Discussion Qualcomm Snapdragon Thread

[qmech]

Yeah but the entire philosophy of boosting chips beyond their sweet spot is inherently bad and it is what has lead x86 CPUs to be so inefficient.

I am not the first one to say this. I recall someone else in this forum saying so, many years ago.

There are periods when one x86 company or the other will ship processors operating well beyond their efficiency peaks, but usually that is to compensate for being behind in performance or as a "halo" product.

Boosting as much as the system's thermal solution can handle is absolutely viable, particularly in a setting where the power is not necessarily a battery (laptop plugged in or a desktop). There are many times when the user just wants faster results (or more fps), regardless of the efficiency. (That is very rarely the case with phones, just to be clear.)

 

[SarahKerrigan]

ARM laptops can’t compete with other Windows laptops until they can run the vast majority of software without issue. This is not the case today.

How does that translate to the primary competition being Apple? And what are you basing the claim of them being unable to run software without issues on? Have you actually used them?

Edited to add: And do you seriously think the average NT/ARM laptop that sells today is a lost sale for Apple, not a lost sale for x86 laptops?

 

[moinmoin]

ARM laptops can’t compete with other Windows laptops until they can run the vast majority of software without issue. This is not the case today.

To add to that, what about hardware support? I haven't looked into it, but I imagine random hardware connected through USB using "Windows" drivers won't automagically be supported under WoA, will it?

 

[eek2121]

How does that translate to the primary competition being Apple? And what are you basing the claim of them being unable to run software without issues on? Have you actually used them?

Edited to add: And do you seriously think the average NT/ARM laptop that sells today is a lost sale for Apple, not a lost sale for x86 laptops?

I never made the claim they were competing with Apple. Regarding WoA, yes, I have used it. Compatibility is a mixed bag.

Don’t forget the lack of proper GPU support. NVIDIA does not have a (public) ARM port of their drivers. Gaming compatibility is very poor without access to a decent GPU.

Non-games are somewhat better, but at the point I tried some applications would not launch, including anything needing AVX.

I am actually wanting to give it another try pretty soon. I plan to attempt to install Windows 11 on my Raspberry Pi 5.

 

[eek2121]

To add to that, what about hardware support? I haven't looked into it, but I imagine random hardware connected through USB using "Windows" drivers won't automagically be supported under WoA, will it?

Correct, if the driver does not have an ARM port it probably won’t work.

I think some folks forget that the reason Apple was able to transition smoothly over to ARM is they controlled the software and the hardware. Nobody controls the software or hardware on the PC side. There is no guarantee, for example, that I am using a Microsoft toolchain to build my application. I could use any number of toolchains.

Apple could make Xcode spit out fat binaries pretty easily, so ports happen quickly. Microsoft can’t do this because they don’t control their ecosystem. For example, I have a driver compiled using mingw.

 

[FlameTail]

Non-games are somewhat better, but at the point I tried some applications would not launch, including anything needing AVX.

And how would that get fixed? ARM does not have AVX, but it does have NEON/SVE.

 

[moinmoin]

And how would that get fixed? ARM does not have AVX, but it does have NEON/SVE.

We are talking about an emulation/binary translation layer. Users of Windows expect Windows programs to run under Windows, so ideally such a translation layer would enable that without glitches. The benchmark there is Apple's Rosetta 2 incidentally, for which Apple's ARM chips were extended to support memory models specific to x86 in hardware (something that's commonly slowing down emulation by quite a lot if done in software).

 

[FlameTail]

Windows 12 will launch in June 2024 with a slew of AI PCs, according to Taiwan’s Commercial Times

Acer and Quanta rubbing their hands in anticipation.

www.tomshardware.com

According to this report, Windows 12 is coming in June 2024, with a slew of AI features.

Boys, I think it's all coming together now.

Remember how Qualcomm said the X Elite will be available in laptops starting from mid-2024?

 

[FlameTail]

I expect/hope that Windows 12 will also have excellent support and integration for ARM processors, in addition to the AI features.

 

[FlameTail]

Assuming that leak is correct about the 8G4 , it would mark the 3rd consecutive generation of massive MT performance uplift. 3900 -> 5200 -> 7200 -> 10 000

This is insane. We haven't seen this rate of massive generational uplifts in the mobile CPU space, since it's early days. In fact, it means the 8G4 is matching the Apple M2 in MT!


Now what can we infer from that leak regarding the composition of cores?

It's rumoured the 8G4 will have a 2+6 CPU.





It's interesting how these cores are named. Phoenix-I is the P-core and Phoenix-M is the E-core. But what's interesting is they both share the Phoenix name. Does this mean Qualcomm isn't designing a ground up new E-core? It sounds like how Zen4C is a derivative of Zen4.

 

[FlameTail]

See, this is why I said the 76W (80W when rounded off) number for the X Elite in CB2024 is not good. The M2 Max is consuming only half as much, which is a staggering difference. Although the M2 Max number is for CBR23.

 

[FlameTail]

So the Oryon CPU is ARMv8 and not v9. We have previously discussed in the Hamoa thread how this might be due to Nuvia initially getting a ARMv8 license, and after being acquired acquired by Qualcomm they may not have been able to obtain the ARMv9 License due to the IP dispute and lawsuit.

Qualcomm themselves officially stated the Snapdragon 8G4 coming next year, will have the Oryon CPU. If Qualcomm can't work out thongs with ARM, the CPU will be ARMv8.

