Discussion Qualcomm Snapdragon Thread

[The Hardcard]

The most obvious way you can tell QC has a better architecture is that, like Apple, they really can just scale well with both core counts and idle power. They don’t even have E Cores yet, and when they do that’ll be another boost here even if is just shrunken P Cores (see AMD’s thing which does give them efficiency boosts at some clocks).

Intel is going to rely very heavily on their E Cores I think way more than Apple does, and they probably won’t quite match Arm’s or Apples and quite frankly, I suspect that Qualcomm’s P Cores (including SoC power anyways) are about as efficient at the same performance level — which is another key difference between Intel’s P and Cortex X, Apple’s cores, and Qualcomm’s which actually still can perform well going pretty low. Apple’s E Cores have a genuine lead in their case over the P cores, but the A7x vs X lead is a bit smaller and a bit more about area and needing every last drop for phones — if any of these three wanted to ship P Cores only in laptops, they could still get solid battery results (diminished for Apple for sure ofc) albeit it’s a less area-efficient way to do things.

This is made most obvious by…. Qualcomm’s own claims. They can ship a 12C standard part that can compete with Intel’s LNL and then AMD’s Strix on both ends.

Anyway, on some level Intel still has to fix their ringbus and P core issues, or they need to move to a different fabric structure. LNL is a good crutch though and an excellent step forward.

View attachment 100665

Before someone says this is what Apple does — that is false, on MacOS. There is a ramping and software (with GCD, a hierarchy of QoS) element to core utilization but “most real workloads” do not stay isolated to the E Cores, you can just watch this yourself in light use. They get a ton of use but it’s not this dramatic. The reason Intel will make a different tradeoff is because their P cores and the ring they’re on are too bloated, and also don’t scale down very well, and their E Cores are faster than Apple’s*.

*Apple E core Spec results are like in the 2.5-3.5 range around .5-.8W, Intel’s MTL LP E Cores are similar at peak, but drawing 5+W. LNL Skymont LPE cores give a 1.7x perf uplift iso-power or a 2x uplift at more power. Putting them at like 5.5 to 6.5 SpecInt performance, depending on power.

Interesting times ahead tbh.

It is like Apple does, it’s just that the Mont cores much more capable. The guy at Eclectic Light has done extensive measurements on core utilization on Apple Silicon. Apple is trying to do the same thing as Intel, in terms of staying on the E cores as much as possible. An in fact their E cores have gained significant increased capability just like the Monts have.

The key difference is Apple keeps their E cores sub watt. The capability and that power is amazing, the Apple E cores are the performance per watt kings. But with half watt peak power, the P cores come in much sooner for even basic mainstream tasks.

Skymont can handle so much more. The Lakefield 1 P 4 E vision makes sense with this design though I think Lunar Lake’s 4 by 4 on N3B makes it much more of a lock, assuming they got the design correct.

 

[SpudLobby]

It is like Apple does, it’s just that the Mont cores much more capable. The guy at Eclectic Light has done extensive measurements on core utilization on Apple Silicon. Apple is trying to do the same thing as Intel, in terms of staying on the E cores as much as possible. An in fact their E cores have gained significant increased capability just like the Monts have.

The key difference is Apple keeps their E cores sub watt. The capability and that power is amazing, the Apple E cores are the performance per watt kings. But with half watt peak power, the P cores come in much sooner for even basic mainstream tasks.

Skymont can handle so much more. The Lakefield 1 P 4 E vision makes sense with this design though I think Lunar Lake’s 4 by 4 on N3B makes it much more of a lock, assuming they got the design correct.

i’ve seen eclectic’s work and this is not actually true. For high priority user interface stuff, the threads run on P Cores and if none are available they go to E Cores.

They are mainstream use for a broad amount of tasks but “most real workloads use the E Cores” is absolutely not how Apple Silicon works. The E cores also did reach like 2 on SpecInt and now like 4.0 plus with M3.

Intel’s LP E Cores with MTL were doing 3.3 or so at most with SpecInt. At a 1.7X uplift they’ll be at around 5.61 at the same cluster power (and at a package level they were around 5-6W for this with everything else shut off in MTL, so that’s interesting). At 2x and more power, they’ll be at 6.6.


