Qualcomm revealed X Elite's benchmark scores

[deconstruct60]

I would find that very surprising. The comparison with the X4 is on another slide.

And. If you go back to your " rolling down hill with hurricane tailwinds" chart is is just 36%. Go to two bars up with "normal" turbo clocks and not in an android. It is in 15-25% zone like first chart above. Look at the number not the size of the bar (which as pointed out are suffering from proportional issues. ).

The chart on right above is more vector (SVE2 implementation uplift) than single threaded. There is no SME here but that in no way means there are not substantial SVE2 implementation changes. More explicit code, but vectors do "matrix math" just fine.

 

[deconstruct60]

The performance claims are very vague. "+36% over 2023 premium Android" could be anything between 2500 and 2900 GB6 points.

If relative to X4 cores (e.g., Snapdragon 8s gen 3) it would be in the 2700-2900 zone ( ~2150 + 30-36% ). It doesn't have to be a latest gen M-series 'killer' chip to sell in high volume. For the average price point in the Windows market, M-series isn't hitting the mark either. SoC cost likely matters.

 

[Xiao_Xi]

The Register may have shed some light on why the X925 doesn't include SME instructions.

Arm's licensees, from Qualcomm to Google, love putting their own AI acceleration in their processors as this helps those designers differentiate their products from one another. And Arm doesn't want to step too much on people's toes and publicly state that it's not a fan of that custom acceleration. Arm staff stressed repeatedly to us that it is not anti-NPU.

Arm really wants chip designers to migrate to using Armv9, which brings in more neural network acceleration on the CPU side. And that's kinda why Arm has this beef with Qualcomm, which is sticking to Armv8 (with NEON) and custom NPUs for its latest Nuvia-derived Snapdragon system-on-chips. You've got the likes of Apple on one side using Armv9 and SME2 in its latest M4 chips, and Qualcomm and others on the other side persisting with NPUs. Arm would be happier without this fragmentation going forward.

CPU designer Arm emphasizes usefulness of CPU cores for AI

As it offers 3nm shake-and-bake Cortex-X925, A725 processor designs for phones, PCs

www.theregister.com

 

[name99]

Pretty good chance they are comparing to the X4. From same article.

There have been rumors about the 'Cortex X5'. This appears that Arm is doing some renaming to a longer naming scheme. ( In part probably because they are going to sell it as a CSS chiplet. )


The X4,X5 probably have some overlap with the Neoverse N1,N2,N3 work. It would be pretty surprising if the Xn got SME ( especially something like SME FP64) before the Neoverse series got it.


Not sure Arm is really trying to match the P/E cores of Apple. They are still trying to throw a triad of cores at their standard cluster.

Nominally, the phone is likely running off the middle ( or smaller A core) one most of the time. Except for folks who stare at their phones all day with it plugged in, there are nominal loads most of the time.
(and again.. the curve shows X3 , X4, 'X5' progression.





Rumor reports earlier this year were that X5 wasn't being a perf/watt champion. Perhaps X5 tried on TSMC N4 instead of N3. Or perhaps an X5 that is on Intel 3. It looks like Arm is trying to kill two birds with one stone. One a top end phone SoC along progression path with somewhat similar PPA as X1,X2,X3 and two trying to blunt Oryon ( relaxed some of the area to chance the General PC market for single threaded scores. ) .





Pretty likely that is "placed in a bigger tablet or PC laptop chassis". A bigger cooling systems isn't necessarily more advanced one; just has a bigger thermal capacity if it has a decent design. Qualcomm may have given up on pushing something like the Spapdrogn 8cx into the PC Winows market, but decent chance another player will show up to fill some lower BOM cost optoin than what Qualcomm is selling.

"This appears that Arm is doing some renaming to a longer naming scheme"

925
725
520

Like AMD, copying i3, i5, i7...

 

[deconstruct60]

The Register may have shed some light on why the X925 doesn't include SME instructions.

Not really.

Arm really wants chip designers to migrate to using Armv9, which brings in more neural network acceleration on the CPU side. And that's kinda why Arm has this beef with Qualcomm, which is sticking to Armv8 (with NEON) and custom NPUs for its latest Nuvia-derived Snapdragon system-on-chips. You've got the likes of Apple on one side using Armv9 and SME2 in its latest M4 chips, and Qualcomm and others on the other side persisting with NPUs. Arm would be happier without this fragmentation going forward.

CPU designer Arm emphasizes usefulness of CPU cores for AI

As it offers 3nm shake-and-bake Cortex-X925, A725 processor designs for phones, PCs

www.theregister.com

This part is a bit too "loose" with terms.

" ... Specifically, we're talking about Armv9's Scalable Matrix Extension (SME2) as well as its Scalable Vector Extension (SVE2) instructions. ..."

The SVE2 was introduced in 2017

Arm releases SVE2 and TME for A-profile architecture

Arm has released early technical details of two new technologies for its A-Profile architecture. These are Scalable Vector Extension v2 and Transactional Memory Extension.

community.arm.com

SME came with Armv9-A in 2021:

" ... SME builds on the Scalable Vector Extensions (SVE and SVE2), adding new capabilities to efficiently process matrices. Key features include: ...