That's a regression from the previous few generations of Snapdragon Mobile CPUs, which featured ARMv9 cores. What are the implications of this? Wouldn't it mean certain ARMv9 features which previous Snapdragon 8 chips supported are now not supported in the 8 Gen 4?

 

[qmech]

Assuming that leak is correct about the 8G4 , it would mark the 3rd consecutive generation of massive MT performance uplift. 3900 -> 5200 -> 7200 -> 10 000

This is insane. We haven't seen this rate of massive generational uplifts in the mobile CPU space, since it's early days. In fact, it means the 8G4 is matching the Apple M2 in MT!

GB6 MT doesn't measure multithreaded performance well at all. The argument put forth by proponents is that it more accurately reflects software that users actuality run. The result is performance that scales extremely poorly with higher number of cores. That's

Has anyone checked how it actually scales with the various components (cores, cache, memory, etc)? The end effect of the various subtests *seems* to be that it pretty much maxes out at 4 threads and really loves main memory bandwidth.

 

[FlameTail]

really loves main memory bandwidth.

LPDDR5X-9600 baby!

 

[Nothingness]

We are talking about an emulation/binary translation layer. Users of Windows expect Windows programs to run under Windows, so ideally such a translation layer would enable that without glitches. The benchmark there is Apple's Rosetta 2 incidentally, for which Apple's ARM chips were extended to support memory models specific to x86 in hardware (something that's commonly slowing down emulation by quite a lot if done in software).

That’s it and that’s why AVX emulation is not an issue in itself. And beyond the memory order extension, having a closed software environment helps a lot.

 

[Nothingness]

GB6 MT doesn't measure multithreaded performance well at all. The argument put forth by proponents is that it more accurately reflects software that users actuality run. The result is performance that scales extremely poorly with higher number of cores. That's

Has anyone checked how it actually scales with the various components (cores, cache, memory, etc)? The end effect of the various subtests *seems* to be that it pretty much maxes out at 4 threads and really loves main memory bandwidth.

GB6 render test scales in a way similar to Cinebench (and yes it scales way beyond 4). Looking at the aggregated score is pointless as is the case for most benchmarks (the extreme being AnTuTu that aggregates CPU, GPU and other things).

 

[qmech]

GB6 render test scales in a way similar to Cinebench (and yes it scales way beyond 4). Looking at the aggregated score is pointless as is the case for most benchmarks (the extreme being AnTuTu that aggregates CPU, GPU and other things).

Obviously the render subtest scales well, but the number being published/leaked is always the composite MT score.

If your argument is that this number is useless, I don't completely disagree (some information is still capable of being extracted), but that's the number we have. It's hard to discuss numbers we don't have.

 

[FlameTail]

Okay Okay! You guys have established your point about GB6. What about the rest of the information in my long comment?

 

[FlameTail]

Snapdragon 8 Gen 3 is doing 32 FPS in 3DMarkWLE.

That means it has exceeded the Radeon 780M!

Insane.

 

[FlameTail]

I wonder, isn't the 8G3 GPU bandwidth bottlenecked?

It's said the Radeon 780M is bottlenecked by the limited memory bandwidth it has. But it can go all the way to 120 GB/s max with LPDDR5X-7500 paired with the 128 bit channel width.

8G3 has only 76 GB/s at max with LPDDR5T-9600 + 64 bit bus.

 

[FlameTail]

Do you all appreciate speculatory concepts of future CPUs ?

I was bored and made some of what a potential X Elite succesor could look like.

 

[Nothingness]

Obviously the render subtest scales well, but the number being published/leaked is always the composite MT score.

If your argument is that this number is useless, I don't completely disagree (some information is still capable of being extracted), but that's the number we have. It's hard to discuss numbers we don't have.

Yes, that's my admittedly trivial argument It's unfortunate most leaks don't provide subtest scores, and that's why I think discussing the MT score is often pointless and misleading.

Okay Okay! You guys have established your point about GB6. What about the rest of the information in my long comment?

Do you mean the one about ARMv8 vs ARMv9? Isn't the main selling point of ARMv9, SVE2? If yes, losing it is no big deal in the current situation given that Qualcomm has disabled it in the current generation of chips

 

[FlameTail]

Do you mean the one about ARMv8 vs ARMv9?

Yes and the one about Oryon cores:

Isn't the main selling point of ARMv9, SVE2? If yes, losing it is no big deal in the current situation given that Qualcomm has disabled it in the current generation of chips

What!?

 

[FlameTail]

Why argue about GB6, when we have GB5 numbers!?

It is indeed matching Apple M2!

Also, details about the cores:


This is concerning. So the 8G4 still uses first gen Oryon core? This means it won't beat Apple A18 in ST, and we may not see the X Elite successor at the next Snapdragon Summit.

 

[soresu]

This is concerning. So the 8G4 still uses first gen Oryon core? This means it won't beat Apple A18 in ST, and we may not see the X Elite successor at the next Snapdragon Summit.

Doesn't concern me at all.

It's an interesting thing to measure with Geekbench, but in reality smartphones really don't see much benefit from better CPUs anymore, even with Android's less than efficient system architecture.

The only thing I'd be concerned about is it burning a hole in my pocket if it uses anywhere near as much juice as all these graphs on SD X Elite are implying 😅