Or for the same 3-3.3 SpecInt the cluster draws 1/3 the power in MTL. Which isn’t bad if that’s moving from 5-6W to 1.5-2W, but Apple’s and Arm’s from a whole platform power pov are doing more than that in Arm’s case around the same power with last generation IP and on N4, or they’re doing it at vastly less power (at least in the 2GHz A17 example.)

intel’s Skymont in LNL based off what we know here looks like the closest Intel has been towards a phone A7x core or a big core from Apple/Arm, but what you’re missing is Apple can draw the kind of power Skymont will with their P Cores and get more performance, and their E Cores don’t scale up as high for performance or certainly didn’t with M1/2. So it very much is a different strategy because they have a different set of cards to play.


Sources here on power/perf:



(And package here is understating things as it doesn’t include RAM or power delivery, but the Arm examples above do and are measured physically).


(A715 is last-gen and on N4 ehre, but even when measured including the whole platform it does phenomenally by contrast to Crestmont and I am skeptical SKT will close that gap even at 1/3 the power — math doesn’t work there, much less Apple’s.)




Based on MTL’s LP E core perf/w which is with Compute Tile off and just the Island, this will put them much closer to Arm cores but even that “1.7x” or “2x” at even more power demonstrates exactly why Apple and others can use their p cores more often.

When Intel says they’re using the E Cores, that cluster is consuming way more power with also up to 50-60 more peak perf than an A7x or Apple E core would.

Which also makes the strategy and difference from Apple or QC much more believable — those two can just use their P cores and get more perf at the same power. But for Intel, using Skymont even at 5W probably makes sense for good enough ST without hurting power from turning on the Ringbus and using P cores.

 

[FlameTail]

Six Five Media did a bunch of interviews with top Qualcomm executives during Computex 2024;

https://twitter.com/x/status/1798093558303637713

CEO Christiano Amon

https://twitter.com/x/status/1798479845808357577

Patel Rahul

https://x.com/TheSixFiveMedia/status/1798505044243042787

Alex Katouzian

https://x.com/TheSixFiveMedia/status/1798728949608145087

Kedar Kondap

https://x.com/TheSixFiveMedia/status/1798735318788776283

Nitin Kumar

https://x.com/TheSixFiveMedia/status/1798741626657103916

CMO of Qualcomm

They also have done interviews with people from the Laptop OEMs, and have more coverage of Qualcomm'a Computex 2024 stuff.

 

[SpudLobby]

It’s really not like Apple does. That lacks enumeration. The difference in degree is meaningful, and the core power/performances also explain why Apple doesn’t need to do this. Apple wants QoS 9 (background tasks) to exclusively go to E Cores, specifically because it’s a better use of energy without hurting UX much, and then they can free up P cores to be used for high QOS user interactivity. This is not synonymous whatsoever with “most real workloads” and the reason is obvious:

Apple’s P cores draw power not dissimilar from Skymont but with much more performance, and their E Cores don’t scale as high. Intel’s tradeoff is entirely different.

 

[FlameTail]

There was a rumour that X Elite G2 will use an overclocked Adreno 830.

FYI, Adreno 830 is the iGPU of the Snapdragon 8 Gen 4.

For context, here's how it compares:

3DMark Wildlife Extreme
M1 : 33 FPS
M2 : 40 FPS
M3 : 48 FPS
M4 : 55 FPS
X Elite : 44 FPS¹
X Plus : 38 FPS
8G2 : 22 FPS
8G3 : 32 FPS
8G4 : 44 FPS²

As you can see, the Adreno 830 is a very respectable iGPU for a smartphone. But the X Elite G2 will be a laptop SoC.

I'd say Adreno 830 with a reasonable 33% overclock, would do about 60 FPS. Any more than that, and the efficiency will go out of the window (we can see this already happening in some degree with the X Elite. The 4.6 TFLOPS GPU SKU consumes 50% more power than the 3.8 TFLOPS SKU)³

60 FPS would be barely better than M4, but it will be certainly worse than Apple M5. And besides, that 60 FPS would only be for the top SKU. And then we have the Intel/AMD competition. I an afraid an overclocked Adreno 830 won't be sufficient to be competitive.

The X Elite in apparently has an overclocked Adreno 740 (the GPU in the 8 Gen 2).

I don't know why Qualcomm can't do what Apple's doing (A15 Bionic -> M2, for example), by simply doubling the GPU cores. Qualcomm did this with the previous gen 8cx Gen 3, which had double the GPU of the Snapdragon 888.

So I hope the rumour about the overclocked Adreno 830 is not true. If it is true, then I hope Qualcomm has changed course. They should go ahead and spend the transistors. It's worth it.