...
A new operating mode is added, Streaming SVE Mode. When in Streaming SVE Mode, the new SME storage and instructions are available, as well as significant subset of the existing SVE2 instructions. When not in Streaming SVE mode, behavior is unchanged from SVE2. Applications can switch between operating modes depending on what is needed. ...
"

Scalable Matrix Extension for the Armv9-A Architecture

In this blog, read the details for Scalable Matrix Extension (SME). This is a new extension from the latest Arm Vision day announcement for Armv9-A.

community.arm.com

SVE2 predates SME it doesn't "belong to it". The "Streaming SVE" stuff 'belongs' with SME. Arm "General Matrix Multiply" is part of Armv8.6 also.

A-Profile Architecture

The Arm Application-profile (A-profile) architecture targets high-performance markets, such as PC, mobile, gaming, and enterprise.

developer.arm.com

Similarly SVE came in with Arm v8 alongside Neon. When 'new' lots of these features are 'optional'. Can be "v9" and not implement SME (implementors can 'cherry-pick' features until they go 'required' ) 8.1 will have a different set of required features than 8.8. Likewise in the 9.x series. SVE2 being 'older' means it is closer to get to the 'required' status than SME is. That is part of the issue for this generation.


If the X925 had SME then the A520 they they put into the same compute tile die would also need SME. That would make the A520 get bigger. ( or you would have to turn it off in the X925 and it would pragmatically 'dead' silicon , e.g., AVX-512 in Intel's gen 12-13. ). If Arm is going to have three layers of highly coupled CPU cores then they need to roll SVE2 out to all of them before moving on to SME. That is just going to take time.

The notion the article presents is that Arm is in SME right now or bust mode. SME not being in the X925 is more demonstrative that isn't true. Besides Apple has a NPU even though they have AMX/SME.

( Converstational capture/generation might want an NPU to run all the time where the CPU may or may not need to run . Depends upon how big and coupled the model is to the rest of the computations running for an interactive app. )


" ... Now Arm has taken that shake-and-bake approach to personal or client devices, and will offer complete physical implementations of the above new Cortex CPU and Immortalis GPU core designs under the banner of Compute Subsystems for Client. These designs were made with the help of TSMC and Samsung, ..."
(also from the linked article.)

I think Intel is in that fab mix also. ( as the article points out the implementation issues get coupled into the system issues. If Arm is packaging mulitple types of cores then they overall system has an impact on which features roll out at which times. )



P.S. Recently rumors that Arm wants to get into the "AI chips" market also.

" ...
SoftBank-owned Arm is reportedly preparing to add AI chips to its product portfolio starting in 2025.

The chip designer plans to set up an AI division with prototypes available in spring 2025 and mass production handled by contract manufacturers starting in fall, reports Nikkei. ..."

SoftBank reportedly preparing for Arm to tackle AI chips

Plus Foxconn defends Wisconsin efforts, Korea to spend on chips, and more

www.theregister.com

they can't be tooooooo far in the anti-matrix-from-CPU-cluster camp.

 

[deconstruct60]

"This appears that Arm is doing some renaming to a longer naming scheme"

925
725
520

Like AMD, copying i3, i5, i7...

More so just being 'internally consistent. ( they had switched to 700 , 500 a while back).

Qualcomm 8 gen 3
X2 715 710 510

Qualcomm 8 gen 4
X4 720 520


https://www.anandtech.com/show/21307/qualcomm-announces-snapdragon-8s-gen-3

The low-mid cores had more options and three digits. The "X" big core was just one digit. Now all sizes have three digits.

But also a jumbled mess. xx5 ( for 5th) generation X . Or perhaps starting over x2x (even though had an X2).

 

[Sydde]

First reports on the new Snapdragon X Elite laptops starting to surface…

nicely done

 

[MrGunny94]

Thing is, most people will be using the Plus and not Elite.. Let's wait and see..

 

[chmania]

Let's wait and see..

Yes.
WWDC will be over by June 14th, but the Copilot+ PCs would be released on the 18th June.

 

[stinksroundhere]

Yes.
WWDC will be over by June 14th, but the Copilot+ PCs would be released on the 18th June.

They are already in the hands of some people, they are already in the warehouses and stockrooms, performance and thermals are already leaking. The hardware and software isn't going to get better in the next 18 days.

Their security problems have already been written about.

Recall: Stealing everything you’ve ever typed or viewed on your own Windows PC is now possible.

Photographic memory comes to Windows, and is the biggest security setback in a decade.

doublepulsar.com

 

[chmania]

They are already in the hands of some people, they are already in the warehouses and stockrooms, performance and thermals are already leaking.

If you want to believe in "leaked" news... 😏
No business entity would let the product out before the release day, that is, if they want to keep their future business intact and not get sued for billions.

 

[stinksroundhere]

If you want to believe in "leaked" news... 😏

Some of that came directly from an MS presentation and you have commented in those threads, which includes this one. But whatever. Having debated tech, benchmarks ad previews since the 90s I always know what kind of people are going to pop up in discussions.

