Discussion Qualcomm Snapdragon Thread

[FlameTail]

I wonder...Next year we'll see Snap 8G4 with the shiny Phoenix Cores + Adreno 8xx.

This year. October. Snapdragon Summit 2024.

Or will QCOM actually divest 7+ and 8s from the 8/8+ and keep using stock Arm designs? Would be a interesting way to keep the premium and the premium mid-range lines separated

Yes, it will be interesting to see how Qualcomm brings Oryon to their midrange chips.

Unlike Apple, who doesn't make any "midrange" chips. They reuse older flagship SoCs.

 

[Ghostsonplanets]

This year. October. Snapdragon Summit 2024.

I actually meant availability. But I remembered that a lot of Chinese and Motorola flagship release in the latter half of the year. My brain was too focused on Galaxy😂

Yes, it will be interesting to see how Qualcomm brings Oryon to their midrange chips.

Pretty much. In the long term, it's in their best interest to stop paying for each Arm design and save costs by using theirs own. The question is how agressive they'll be.

Stuff like Snap 6 and Snap 4 will take 2 - 3 years to see a QCOM custom core. And QCOM will probably reuse an old design (Like Snap 6 Gen 5 in 2026 using Phoenix rather than Pegasus). But Snap 7 is where all the eyes will be. Specially given QCOM has 3 premium midrange lines (7G, 7G+ and 8s).

 

[uzzi38]

Why do you keep increasing this? Kepler_L2 said 40%.

Nobody stops the hype train I guess?

 

[hemedans]

7s Gen 3?
My head is starting to spin with these confusing names.

There is 7s gen 2 already fabbed by Samsung.

 

[SpudLobby]

I actually meant availability. But I remembered that a lot of Chinese and Motorola flagship release in the latter half of the year. My brain was too focused on Galaxy😂

Pretty much. In the long term, it's in their best interest to stop paying for each Arm design and save costs by using theirs own. The question is how agressive they'll be.

Yeah.

 

[SpudLobby]

Well, if the rumours of AMD achieving +50% ST with Zen 5 us true, then Qualcomm is in trouble....

What they get on desktops may not be what they get on laptops depending on the frequency and what they’re willing to push, but to be fair they have room and N4P will help for 5% more perf iso-power on paper.

I think 40% is what’s been quoted btw.

Keep in mind the 7940HS at best in the most extreme gaming or beefy profiles looks like about a 2600-2700 GB6, at 5.2GHz, and almost 22-23W platform, and Qualcomm still has ~ 10% perf on them at peak, obviously at a lower power level (14W).

I will bet you that with a frequency bump and some crazy 30% IPC improvement, they’ll consume more power than they already do at peak performance.

IPC improvements in wider cores can push power down for the same performance, this is true, but at the same frequency with a similar node, it very well might be higher.




At 9W, QC has about a 30% more performance gain than AMD’s 7940HS on ST. At 14W, it looks smaller, more like 20%, but still there. QC can scale down more than AMD too and while they don’t have e-cores yet I would bet Oryon is still more efficient than Zen 5C for background tasks.

Does AMD close or substantially narrow this? To me that’s the most interesting, that and battery life. I don’t care if they get 30-40% more ST performance and at the top end they’re using another 5-10W on mobile, I want to see the curves.

I feel the same with LNL.

 

[FlameTail]

I think 40% is what’s been quoted btw.

Keep in mind the 7940HS at best in the most extreme gaming or beefy profiles looks like about a 2600-2700 GB6

So
2700 × 140% = 3800

Does AMD close or substantially narrow this? To me that’s the most interesting, that and battery life. I don’t care if they get 30-40% more ST performance

Well, you don't care, but many others will. For Qualcomm, it is essential they keep up with the industry's leading edge in ST performance.

X Elite is able to keep up with M3, Hawk Point/Phoenix, and Meteor Lake.

If X Elite G2 launches in late 2025, or early 2026, it's competition is going to be Zen6 and Apple M5.