____

¹This is for the X1E-84-100 SKU
²This is the leaked benchmark figure from Reve.
³According to Qualcomm's official GPU power curves

 

[SpudLobby]

Six Five Media did a bunch of interviews with top Qualcomm executives during Computex 2024;

https://twitter.com/x/status/1798093558303637713

CEO Christiano Amon

https://twitter.com/x/status/1798479845808357577

Patel Rahul

https://x.com/TheSixFiveMedia/status/1798505044243042787

Alex Katouzian

https://x.com/TheSixFiveMedia/status/1798728949608145087

Kedar Kondap

https://x.com/TheSixFiveMedia/status/1798735318788776283

Nitin Kumar

https://x.com/TheSixFiveMedia/status/1798741626657103916

CMO of Qualcomm

They also have done interviews with people from the Laptop OEMs, and have more coverage of Qualcomm'a Computex 2024 stuff.

Thanks for this! Awesome stuff.

 

[SpudLobby]

There was a rumour that X Elite G2 will use an overclocked Adreno 830.

FYI, Adreno 830 is the iGPU of the Snapdragon 8 Gen 4.

For context, here's how it compares:

3DMark Wildlife Extreme
M1 : 33 FPS
M2 : 40 FPS
M3 : 48 FPS
M4 : 55 FPS
X Elite : 44 FPS¹
X Plus : 38 FPS
8G2 : 22 FPS
8G3 : 32 FPS
8G4 : 44 FPS²

As you can see, the Adreno 830 is a very respectable iGPU for a smartphone. But the X Elite G2 will be a laptop SoC.

I'd say Adreno 830 with a reasonable 33% overclock, would do about 60 FPS. Any more than that, and the efficiency will go out of the window (we can see this already happening in some degree with the X Elite. The 4.6 TFLOPS GPU SKU consumes 50% more power than the 3.8 TFLOPS SKU)³

60 FPS would be barely better than M4, but it will be certainly worse than Apple M5. And besides, that 60 FPS would only be for the top SKU. And then we have the Intel/AMD competition. I an afraid an overclocked Adreno 830 won't be sufficient to be competitive.

Tbh I think it would be fine on graphics in isolation from drivers.


This again is too neurotic on these merits alone.

Apple’s stuff is useless for games. It’s not even in the ballpark. Even Qualcomm especially soon will blow them out.

Compute is different.

Here’s what matters from a pure engineering pov for Windows in the short term, putting software aside: can Qualcomm stay on par or ideally ahead of AMD/Intel on efficiency?

Intel’s LNL graphics on N3B achieve about M1 12W GPU-only power draw parity per their own documents and based on what they showed, that seems right. AMD’s RDNA3.5 is an improvement but Lunar should be better below 15W honestly and looks like a better overall setup (see XMX and stuff).

And AMD has such a gap with Qualcomm on power that even 3.5 won’t get them total parity I think.

So

The X Elite in apparently has an overclocked Adreno 740 (the GPU in the 8 Gen 2).

I don't know why Qualcomm can't do what Apple's doing (A15 Bionic -> M2, for example), by simply doubling the GPU cores. Qualcomm did this with the previous gen 8cx Gen 3, which had double the GPU of the Snapdragon 888.

See above. Area, their focus, and their competitors not being that good in laptops GPU efficiency already. Why would they go balls to the wall when they’re not mainly focused on gaming, there are early driver issues, and they can still compete with higher clocked phone parts even pending some porting? It doesn’t make literally any sense for Qualcomm to be spending out the roof on GPU area.

See again also above wrt where they already are vs Intel and AMD at least on paper wrt games. So the next Adreno would be more than fine.

Where Qualcomm has to worry and fix is drivers and getting games ported. Long term, the other thing is compute tor some local AI stuff or other applications, making the GPU good enough there is important and there is certainly a humongous gap between Adreno and Apple/Intel/AMD/Nvidia and probably Immortalis too. Less important for now but might be down the line.

So I hope the rumour about the overclocked Adreno 830 is not true. If it is true, then I hope Qualcomm has changed course. They should go ahead and spend the transistors. It's worth it.

____

¹This is for the X1E-84-100 SKU
²This is the leaked benchmark figure from Reve.
³According to Qualcomm's official GPU power curves

 

[FlameTail]

View attachment 100663


Half the power (has to be package or platform, which is good that they’ve improved there because the P Core didn’t get that much alone) with similar ST to Meteor Lake is a great result and it’s awesome to see Intel emphasizing stuff like this — SoC and packaging (power delivery) and ST perf/W which is very important, unlike AMD.

However this also, as I predicted, means they aren’t going to match an M3’s ST perf/W, which is a similar size on N3B. In fact this will just get them in line with the X Elite and maybe not even that depending on H and U skus and other things. But it does put them way ahead of Zen 4 and at at least a tolerable and sane level of performance/W.