No business entity would let the product out before the release day

And yet there's a link in my post above and a dozen previews on YouTube.

 

[chmania]

Some of that came directly from an MS presentation and you have commented in those threads, which includes this one. But whatever. Having debated tech, benchmarks ad previews since the 90s I always know what kind of people are going to pop up in discussions.

And yet there's a link in my post above and a dozen previews on YouTube.

It is up to you, if you want to believe in "leaked" news...

 

[MrGunny94]

Looks like the supplied Linux image for Dell Latitudes for these will also be certified Ubuntu Images just like their x86 counter parts.

I'm quite curious about these models but until we have everything upstream on the Linux Kernel (at this stage looks too early)

More info: https://www.qualcomm.com/developer/...nux-kernel-support-for-the-snapdragon-x-elite

For kernel versions 6.10 and 6.11, we’re also working on merging more features in these areas:

USB host On-board display (eDP) GPU Memory DCVS CPUFreq Speakers/MIC/Headset

Battery External DP Suspend/Resume Camera Video

In short, our roadmap for the next six months includes work in these areas:

• End-to-end hardware video decoding, on Firefox and Chrome
• Implementation of the libcamera-SoftISP camera solution
• GPU and CPU performance optimizations
• Power optimizations (Suspend/DCVS)
• Making our firmware openly available (in Linux-firmware)
• Access to easy installers (Ubuntu and Debian)

 

[Random_Matt]

Price will dictate their success or not, they are not exactly cheap are they?

 

[senttoschool]

I would find that very surprising. The comparison with the X4 is on another slide.

It's a shame Arm hasn't posted the slides.

+14% GB6 = 2,466

A17 Pro = 2,907

A18 Pro estimated = 3,600

Based on this, iPhone 16 would be 46% faster than the fastest X925 Android phone in SC.

 

[leman]

+14% GB6 = 2,466

A17 Pro = 2,907

A18 Pro estimated = 3,600

Based on this, iPhone 16 would be 46% faster than the fastest X925 Android phone in SC.

Those improvements refer to same frequency and memory. I’d assume the new cores are clocked higher.

 

[chmania]

Based on this, iPhone 16 would be 46% faster than the fastest X925 Android phone in SC.

Meaning based on nothing, but hearsay...

 

[senttoschool]

Those improvements refer to same frequency and memory. I’d assume the new cores are clocked higher.

You’re right. Based on their other slides, I initially concluded that it’s slightly faster than the A17 Pro at 3.8ghz but wasn’t sure if real Android phones would clock that high.

 

[Xiao_Xi]

Qualcomm has started the competition to see who has the most efficient NPU.

 

[MrGunny94]

Qualcomm has started the competition to see who has the most efficient NPU.

View attachment 2384601

I wonder if any of this pushing from Qualcomm will force Apple's hand to release the M4 early at this stage...

I mean at the end of the day nobody outside of the enthusiast sphere really cares about these power point presentations. No casual user is buying a computer based on these.

 

[Icelus]

The Qualcomm Computex 2024 Keynote Live Blog (10:30pm PT/05:30 UTC)

www.anandtech.com

 

[senttoschool]

Qualcomm has started the competition to see who has the most efficient NPU.

View attachment 2384601

That's a huge advantage over M3 NPU in perf/watt. I wonder if it's real.

 

[crazy dave]

That's a huge advantage over M3 NPU in perf/watt. I wonder if it's real.

A third party (albeit one with early access to the hardware) said about 2.1x more efficient. It's certainly plausible - most NPU tasks are embarrassingly parallel or close enough at those TOPs and so you take the same approach that Apple (and to be fair the Nvidia MaxQ line) takes for its GPUs vs Nvidia "mobile" GPUs where you can get the same performance for 1/3 the power cost: more cores, lower clocks, huge efficiency gains. The downside is that you have to spend die area, which is expensive, to do it. So it has to be a priority.

I wonder if any of this pushing from Qualcomm will force Apple's hand to release the M4 early at this stage...

Along with all the other NPU announcements coming from other vendors? I'm beginning to suspect it might've been. I had joked earlier that this is why they actually released the M4 this early, now I'm somewhat more serious.

I mean at the end of the day nobody outside of the enthusiast sphere really cares about these power point presentations. No casual user is buying a computer based on these.

To be fair, Computex is an industry trade show for enthusiasts and even more importantly OEMs. This is about chip makers generating hype for getting design wins.

 

[Xiao_Xi]

Chips and Cheese wrote an article about Qualcomm's earlier version of the NPU.

Qualcomm’s Hexagon DSP, and now, NPU

Cell phones have to handle telecommunications and audiovisual processing while maximizing battery life. Digital Signal Processors, or DSPs, use specialized hardware to offload these tasks from the …

chipsandcheese.com

I wonder if Gerard Williams will talk about the NPU or just the Oryon core during his presentation at the upcoming Hot Chips.

Advance Program

A Symposium on High Performance Chips

www.hotchips.org