Discussion Qualcomm Snapdragon Thread

[FlameTail]

Geekerwan still hasn't released their review of the X Elite (for reasons). They already have review videos out for Strix Point and Ryzen 9700X- both of which were released weeks after the X Elite!

Anyways, here's a wishlist of things I want to see in the Geekerwan review.

• SPEC2017 Single Thread Power Curves

• Cinebench 2024 Multi Core Power Curves

Would be interesting to see how it stacks up against the M3 Max and Ryzen AI HX 370. Toss in an Intel Meteor Lake chip too.

• Running Cinebench 2024 Multi Core using liquid nitrogen

Qualcomm showed off their 80W reference scoring 1200 points in CB2024 MT. However, I have a suspicion that this doesnt actually reach the 3.8 GHz all core clock, and the X Elite can score even higher, if provided with more power and lifting the thermal ceiling.

• GPU Power Curves in 3DMark Wildlife Extreme and Steel Nomad Light

How good is the efficiency of rhe 1.5 GHz Adreno?

 

[FlameTail]

[Double post; Delete]

 

[mvprod123]

Geekerwan досі не опублікував свій огляд X Elite (з причин). У них уже є оглядові відео для Strix Point і Ryzen 9700X, обидва з яких були випущені через кілька тижнів після X Elite!

У будь-якому випадку, ось список бажань речей, які я хочу побачити в огляді Geekerwan.

• SPEC2017 Однопотокові криві потужності

• Багатоядерні криві потужності Cinebench 2024

Було б цікаво подивитися, як він протистоїть M3 Max і Ryzen AI HX 370. Також додайте чіп Intel Meteor Lake.

• Запуск Cinebench 2024 Multi Core з використанням рідкого азоту

Qualcomm продемонструвала свою еталонну потужність 80 Вт, набравши 1200 балів у CB2024 MT. Однак у мене є підозра, що це насправді не досягає тактової частоти всіх ядер у 3,8 ГГц, і X Elite може отримати навіть вищу оцінку, якщо забезпечити більше потужності та підняти теплову стелю.

• Криві потужності GPU в 3DMark Wildlife Extreme та Steel Nomad Light

Наскільки ефективний rhe 1,5 ГГц Adreno?

They said they won't do the X elite video in a while because they are busy on other videos and no one would buy them now in China. It is more expensive and can do less. The only reason to buy them is the battery life but people would rather get an old Macbook for that in China. (From the comments under the YouTube video)

 

[FlameTail]

The difference in the boost clock speeds between the lowest Snapdragon X SKU and the highest one is absurd (3.4 GHz vs 4.2 GHz- 23% difference).

I am still unsure if this is due to binning/yield issues, or whether it is artifical segmentation.

If it is the former, then that's something they will have to fix in the 2nd gen X Elite.

If it is the latter...

Hawk Point's SKU lineup for reference.

The turbo frequency difference is <10% between the lowest and highest SKU.

Lunar Lake leaked SKU lineup:

13% difference in boost frequency between lowest and highest SKU.

 

[POWER4]

I am still unsure if this is due to binning/yield issues, or whether it is artifical segmentation.

It is the former. IIRC, Oryon was not supposed to even reach 4.0 GHz in its early versions. It was a "why not try?" situation for ST benchmark show-off. If the word that got out back then was accurate, they never got yields as good as originally expected, for whatever reason.

 

[ikjadoon]

Saw in my feed today that, according to Mercury Research, ARM lost market share in 2Q24. Specifically for notebooks,



These are probably before X-Elite had time to get a lot of volume into the market, but if MS is predicting further Surface sales in the next quarter, that's probably not a good sign for QC Windows laptops in general.

AMD, Intel in focus as data shows gains over Arm

AMD (AMD) and Intel (INTC) were in focus on Wednesday after Citi maintained its respective ratings on the semiconductor companies. Read for more.

seekingalpha.com

That is interesting. I might assume this is global + any-OS.