View attachment 100664


As predicted, LNL is not really a threat on the merits alone, and Qualcomm probably has a cost advantage. With no emulation though and better GPU drivers, and the same design wins + more, LNL will be a good chip. But strategically here I’d rather be QC — they clearly have a better architecture at a core and SoC level.

Yes yes, very good.

Lunar Lake vs X Elite.

Qualcomm will have the win in CPU Performance, CPU Efficiency, NPU efficiency.

But when it comes to the GPU...

It's not looking good.

In fact I would go as far as to say it's weakest point of the X Elite.

● GPU drivers are evidently not polished enough
● GPU doesn't have Ray tracing, Mesh Shaving or Hardware Acceleration for SuperSampling.
● Adreno's GPU Compute performance is lacking, despite the graphics performance being good.
● X Elite's iGPU is better than Intel Meteor Lake and AMD Hawk Point in graphics benchmarks, but it's a mixed story in actual gaming.
● I estimate that Intel Lunar Lake iGPU and AMD Strix Point iGPU will surpass X Elite's iGPU in graphics benchmarks. You can imagine how it's going to be in actual games : |

Qualcomm needs to make big improvements to the GPU with X Elite G2. (See my above post)

 

[SpudLobby]

You’re way overthinking this I think!

Even in 2025/early 2026, anyone going to buy a Qualcomm laptop vs a Mac on the merits of the hardware nerd situation is 80% of the time going for Windows anyways or not comparing off that particualrly GPUs because of gaming. What would Apple’s GPU get them? Nothing, unless it’s a beefier laptop at which point QC doesn’t have a 256/512-bit caliber part anyway and they ain’t using that for gaming, they’re using that for Blender, AI, etc.

the cross section of people that care about GPU perf/W and perf and might compare an M chip to a Qualcomm chip for gaming (it won’t be compute lol), is effectively 0 people on planet earth.

And for Intel/AMD, Panther Lake is really a 2026 part with how Intel is, and AMD Zen 6 the same. Seeing how those two are doing currently on graphics perf/W, Qualcomm is fine. And in the near term, even with WoA improving, the most devout gamers will still go Nvidia/AMD/Intel before QC. More area won’t change that.

QC’s goal should be drivers drivers drivers and WoA porting.

Later on if software improves, they’ll need to look at spending more area, and also adding stuff like general matrix units and some better compute functionality.

 

[SpudLobby]

Now, if games worked very well with ports and even Intel-quality drivers as of today, then yes, something like how Apple’s M3 can match the M1 GPU TFLOPS at just half the power (so like 6W!) or gets like twice the perf at a bit more power would be appealing and I’d be annoyed QC cheaped out with the 740.

But that isn’t today and by Oryon V2 that will probably be here with the 830 or similar enough, which while is the same strategy it’s objectively a massive improvement and good enough once again.

 

[SpudLobby]

Yes yes, very good.

Lunar Lake vs X Elite.

Qualcomm will have the win in CPU Performance, CPU Efficiency, NPU efficiency.

But when it comes to the GPU...

It's not looking good.

In fact I would go as far as to say it's weakest point of the X Elite.

● GPU drivers are evidently not polished enough
● GPU doesn't have Ray tracing, Mesh Shaving or Hardware Acceleration for SuperSampling.
● Adreno's GPU Compute performance is lacking, despite the graphics performance being good.
● X Elite's iGPU is better than Intel Meteor Lake and AMD Hawk Point in graphics benchmarks, but it's a mixed story in actual gaming.
● I estimate that Intel Lunar Lake iGPU and AMD Strix Point iGPU will surpass X Elite's iGPU in graphics benchmarks. You can imagine how it's going to be in actual games : |

Qualcomm needs to make big improvements to the GPU with X Elite G2. (See my above post)

I agree the GPU is the weakest point, but mainly due to Arm gaming and mediocre Adreno drivers. As an aside Intel’s Xe2 is a complete package even down to like the encoding stuff or XMX units. Qualcomm’s is a very focused gaming GPU.

The perf/W itself is not going to differ that much at the 5-15W range, QC might even be ahead slightly.

But besides the gaming sw issues which is #1 — where QC blows is compute and stuff. Adreno just for gaming, whereas Xe2 looks almost like Apple’s GPU in some ways or like what Nvidia would have, and Intel will have real software to take advantage of it. Again, that’s also with Lunar Lake, a 128-bit part, so this is why it isn’t terrible for QC, if QC made a 256/512-bit part then the computer issue would really hurt, but I do agree with you.

