've turned off 1/2 the E-cores. If I reduce the clocks with a windows power plan I can run most games at a smooth 60 fps using around 20 watts with the CPU. But not Phantom Liberty driving the roundabout next to Heavy Hearts, that takes like 40w.

 

[kjfatl]

Thank your for your clear advice as to exactly what I should do!

The amount or RAM needed depends heavily on the user and his application. For many, the Chrome Book type users or those who use the web for streaming, 16GB should be fine for a decade. The big issue they have is battery life with battery life of 24 hours or more. When I am running large circuit simulations, 128GB is tolerable and 2 to 3 hours of battery life is at the high end. My preference though is to use the server farm for these tasks. RAM mounted on the substrate may actually give the user a little leverage on the price side. It will result is a thinner, lighter, more reliable machine.

 

[DavidC1]

That's crazy. Sure the Skymont cores are running at lower clocks, but Raptor Cove non-HT becoming the new baseline for the weakest threads on Arrow Lake with first at 24 then maybe 40 of them? Probably hundreds in a Xeon?

This core is a monster for what it is. I'm absolutely ecstatic for what's coming. The P core design's days are numbered.

And Lunarlake is a HUGE improvement in terms of power. They got a winner here.

 

[baboma]

For a preview of how LNL will fare against Ryzen AI & QC X, we can look at the full Asus laptop line-up shown at Computex (courtesy of Videocardz):

https://videocardz.com/newz/asus-in...-and-intel-core-ultra-200v-lunar-lake-laptops

Of the 12 listed models:
. 3 ProArt (creator series) - Ryzen AI has 2 wins, both with dGPUs; QC has 1.
. 2 Zenbook S (premium ultraportable) - Ryzen AI has both wins.
. 3 Vivobook S (all-around midranger) - Ryzen AI has 2 wins; QC has 1.
. 1 ExpertBook (premium business) - Lunar Lake has the win.
. 2 TUF Gaming (value gaming) - Ryzen AI has both wins, both with dGPUs.
. 1 ROG Zephyrus (premium gaming) - Ryzen AI has the win, with dGPU.

Aside from the sheer number of design wins, one thing to note is that only Ryzen is used in tandem with dGPUs, ie in use cases where power takes priority over power efficiency. This is a large proportion of PC laptops, and is an argument why ARMs won't make much inroads into the x86 laptop market.

Yet, even in use case where power efficiency is ostensibly more important, ie premium ultraportable, Ryzen still gets the win over LNL. LNL's sole win is in the premium business segment.

LNL will likely get more wins with other vendors, if only because of Intel's massive negotiating leverage with OEMs to ply its wares. But this is a good first look of how OEMs are judging the LNL/Ryzen/QC contest.

 

[dalauder]

I'd like to see extensive power consumption/performance benchmarks for the next gen. Intel, AMD, ARM CPUs and how slower and faster SSDs affect the battery life.

You won't. You'll see an Intel CPU pulling 350W on an open table with a 360mm water cooler compared to an AMD system pulling 120W and an ARM system that costs half as much as either pulling 30W. Then Intel will win the performance crown.

Far too long I've been asking for power vs performance vs thermal results.... All anyone keeps measuring though is this chip greater than/less than the last chip in some meaningless synthetic benchmark. So I wouldn't hold your breath waiting on this data if I were you. Likely the only way to get the data you seek is to formulate, test and measure everything yourself.

Exactly. You'll never see it.

Intel is the king of marketing. AMD needs to release a 9990X Black Edition that pulls 600W and leave it to the benchers to cool it. It's what NVidia does with its GPUs.