I don't think it's caused by Apple sales (up +19% QoQ in Q2 2024 vs Q1 2024). Other theories:

1. higher overall AMD & Intel shipments in Q2 2024 and / or
2. lower Arm-based Chromebooks (Qualcomm & Mediatek) shipments in Q2 2024 and / or
3. Osborne effect of WoA folks waiting for X Elite / Plus laptops in Q2 2024

FWIW, Mercury's data so far (source 1, source 2, source 3). They do make revisions, so some may be off by a bit, as I don't have the actual reports.

Quarter	Global Arm Desktop & Laptop Market Share
Q2 2020	<2%
Q3 2020	N/A
Q4 2020	3.4%
Q1 2021	5.9%
Q2 2021	~7.0%
Q3 2021	8.9%
Q4 2021	10.3%
Q1 2022	11.3%
Q2 2022	9.4%
Q3 2022	14.6%
Q4 2022	13.3%
Q1 2023	14.8%
Q2 2023	11.1%
Q3 2023	10.7%
Q4 2023	10.3%
Q1 2024	11.1%
Q2 2024	9.7%

It's interesting that the power consumption under load is so high (and the battery life is so low) given the high battery life of the other tests. It makes me think that Dell might be frequency limiting the CPU on battery (the HP model was doing this and had terrible on battery performance as a result). Some plugged/unplugged tests would be interesting.

That is a great point and a key missing data point in Notebookcheck's reviews.

IIRC, even Signal65's testing showed some random power saving features get activated (or power guzzling features get deactivated) in long battery life tests.

 

[FlameTail]

Revisiting old rumour:

 

[FlameTail]

Revisiting old rumour:

8850 = Snapdragon 8 Gen 5

 

[FlameTail]

It's going to be a pretty interesting situation if both Qualcomm and Apple are hitting ~5 GHz with their next gen chips (X2 Elite and M5). That would mean they have essentially caught up to Intel/AMD's laptop part clock speeds.

 

[majord]

That is interesting. I might assume this is global + any-OS.

I don't think it's caused by Apple sales (up +19% QoQ in Q2 2024 vs Q1 2024). Other theories:

1. higher overall AMD & Intel shipments in Q2 2024 and / or
2. lower Arm-based Chromebooks (Qualcomm & Mediatek) shipments in Q2 2024 and / or
3. Osborne effect of WoA folks waiting for X Elite / Plus laptops in Q2 2024

FWIW, Mercury's data so far (source 1, source 2, source 3). They do make revisions, so some may be off by a bit, as I don't have the actual reports.

Quarter	Global Arm Desktop & Laptop Market Share
Q2 2020	<2%
Q3 2020	N/A
Q4 2020	3.4%
Q1 2021	5.9%
Q2 2021	~7.0%
Q3 2021	8.9%
Q4 2021	10.3%
Q1 2022	11.3%
Q2 2022	9.4%
Q3 2022	14.6%
Q4 2022	13.3%
Q1 2023	14.8%
Q2 2023	11.1%
Q3 2023	10.7%
Q4 2023	10.3%
Q1 2024	11.1%
Q2 2024	9.7%

That is a great point and a key missing data point in Notebookcheck's reviews.

IIRC, even Signal65's testing showed some random power saving features get activated (or power guzzling features get deactivated) in long battery life tests.

Yes, it's important to test and tabulate balanced and power saving mode performance before doing these tests.. then try and correlate the resulting performance with the battery life . only ultrabook review seem to do first bit.

Notebookcheck use high perf for the full load, and balanced for the wifi , browser tests. they also don't specify if they turn off any automatic battery saver functionality (that kicks in at 10-20% iirc by default) , Screen brightness is stated, but not key backlighting , which can use a surprising amount of power on some machines.

In the case of this laptop, the smallish 1200P IPS screen would be a huge player in the seemingly very impressive idle and wifi tests , it's good, but not shockingly good considering this.

 

[FlameTail]

Charlie dropped an article (it's paywalled):

What is Qualcomm launching at IFA this year?

So what is Qualcomm launching at IFA this year?

www.semiaccurate.com

Bet it will be Purwa.