 

[FlameTail]

Apple’s stuff is useless for games. It’s not even in the ballpark. Even Qualcomm especially soon will blow them out.

Compute is different.

Here’s what matters from a pure engineering pov for Windows in the short term, putting software aside: can Qualcomm stay on par or ideally ahead of AMD/Intel on efficiency?

What about the consumer pov?

Intel’s LNL graphics on N3B achieve about M1 12W GPU-only power draw parity per their own documents and based on what they showed, that seems right. AMD’s RDNA3.5 is an improvement but Lunar should be better below 15W honestly and looks like a better overall setup (see XMX and stuff).

And Qualcomm's 3.8 TFLOP does 20W, and 4.6 TFLOP does 30W.

And AMD has such a gap with Qualcomm on power that even 3.5 won’t get them total parity I think.

If you are going by Qualcomm's official curve...

Apparently the laptop with the Ryzen 7940HS has single-channel RAM. So it's performance (and thereby efficiency) is getting gimped. How else do you explain the fact that the 780M's performance/efficiency is so worse than Intel Arc Xe? 3rd party reviewers have found them to be pretty similar.

See above. Area, their focus, and their competitors not being that good in laptops GPU efficiency already. Why would they go balls to the wall when they’re not mainly focused on gaming, there are early driver issues, and they can still compete with higher clocked phone parts even pending some porting? It doesn’t make literally any sense for Qualcomm to be spending out the roof on GPU area.

Software is malleable. Hardware is not. You put good hardware today, and when the software improves it will take advantage of it.

You’re way overthinking this I think!

I certainly am, but I think I have good reasons for doing so.

Even in 2025/early 2026, anyone going to buy a Qualcomm laptop vs a Mac on the merits of the hardware nerd situation is 80% of the time going for Windows anyways or not comparing off that particualrly GPUs because of gaming. What would Apple’s GPU get them? Nothing, unless it’s a beefier laptop at which point QC doesn’t have a 256/512-bit caliber part anyway and they ain’t using that for gaming, they’re using that for Blender, AI, etc.

Yeah, it's mainly Qualcomm vs Intel/AMD from consumer POV. The comparison to Apple is for technical reasons.

And for Intel/AMD, Panther Lake is really a 2026 part with how Intel is, and AMD Zen 6 the same. Seeing how those two are doing currently on graphics perf/W, Qualcomm is fine. And in the near term, even with WoA improving, the most devout gamers will still go Nvidia/AMD/Intel before QC. More area won’t change that.

It's about being good enough for casual gamers.

But if QC CEO is serious about getting 50% marketshare in 5 years, this is not going to cut it.

QC’s goal should be drivers drivers drivers and WoA porting.

Later on if software improves, they’ll need to look at spending more area, and also adding stuff like general matrix units and some better compute functionality.

I have often been wondering if there is a need to add matrix units to the GPU itself? Why not do SuperSampling on the NPU?

 

[FlameTail]

I agree the GPU is the weakest point, but mainly due to Arm gaming and mediocre Adreno drivers. As an aside Intel’s Xe2 is a complete package even down to like the encoding stuff or XMX units. Qualcomm’s is a very focused gaming GPU.

The perf/W itself is not going to differ that much at the 5-15W range, QC might even be ahead slightly.

But besides the gaming sw issues which is #1 — where QC blows is compute and stuff. Adreno just for gaming, whereas Xe2 looks almost like Apple’s GPU in some ways or like what Nvidia would have, and Intel will have real software to take advantage of it. Again, that’s also with Lunar Lake, a 128-bit part, so this is why it isn’t terrible for QC, if QC made a 256/512-bit part then the computer issue would really hurt, but I do agree with you.

So we both agree on common terms.🤝🤝

 

[FlameTail]

Meanwhile Microsoft is shooting themselves in the foot, and Qualcomm with it...

A PR disaster: Microsoft has lost trust with its users, and Windows Recall is the straw that broke the camel's back

The world is up-in-arms over Windows Recall, but why? It stems from Microsoft's seeming lack of care for Windows and its users.

www.windowscentral.com

 

[FlameTail]

Then there is the other question of whether Qualcomm should make a Strix Halo/Mx Pro like part with 256 bit memory bus.

Is it worth it? Making such a large die will certainly be an expensive endeavour.

(1) In terms of marketshare/profit, probably not. The volume of such parts are low, and even among the people who'd look to buy such a 256-bit part, they'd probably choose Apple or AMD's option instead of Qualcomm.