 

[SpudLobby]

So
2700 × 140% = 3800

That is wild and impressive, and for desktops I imagine Zen 5 is going to be a killer deal even when it’s expensive. Would much rather have 6-8 Zen 5 cores than a mess of Intel’s, even with ARL, which probably would still consume too much power and suck for some stuff.

But the difference between that 3800 and say, when the M1 came out and had an ST lead, is that the power consumption for that is going to be ridiculous for mobile, I will guarantee it’ll be higher than the score at 2700 which is already like 22W. More IPC will often reduce power iso-performance (AKA lower clocks) but iso-frequency, and without major node gains, major cache increases (Zen 5 will have a slightly larger L1 but nothing crazy like Apple or QC, and a bigger L3 — for the mobile stuff (16 -> 32), etc, it’ll be a steamer at that peak.

Again though I expect they’ll improve nontrivially from the 5-15W range in ST perf, would be shocked if they don’t. Whether or not they’ll pull their floor lower is another question though.

Well, you don't care, but many others will. For Qualcomm, it is essential they keep up with the industry's leading edge in ST performance.

Keep up with, yes, take the crown, no. Even Apple is in a game of leapfrog when it comes to that and Intel/AMD have been nipping at their heels if not occasionally surpassing them for mobile parts.

Let me put it this way: LNL despite high clocks and power also won’t be anywhere near Zen 5, fwiw, they’re going to be matching the M3 at likely 2-3x++ the power if the 4.9GHz and 20% IPC pans out.

QC, Apple, and Intel will all be around the 2900-3200 range it seems, with Apple’s best entries at the top towards 3200 GB6 and at less power Intel and to a lesser extent QC. Qualcomm’s top SKU around 2900+, Intel’s top LNL SKU probably in the middle of QC/Apple, again at the most power though I think.



Regardless you see QC isn’t unique here. Apple, QC, and Intel’s real mobile parts would in this case cluster in the same group on peak ST performance with a deficit to AMD. And Intel’s even using more watts on a better node than QC, and has less MT oomph.


****ARL is garbage and has the same issues MTL has and so I won’t even mention that, LNL is what actually matters for volume and competing with Qualcomm and arguably Apple, and eventually even AMD’s Zen 5 Kraken 4+4 stuff.

X Elite is able to keep up with M3, Hawk Point/Phoenix, and Meteor Lake.

If X Elite G2 launches in late 2025, or early 2026, it's competition is going to be Zen6 and Apple M5.

 

[SpudLobby]

On the X Gen 2, if it launches in Q1 2026 I think that’s not ideal given the delay we had with Gen 1 but it’s not the end of the world.

One thing to remember is AMD’s Kraken the 4+4 mainstream Zen 5 mobile part for example doesn’t launch until 2025, and Lunar Lake in practice will be volume 2025 too — it’s been moved up in place of ARL on priority (per the rumors which makes sense) but it won’t be drastically different from MTL I don’t think.

And so with those two I think QC will be mostly fine. It’s also likely QC will have a significant IPC tweak for Pegasus. If they get 10-15%, and 10% off the node, and keep power the same or even reduce it with more cache and new phydes, that’s great.

But you have to realize they’re not going to do a fat core that reaches insane frequencies that require they bloat the design too much. They’ll take the gains they can from newer nodes on similar libraries and transistors without changing their uArch drastically.

Like, a + 5-10% clocks from N4P to N3E? Yeah I could see 4.6-4.7GHz. But you get the idea here.

Also remember from where they are, adding a “mere” +13% IPC and + 10% frequency would take their 2950 (about the best they can get on Windows or so) to 3,666. That’s fine. Doubt the M4 is going to run past that and even M5 may not be that far ahead if they’re still on N3 which they will be if they’re going annually.

 

[SpudLobby]

Honestly I feel like people are being hyperbolic, though I do agree the part is late and QC will need some price competition for the X Plus, or that they need probably a 20% ST uplift for 2026, sure.