 

[FlameTail]

Nah. FlameTail is making sure the hype train continues there 🤭

Hey, I am not throwing around absurd numbers like +40%. I am wary of hype trains, particularly after what happened with Zen5.

We have scarce information about the next generation Oryon V2. Now that the X Elite launch is past us, we have 2 things to look forward to;
• Snapdragon 8 Gen 4
• Cheaper Snapdragon X SKUs

 

[gdansk]

Hey, I am not throwing around absurd numbers like +40%. I am wary of hype trains, particularly after what happened with Zen5.

I'm equally disappointed by SDXE. I genuinely hoped I could buy something as good as a MacBook Air and fanless but with an OLED display by this time.

Zen 5 missed my expectations by about 300MHz. SDXE missed my expectations by about 3000RPM.

 

[FlameTail]

Linux On The Snapdragon X1 Microsoft Surface Laptop 7, But Critical Features Missing - Phoronix

www.phoronix.com

With Linux 6.11 there is initial support for the ASUS Vivobook S15 and Lenovo Yoga Slim7x, more recently has been patches for enabling the X1 Elite powered ThinkPad T14s Gen 6, and now as of this morning also DT support for the Microsoft Surface Laptop 7 devices.

Slow progress.

 

[POWER4]

Bet it will be Purwa.

I guess we have a winner!

It will be.

 

[Hesperax]

Interesting section in one of the legal documents that ikjadoon posted in the ARM thread:

Qualcomm then began intensive work on designing three new version 8-A CPUs: [REDACTED] for the compute (e.g., laptops/PCs), mobile (e.g., cellphones), and automotive (e.g., digital cockpit) markets, respectively, and a fourth series of custom cores compliant with version 9-A of Arm’s architecture [REDACTED].

Maybe Pegasus will be ARMv9?

 

[FlameTail]

Interesting section in one of the legal documents that ikjadoon posted in the ARM thread:

Maybe Pegasus will be ARMv9?

Well spotted. I think more nuggets of information about Qualcomm's future roadmap can be found if we scour those documents.

Who is bored ?

 

[FlameTail]

I wouldn't be wholly surprised if Pegasus is indeed ARMv9. ARM's own Cortex cored have been ARMv9 for a while now, and Apple adopted v9* for the first time with the M4. Qualcomm would the last of the 3 major ARM designers to do so.

 

[poke01]

I wouldn't be wholly surprised if Pegasus is indeed ARMv9. ARM's own Cortex cored have been ARMv9 for a while now, and Apple adopted v9* for the first time with the M4. Qualcomm would the last of the 3 major ARM designers to do so.

Apple is still ARM v8.7

 

[FlameTail]

Apple is still ARM v8.7

It's somewhat hazy.

Apple M4 Support Added To The LLVM Compiler, Confirming Its ISA Capabilities - Phoronix

www.phoronix.com

In recent weeks there's been much speculation around the M4's ISA capabilities and this suspected to be Apple's first SoC relying on the ARMv9 architecture with some suggesting even ARMv9.4. So seeing the Apple-contributed compiler support for LLVM adding the M4 as an "ARMv8.7" based design was coming as a surprise. But in digging through the patch there is this code comment that explains the situation:
"Technically apple-m4 is ARMv9.2a, but a quirk of LLVM defines v9.0 as requiring SVE, which is optional according to the Arm ARM and not supported by the core. ARMv8.7a is the next closest choice."

 

[FlameTail]

Snapdragon 8G4 leak (translated from Chinese):

Several running points and power consumption data currently available + estimated:
About 8W multi-core running score 7300
Around 8W GPU running points AZ project tied 9300
The P core scores 3200, and the power consumption of a single core is less than 6W. It is estimated that the power consumption of a single core of the whole machine - the power consumption at no-load = about 9W (the so-called single-core power consumption),
which is 6x1.5=9
E core running points 1400-1500
The data is based on this summary
Some people have already obtained the process-tested engineering machine, and more real data will available later.

That mention of the E-core is very interesting. I assume that's GB6 Single Core.