(2) In terms of mindshare, their is certainly a benefit. It will show that Qualcomm'a got guts. It's the same justification as to why so many people talk a lot about the RTX 4090, although they'll probably have something like a 4060. It's because such Halo products act as the flagship of the brand.

 

[coercitiv]

Meanwhile Microsoft is shooting themselves in the foot, and Qualcomm with it...

Microsoft has made it a habit to force things on their users. This erodes consumer trust and sooner or later the brand will pay in some way or another.

Recall however, is more than a problem of brand trust. They screwed it up big time, the feature is VERY insecure and would make the work of malicious actors so much easier. Not only would Recall data be a boon for quickly finding sensitive information about the user, it might be used to find out info that could never be found otherwise (because there might not be other records left on that particular machine).

Then there is the other question of whether Qualcomm should make a Strix Halo/Mx Pro like part with 256 bit memory bus.

Is it worth it? Making such a large die will certainly be an expensive endeavour.

If QC is smart, they will go after the professionals. Keep the chips lean, offer strong performance and battery life, make sure drivers and apps are top tier, maybe even add an additional layer of software features aimed at professionals and small business owners. You know what was the crappiest part of my Ideapad 5 Pro x86 laptop I bought two years ago? The Qualcomm Atheros WiFi card. During device sleep it would occasionally disappear and stay that way until a system reboot. Had to replace it with an Intel card, never had a problem since.

After they get their beachhead in the form of satisfied users in one segment, go wild with some more powerful stuff. A strong and versatile GPU + wide memory bus is probably the future in terms of AI features as well.

 

[FlameTail]

Microsoft AutoSR detailed:

Automatic Super Resolution: The First OS-Integrated AI-Based Super Resolution for Gaming - DirectX Developer Blog

Auto SR automatically enhances select existing games allowing players to effortlessly enjoy stunning visuals and fast framerates on CoPilot+ PCs equipped with a Snapdragon® X processor. DirectSR focuses on next generation games and developers, Together, they create a comprehensive SR solution...

devblogs.microsoft.com

It ships with Copilot+ PCs powered by Snapdragon X.

Auto SR stands out in two exciting ways: it’s applied automatically and enhances existing games. Like other SR technologies, Auto SR uses AI to improve frame rates and image quality. However, Auto SR focuses on bringing the benefits of super resolution to your existing game library with no manual configurations needed. This makes it ideal for players who prefer a straightforward experience. Simply start your game, and Auto SR instantly enhances it, allowing you to effortlessly enjoy visuals that surpass native 1080p quality with the fast frame rates typically seen at lower resolutions. Auto SR boosts detail and performance on compatible hardware, transforming your gameplay and letting you experience select titles in a new light.

At the heart of Auto SR lies a sophisticated AI model—a convolutional neural network (CNN) expertly trained on gaming content. Its mission: to deliver visuals of exceptional quality, surpassing even native 1080p from visuals rendered as low as 700 vertical lines. This AI is tailored to run on a Neural Processing Unit (NPU), which is hardware built to efficiently process the large number of arithmetic operations and memory transfers required by CNNs, like extensive multiply-adds. This setup allows us to use larger AI-based SR models to make big visual enhancements.

It uses the NPU!!!

 

[FlameTail]

Microsoft has made it a habit to force things on their users. This erodes consumer trust and sooner or later the brand will pay in some way or another.

Recall however, is more than a problem of brand trust. They screwed it up big time, the feature is VERY insecure and would make the work of malicious actors so much easier. Not only would Recall data be a boon for quickly finding sensitive information about the user, it might be used to find out info that could never be found otherwise (because there might not be other records left on that particular machine).

Good assessment. Their is talk that MS is considering making some changes to Recall in a software update, before devices launch on June 18th.

If QC is smart, they will go after the professionals. Keep the chips lean, offer strong performance and battery life, make sure drivers and apps are top tier, maybe even add an additional layer of software features aimed at professionals and small business owners. You know what was the crappiest part of my Ideapad 5 Pro x86 laptop I bought two years ago? The Qualcomm Atheros WiFi card. During device sleep it would occasionally disappear and stay that way until a system reboot. Had to replace it with an Intel card, never had a problem since.

What's the difference between Qualcomm Atheros and Qualcomm Fastconnect? (Snapdragon X Elite comes bundled with the latter).

 

[The Hardcard]

i’ve seen eclectic’s work and this is not actually true. For high priority user interface stuff, the threads run on P Cores and if none are available they go to E Cores.

They are mainstream use for a broad amount of tasks but “most real workloads use the E Cores” is absolutely not how Apple Silicon works. The E cores also did reach like 2 on SpecInt and now like 4.0 plus with M3.