But like we don’t even know when LNL’s successor is coming, M4 is on N3E and M5 probably just N5P if it even arrives in late 2025, neither of which give Apple a crazy amount of headroom for frequencies while keeping power down. Minimal logic gains and they will have made about 3% annual IPC gains since 2020 by 2024 November, at least pre-M4. A large denominator to start from and still leading on perf/W, but still.

If M3 -> M5 with N3P got them 5% higher frequencies while keeping power constant, they’d be right at about 4.25GHz, and if it got them 5% more IPC one year, their GB6 ST would be like 3500-3550.

Separately — it is not impossible that, because N3B yields could impact things, maybe N3E and N3P offer more like + 5 and then + 10% frequency, but even then it’s like +10%. And when you think about Apple’s target market — they can’t keep raising power too much at this point and so they’re really going to need some IPC upgrades. If M4 and M5 are on an annual cadence since M3, and looking at historical improvements since they lost Nuvia, why would we expect like + 20-30% cumulative IPC from 2023 M3 to 2025 M5? People get so excited for that but it’s just crazy. Even the A12 to A14 was just 14-16% cumulative and those were huge years.

 

[uzzi38]

On the X Gen 2, if it launches in Q1 2026 I think that’s not ideal given the delay we had with Gen 1 but it’s not the end of the world.

One thing to remember is AMD’s Kraken the 4+4 mainstream Zen 5 mobile part for example doesn’t launch until 2025, and Lunar Lake in practice will be volume 2025 too — it’s been moved up in place of ARL on priority (per the rumors which makes sense) but it won’t be drastically different from MTL I don’t think.

And so with those two I think QC will be mostly fine. It’s also likely QC will have a significant IPC tweak for Pegasus. If they get 10-15%, and 10% off the node, and keep power the same or even reduce it with more cache and new phydes, that’s great.

But you have to realize they’re not going to do a fat core that reaches insane frequencies that require they bloat the design too much. They’ll take the gains they can from newer nodes on similar libraries and transistors without changing their uArch drastically.

Like, a + 5-10% clocks from N4P to N3E? Yeah I could see 4.6-4.7GHz. But you get the idea here.

Also remember from where they are, adding a “mere” +13% IPC and + 10% frequency would take their 2950 (about the best they can get on Windows or so) to 3,666. That’s fine. Doubt the M4 is going to run past that and even M5 may not be that far ahead if they’re still on N3 which they will be if they’re going annually.

MTL actually launched late Q4 as well, iirc mid-December.

Lunar Lake isn't that, it's launching in Q3, similar timeframe to X Elite. You're just better off assuming both parts are the same timeframe for the most part. Also Kraken is the mainstream part sure, but both LNL and X Elite are more akin to Strix in thatr they're "premium" ultrathin SoCs.

I'm not going to comment on performance numbers though for now. Just going to say lets see where all of these first parts land later this year first, then we can worry about next generation products after that.

 

[SpudLobby]

MTL actually launched late Q4 as well, iirc mid-December.

I know it launched Q4 fwiw, that’s what I was expecting for LNL. Are you saying it’s more like Q3 now?

Interesting.

Lunar Lake isn't that, it's launching in Q3, similar timeframe to X Elite.

Yeah I see.

You're just better off assuming both parts are the same timeframe for the most part. Also Kraken is the mainstream part sure, but both LNL and X Elite are more akin to Strix in thatr they're "premium" ultrathin SoCs.

That’s true, but X Plus is more mainstream part from QC, and will still share the efficiency characteristics of the Elite (minus two cores and minus some clocks), which is why I brought this up.

I'm not going to comment on performance numbers though for now. Just going to say lets see where all of these first parts land later this year first, then we can worry about next generation products after that.

That’s how I feel as well but speculation is about half of what we do here and I felt like addressing Flame. That said, I basically agree. LNL is what interests me vs X Elite and X Plus, Strix I am sure will be great though is also a bit later afaict.

 

[igor_kavinski]

****ARL is garbage and has the same issues MTL has

Oh no. They seem to be going all in with that garbage



Guess WoA will have ample chance to gain tons of marketshare, eh?