1400-1500 means it has a little below half the performance of the P-core (~3200).

 

[FlameTail]

The Qualcomm Snapdragon X Architecture Deep Dive: Getting To Know Oryon and Adreno X1

www.anandtech.com

But while x86 emulation is largely a software task – it’s Prism that’s doing a lot of the heavy lifting – there are still certain hardware accommodations that Arm CPU vendors can make to improve x86 performance. And Qualcomm, for its part, has made these. The Oryon CPU cores have hardware assists in place to improve x86 floating point performance. And to address what’s arguably the elephant in the room, Oryon also has hardware accommodations for x86’s unique memory store architecture – something that’s widely considered to be one of Apple’s key advancements in achieving high x86 emulation performance on their own silicon.

Would the Oryon core in the Snapdragon 8 Gen 4 also support Total Store Reordering?

If the Snapdragon 8 Gen 4 isn't going to run Windows and has no need of x86 emulation, then they may as well remove it, right? From the recent Geekbench listing, we know that the 8 Gen 4 has a different revision of Phoenix core compared to the one in X Elite.

Now that I think about it, What does Apple do? Do thr CPUs in their A series chips support TSO like the ones in the M series?

 

[Doug S]

Now that I think about it, What does Apple do? Do thr CPUs in their A series chips support TSO like the ones in the M series?

They probably have the transistor support for it, but no way to actually enable it. They aren't going to change the design to remove something like that which has zero performance impact when not enabled. The number of transistors would be pretty small, we're not talking 0.1% but probably more like 0.01% of core area if not less.

 

[FlameTail]

Snapdragon 8G4 leak (translated from Chinese):

That mention of the E-core is very interesting. I assume that's GB6 Single Core.

1400-1500 means it has a little below half the performance of the P-core (~3200).

That lines up with the math I did 8 months ago:

Imma do some math.

We are gonna use GB5 Multi-core, since GB5's multi-core scales properly compared to GB6.

My goal: To attempt to guess how powerful Phoenix-M, the efficiency core of the 8 Gen 4 is.

Let's take the Apple A17 Pro.

GB5 MC: 6500 points

We know that the P core in the A17 Pro has 3x the performance of the E core. (Source Geekerwan SPEC2017).

From that we can write an equation

E-core = a
P-core = 3a

2 P-cores + 4 E-cores = 6500

2 (3a) + 4 (a) = 6500

10a = 6500

a = 650.

Therefore,
P-core = 1950 points
E-core = 650 points

Now I'll tell you the reason why I did all this calculation for A17 Pro. It is because the P-core in A17 Pro and P-core in SD8G4 are very similar, judging by the leaked benchmarks.

Hence we'll assume that the SD8G4 P-core (Phoenix-L) has the same score as A17 Pro P-core.

SD 8G4
GB5 MC : 9000 points (leaked)

2 (Phoenix-L) + 6 (Phoenix-M) = 9000
2 (1950) + 6 (Phoenix-M) = 9000
3900 + 6 (Phoenix-M) = 9000
6 (Phoenix-M) = 5100
Phoenix-M = 850

It's some rough math, but we ended up with Phoenix-M = 850.

That means Phoenix-M is about 30% faster than A17 Pro E-core.

I got Phoenix-L = 1950 and Phoenix-M = 850 point, which means again that Phoenix-M has a little less than half the performance of Phoenix-L.

Core	Clock speed	GB6 1T	Performance Per Clock
Phoenix-L	4.37 GHz	3250	743.707
Phoenix-M	3.0 GHz	1400	466.666

Could Phoenix-M be a different uarch after all?

 

[ridham05]

The Qualcomm slide does not necessarily say 45 TOPS INT4. It could just as easily be read as two different, completely unrelated, features: Supports INT4 and has a throughput of 45 TOPS. https://iconic-vision.co.uk For NPUs the performance is usually given in INT8, so using INT4 without clearly saying that it was INT4 could easily be considered misleading.

This lack of clarity is why my comment in the Zen4 APU thread hedged quite strongly. I have not seen any definitive source pointing at one over the other.