Intel’s LP E Cores with MTL were doing 3.3 or so at most with SpecInt. At a 1.7X uplift they’ll be at around 5.61 at the same cluster power (and at a package level they were around 5-6W for this with everything else shut off in MTL, so that’s interesting). At 2x and more power, they’ll be at 6.6.


Or for the same 3-3.3 SpecInt the cluster draws 1/3 the power in MTL. Which isn’t bad if that’s moving from 5-6W to 1.5-2W, but Apple’s and Arm’s from a whole platform power pov are doing more than that in Arm’s case around the same power with last generation IP and on N4, or they’re doing it at vastly less power (at least in the 2GHz A17 example.)

intel’s Skymont in LNL based off what we know here looks like the closest Intel has been towards a phone A7x core or a big core from Apple/Arm, but what you’re missing is Apple can draw the kind of power Skymont will with their P Cores and get more performance, and their E Cores don’t scale up as high for performance or certainly didn’t with M1/2. So it very much is a different strategy because they have a different set of cards to play.


Sources here on power/perf:

View attachment 100670
View attachment 100671
(And package here is understating things as it doesn’t include RAM or power delivery, but the Arm examples above do and are measured physically).

View attachment 100672
(A715 is last-gen and on N4 ehre, but even when measured including the whole platform it does phenomenally by contrast to Crestmont and I am skeptical SKT will close that gap even at 1/3 the power — math doesn’t work there, much less Apple’s.)

View attachment 100673


Based on MTL’s LP E core perf/w which is with Compute Tile off and just the Island, this will put them much closer to Arm cores but even that “1.7x” or “2x” at even more power demonstrates exactly why Apple and others can use their p cores more often.

When Intel says they’re using the E Cores, that cluster is consuming way more power with also up to 50-60 more peak perf than an A7x or Apple E core would.

Which also makes the strategy and difference from Apple or QC much more believable — those two can just use their P cores and get more perf at the same power. But for Intel, using Skymont even at 5W probably makes sense for good enough ST without hurting power from turning on the Ringbus and using P cores.

Maybe, it depends on the software. I don’t currently have an easy way to run my M1 Max, but during my use, my experience matched his across most of the tasks I did. That is E cores loading first. Very rarely were my P cores firing up without the E cores being mostly loaded.

 

[DavidC1]

Intel’s LP E Cores with MTL were doing 3.3 or so at most with SpecInt. At a 1.7X uplift they’ll be at around 5.61 at the same cluster power (and at a package level they were around 5-6W for this with everything else shut off in MTL, so that’s interesting). At 2x and more power, they’ll be at 6.6.

You have to compare it core-level not package level as I suspect Meteorlake and previous generations have horrible idle power management figures.

So you might have a 1W core, but 3W uncore, getting it to 4W. Remember, if they can get that 3W down to 0.5W, it'll change the equation, even if it's still behind.

We have to see how Lunarlake does. It might be 1.7x the performance uplift over MTL's LP-E core while cutting package power in real world situations drastically. I mean in one video playback the board-level power drops by nearly half, from 9W to 5.5W.

Also, even if load-power isn't a leadership, you can still get better battery life in bursty situations such as web browsing with Lunar Lake. So wait for actual tests. They are claiming leadership with Teams for example.

 

[SpudLobby]

You have to compare it core-level not package level as I suspect Meteorlake and previous generations have horrible idle power management figures.

So you might have a 1W core, but 3W uncore, getting it to 4W. Remember, if they can get that 3W down to 0.5W, it'll change the equation, even if it's still behind.

True. With the package here, I assumed compute tile was off for the LP E cores, and the RAM is not on package with Meteor Lake, nor is it measuring the VRM losses or anything. I suspect the package and PD will still put them behind even if that is true, and the Arm figures and Apple figures do include all of that for them minus static.



Though FWIW it says “E Core Cluster” so I think they’re including the power of the cluster, just not the rest of the uncore, but you are right here, if the gains on package and other uncore are even larger as a constant thing, then it really will change the equation even if they are behind.

We have to see how Lunarlake does. It might be 1.7x the performance uplift over MTL's LP-E core while cutting package power in real world situations drastically. I mean in one video playback the board-level power drops by nearly half, from 9W to 5.5W.

I fully believe they will have similar idle to Qualcomm and similar battery life for a lot of mixed use, albeit QC doing so with a part that has 12C and the RAM isn’t on package. The Meteor Lake to QC video battery life from Dell’s XPS seem to indicate something like this: the exact same laptop goes from 18 to 27H on video. Intel claims like 60% video battery over Meteor Lake which also roughly matches QC.