 

[FlameTail]

QC, Apple, and Intel will all be around the 2900-3200 range it seems, with Apple’s best entries at the top towards 3200 GB6 and at less power Intel and to a lesser extent QC. Qualcomm’s top SKU around 2900+, Intel’s top LNL SKU probably in the middle of QC/Apple, again at the most power though I think.

Regardless you see QC isn’t unique here. Apple, QC, and Intel’s real mobile parts would in this case cluster in the same group on peak ST performance with a deficit to AMD. And Intel’s even using more watts on a better node than QC, and has less MT ooph

Are you forgetting that the Apple M4 lineup is said to be releasing Q4 of this year (According to Gurman). That's ~5 months after X Elite releases.

I estimate that M4 will bring atleast ~15% ST increase.

M3 does 3100-3200
Therefore,
M4 = 3100 × 115% = 3565.

So Apple M4 will be on a good footing vs AMD Strix Point. You know what? When the Gurman report said that M4 is coming in Q4 2024, I was like: "Oh wow, it's arriving just at the right time to fight off Lunar Lake / Strix Point".

Even if Strix is announced in June 2024 at Computex, we all know that it will take -~6 months for real availability.

 

[FlameTail]

+ 20-30% cumulative IPC from 2023 M3 to 2025 M5? People get so excited for that but it’s just crazy. Even the A12 to A14 was just 14-16% cumulative and those were huge years.

A12 -> A13 is a substantial 22% PPC increase. You can see the that in the preceding years too they have got huge uplifts. It's only with the A14 and after, that the PPC gains slowed down drastically.

____

You seem to be skeptical that Nuvia engineers can't pull off a 40-50% IPC increase in the span of 2 years. (If X Elite G2 with Pegasus core is announced in 2025Q4, that's 2 years after the announcement of X Elite with Phoenix core in 2023Q4).

I believe otherwise. I think Nuvia can do the iimpossible. As Gerard Williams stated, they are not breaking physics, but they are bending it.

 

[poke01]

It's only with the A14 and after, that the PPC gains slowed down drastically.

The question we should we asking is who is Apple's current CPU architect?

 

[SpudLobby]

Are you forgetting that the Apple M4 lineup is said to be releasing Q4 of this year (According to Gurman). That's ~5 months after X Elite releases.

I estimate that M4 will bring atleast ~15% ST increase.

Doubt this honestly. They pushed power up just a tad with the M3 from M2, though efficiency is fine, and we saw their new core which was functionally like an Everest +++++.

I am really skeptical Apple will have “at least” a 15% upgrade in part because clock gains from N3B to N3E are not going to be huge, and I doubt they have a 5-10% IPC upgrade in store.


They could say screw it and go balls to the wall with clocks maybe but they won’t do that.

I will own it if I’m wrong. But I bet I’m right and people are going to be pretty disappointed again.

 

[SpudLobby]

Oh no. They seem to be going all in with that garbage

View attachment 96988

Guess WoA will have ample chance to gain tons of marketshare, eh?

Yep it’s crap, wasting N3 and 20A on an insane SoC architecture.

LNL is at least good relative to Intel stuff and something I’d probably consider buying

 

[FlameTail]

Oh no. They seem to be going all in with that garbage

What's wrong with Arrow Lake?

Guess WoA will have ample chance to gain tons of marketshare, eh?

Qualcomm, AMD, ARM, Nvidia all dining over Intel's dead body.

 

[SpudLobby]

What's wrong with Arrow Lake?

Did you see the leaks a while back?

It’s MTL, with a new node and core architectures but not enough to push the envelope substantially over ADL, ST apparently barely gains in some stuff — power is also ridiculous.

It’s just MTL+++, which might be good to some people but honestly it’s a waste of high end sand to me. Makes me almost sad to think about the utter waste lol

Like LNL is pathetic in some sense, but also at least will be good for battery life and efficiency relative to ARL and not a mess. An Intel part worth buying

Qualcomm, AMD, ARM, Nvidia all dining over Intel's dead body.

 

[DrMrLordX]

Did you see the leaks a while back?