What will be most interesting is mixed automated tests. Like what does web browsing look like at 150 nits for similar laptops? How does responsiveness feel on like balanced windows power plans?

Either way it’s a huge step forward just self evidently, and AMD I don’t think is going to do well next gen.

 

[DavidC1]

True. With the package here, I assumed compute tile was off for the LP E cores, and the RAM is not on package with Meteor Lake, nor is it measuring the VRM losses or anything. I suspect the package and PD will still put them behind even if that is true, and the Arm figures and Apple figures do include all of that for them minus static.

Though theoretically Meteorlake should have been better, I highly suspect they missed their targets with massive political upheaval and delays to 7nm process. Thus it wouldn't have worked as well as they'd like, hence the wildly varying battery life and even performance figures. I doubt they are turning off tiles even 25% often as they planned.

I know for a fact they are behind, because it has been seen previously with people in denial that it couldn't be improved. I knew it could be improved, because it was Intel's reluctance to actually address the system power that battery life sucked pre-Haswell laptops. The battery life figure did not matter, whether it was a 3W Atom nor a 45W -H chip.

Then Haswell proved them wrong, because the package power dropped to less than half. It went from 2.5W minimum package power to 0.7W. I ignore pre-Haswell laptops when buying used for this reason - for me they are nonexistent.

While Haswell did improve it, and Broadwell did it further(total = 2x), not only they had regressed starting with Icelake, even with Broadwell they were significantly behind ARM parts.

The battery life on Haswell was better even on load, because it affects all scenarios, while reducing idle power by X amount results in X being reduced from all workloads too, including load.

Now with Lunarlake and beyond, they need to address load power too, along with further improving system and SoC design all the while improving the cores.

 

[SpudLobby]

Honestly, the P core gains from Intel 4 to N3 look reasonable, and some of this is architectural advancements (see: new EDA tools and an IPC gain), too. So our upper bound for how much better N3 is than Intel 4 from a pure performance iso-power perpsective is like 10-5%, the 18% looks larger though that’s really at the bottom and I suspect also about some design changes as much as it might be TSMC low power.




Which is like a standard node gain from a pure power/performance POV — I’d say like 12-15% more frequency iso-power, 24-30% power iso-frequency for the same core from I4 to N3B.

Significant? Absolutely. But that’s actually not that bad. Would imply that Intel 4 really is N5 or N5P-vanilla caliber on power.

Makes you wonder what Intel 3 with HD libraries or other enhancements will be like, since supposedly they get another 10% perf/W from Intel 3 standard alone, and also makes Intel’s move to 18A with Panther Lake make more sense. It [18A] probably is legitimately better than the N3 family.

 

[StinkyPinky]

I think this is a good practical snapdragon laptop.

https://www.bhphotovideo.com/c/product/1826504-REG/lenovo_21n10000us_thinkpad_g6_qualcomm_32gb_1tb_w11p_14.html

Pros: 32GB ram, 2 x thunderbolt 4 ports, 2 x USB A 3,2 ports, HDMI 2.1, headphone jack, wifi 7, Matte display
Cons: IPS display, 58W battery, slowest of the Elite soc's

 

[DavidC1]

True. With the package here, I assumed compute tile was off for the LP E cores, and the RAM is not on package with Meteor Lake, nor is it measuring the VRM losses or anything. I suspect the package and PD will still put them behind even if that is true, and the Arm figures and Apple figures do include all of that for them minus static.

https://www.notebookcheck.net/MSI-Prestige-16-B1MG-laptop-review-From-Core-i7-Xe-to-Core-Ultra-7-Arc.785587.0.html#10648157

CPU package power: 6.5W
IA Cores power: 2.4W
System Agent Power: 1.9W

So that means we have 2.2W unaccounted for. And SA is also uncore and that's something they can improve. I believe package power is actually lot lower than 6.5W and they didn't close enough background applications, but still. That means in this case the CPU power takes up only little over 1/3rd. Cutting 2.4W to 0.8W is not too significant.

Makes you wonder what Intel 3 with HD libraries or other enhancements will be like, since supposedly they get another 10% perf/W from Intel 3 standard alone, and also makes Intel’s move to 18A with Panther Lake make more sense. It [18A] probably is legitimately better than the N3 family.

You got it confused with 18A. 18A is 10% over 20A. Intel 3 is 18% over Intel 4, meaning based on your calculations it'll be on par with N3B. 20A is further 15% better, and 18A is 10% on top of that.