It’s MTL, with a new node and core architectures but not enough to push the envelope substantially over ADL, ST apparently barely gains in some stuff — power is also ridiculous.

The oldest leak I remember was a leaked slide showing Arrow Lake-S @ PL2 = ~250W being maybe 5% faster than Raptor Lake in ST and uh 20% faster (I think?) in MT (same PL2 for Raptor Lake). That was for 8p+16e parts. Based on what you're saying now, that slide could be accurate.

 

[SpudLobby]

The oldest leak I remember was a leaked slide showing Arrow Lake-S @ PL2 = ~250W being maybe 5% faster than Raptor Lake in ST and uh 20% faster (I think?) in MT (same PL2 for Raptor Lake). That was for 8p+16e parts. Based on what you're saying now, that slide could be accurate.

That’s what I’m basing it off of, to be fair, that and other reports. Lot of gossip but it makes sense to me based on a few different things.

 

[FlameTail]

M3 is more than 10% faster (148 vs 133) than X Elite in Cinebench 2024 ST, a gap that is larger than the one in Geekbench 6 (3200 vs 3000).

Seems like CB2024 really likes the extra cache in the M3.

 

[Tigerick]

Snapdragon X2 series Speculation/Concept Version 2:

View attachment 96753
After taking into consideration your feedback that Qualcomm is unlikely to make an SoC with a >128 bit memory bus, I eliminated those SKUs. Now we have 2 dies and 4 SKUs.

Other changes include use of LPDDR6 instead of LPDDR5T. @Tigerick Happy?

I assumed LPDDR6 would double it's density vs LPDDR5, hence why X2 Ultimate can have 128 GB RAM in only a 128 bit bus.

Release date has been adjusted to 2026Q1.

The P-core now has +50% IPC over that of X Elite. I believe it is essential to compete with Zen6 and Apple M5.

What do you guys think of this new concept? More reasonable/realistic?

Why should you care about my happiness if you are wrong in the beginning?

If you don't believe what we said, then DON'T change your speculation...

If you do, at least shown some appreciation in our reply.

And I have been saying flagship phone SoC going to be used as WoA 64-bit ARM SoC. That's mean future 8G4 and 8G5 will be used as X G1 and X G2. That's mean they should use 2+6 CPU configuration. Thus your 4+4 config is incorrect. And I remember what you said:

Of course that's always an open possibilty. But I am skeptical about it, especially with regards to Qualcomm doing so.

You see, these flagship 64-bit SoCs come integrated with cutting edge 5G modems. They take up about 15-20 mm² of die area.

Not all brands want to put 5G in their laptops. So the added cost of the 5g modem but it's redundancy when not used, is a waste of money. Hence why all Snapdragon 8cx chips and now the Snapdragon X Elite don't have integrated modems.

So you don't believe Qualcomm will use 8G4 and 8G4 as X G1 and X G2. I will let you think first before saying anything...

Just like you said about 8 core: "Have you heard of the Snapdragon X Plus?" I will reply to you promptly in the future. Stay tune

 

[Tigerick]

View attachment 96990
A12 -> A13 is a substantial 22% PPC increase. You can see the that in the preceding years too they have got huge uplifts. It's only with the A14 and after, that the PPC gains slowed down drastically.

____

You seem to be skeptical that Nuvia engineers can't pull off a 40-50% IPC increase in the span of 2 years. (If X Elite G2 with Pegasus core is announced in 2025Q4, that's 2 years after the announcement of X Elite with Phoenix core in 2023Q4).
View attachment 96991
I believe otherwise. I think Nuvia can do the iimpossible. As Gerard Williams stated, they are not breaking physics, but they are bending it.

No way, in case you forgotten about the table I created to calculate PPC which include X Elite here. Phoenix core has about 700 PPC. If improvement over 40%, that's mean PPC would be 980. Do you really think Nuvia team able to leapfrog Cortex-X5 so much which already shown pretty impressive results.

Based on my estimate, if Pegasus core manage to show 20% PPC improvement, ie. 840, then should be pretty good already.