News Arm-powered Snapdragon X Elite laptop shown outperforming Intel Core Ultra by up to 10X in AI tests — Qualcomm fires early NPU shots at Intel

[Admin]

Qualcomm has published a short video which it says shows a Snapdragon X Elite laptop can outpace an Intel Core Ultra 7 laptop by up to 10X in AI workloads.

Arm-powered Snapdragon X Elite laptop shown outperforming Intel Core Ultra by up to 10X in AI tests — Qualcomm fires early NPU shots at Intel : Read more

 

[Roland Of Gilead]

Will be good to see some real benches and side by sides. I hope AMD's version can be tested along the other two. See the difference and similarities each have and how they perform.

Interesting times ahead with this AI malarky.

 

[setx]

Very nice demonstration that if you lose CPU performance and start substituting it with accelerators, someone is going to slap you hard, since it's not a walled garden of x86.

All "AI technologies" should be marketed and measured separately as they are not an indication of general performance.

 

[Neilbob]

I'm so weary of the whole AI thing at the moment. Everything seems to be about it and disregards the fact that most things people do are still in the category of 'general workload.'

I kind of prefer to think of the current jiggery-pokery as Artificial Augmentation, plus I like the AA acronym, so... sure, I completely understand that AA will be an increasing thing in the future and can do things like clever image manipulation and I've been shown one or two cool things it can do right here, but for now I'm unenthused about the so-called benefits it has with things like targeted search results, real-time chat and text responses and yada yada yada. I can do these things myself. Is it only me who doesn't care in the slightest about that stuff?

When the 'intelligence' actually manages to have an original thought and work with something other than what already exists, then I might be impressed... for the few seconds we have before the AI launches nuclear weapons at it's targets in Russia and they counterattack accordingly (that's a T2 reference just in case anyone is unaware, or I'm being political).

Perhaps this is just me being my grouchy old self being completely past it when it comes to modern tech. I have to keep up my thing of everything I say coming from a place of extreme cynicism and bitterness.

Yep, I'm the duffer waving a rake and yelling at the local brats to get off my lawn.

 

[Roland Of Gilead]

for the few seconds we have before the AI launches nuclear weapons at it's targets in Russia and they counterattack accordingly (that's a T2 reference

Brilliant 🤣

 

[Notton]

I don't really care for the AI stuff, but I would like to see how it performs against Strix Point, which is speculated at 48+ TOPS.

 

[jojosbuddy]

Good chance QC is benchmarking the SOM vs just the npu core on the ultra. Qc's famous for repurposing their kyro/hexagon GPU/dsps and this is likely the case. The GPU will edge out IntelAI boost, an fpga, as well as consume less power. But it's all from software: more prone to bugs and more expensive. nVidia is in the same camp and why it's about gpus today. Also if they benchmark a core ultra w/integrated arc GPU would be a way different story.

But once Intel figures out better parallelism in their fpga design, the NPUs going to blow gpus away and much easier and 10x cheaper to scale.

 

[sonichedgehog360]

But Intel has Arrow Lake and Lunar Lake waiting in the wings.

COPE.

Intel just released Meteor Lake, and we are comparing a product coming out in a few months (ETA June for Snapdragon X Elite) with two from Intel that quite possibly won't be out until end of year. Further, we have no idea how much better than AI performance is on Arrow Lake and Lunar Lake. Additionally, Intel is notorious for delays. Q3 and Q4 2024 is a best case scenario for Intel.

 

[Loadedaxe]

I'm so weary of the whole AI thing at the moment. Everything seems to be about it and disregards the fact that most things people do are still in the category of 'general workload.'

I kind of prefer to think of the current jiggery-pokery as Artificial Augmentation, plus I like the AA acronym, so... sure, I completely understand that AA will be an increasing thing in the future and can do things like clever image manipulation and I've been shown one or two cool things it can do right here, but for now I'm unenthused about the so-called benefits it has with things like targeted search results, real-time chat and text responses and yada yada yada. I can do these things myself. Is it only me who doesn't care in the slightest about that stuff?

When the 'intelligence' actually manages to have an original thought and work with something other than what already exists, then I might be impressed... for the few seconds we have before the AI launches nuclear weapons at it's targets in Russia and they counterattack accordingly (that's a T2 reference just in case anyone is unaware, or I'm being political).

Perhaps this is just me being my grouchy old self being completely past it when it comes to modern tech. I have to keep up my thing of everything I say coming from a place of extreme cynicism and bitterness.

Yep, I'm the duffer waving a rake and yelling at the local brats to get off my lawn.

No need for me to post, spot on what was in my head.

 

[Alvar "Miles" Udell]

Please note that the Qualcomm machine seems to have benefitted from a specially optimized Stable Diffusion release, though.

ARM in a nutshell, they run great when the software is designed for them.

But there are many legitimate uses for machine learning algorithms in many fields which will have a great positive effect, and the ability for an ARM chip to efficiently run them does bode well for them being able to be run natively , efficiently, and quickly locally on phones, tablets, and ARM based laptops which will extend their uses to medical and legal fields, among others, where privacy laws prevent their use on cloud based systems.

 

[usertests]

Has anyone given a TOPS number for Intel's first-gen NPU? Because if Qualcomm's future 45 TOPS NPU is beating an 8-10 TOPS showing from Intel, the results aren't really a surprise.

 

[OriginFree]

When the 'intelligence' actually manages to have an original thought and work with something other than what already exists, then I might be impressed... for the few seconds we have before the AI launches nuclear weapons at it's targets in Russia and they counterattack accordingly (that's a T2 reference just in case anyone is unaware, or I'm being political).

That's a WOPR of a statement.

 

[CmdrShepard]

This "benchmark" is suspect.

The only thing I am finding when I Google "Fast Stable Diffusion" is this repo.

It mentions RunPod, Paperspace, and Colab which if I am not mistaken means it runs in cloud, not locally.

 

[bit_user]

Good chance QC is benchmarking the SOM vs just the npu core on the ultra.

The article made it clear they're utilizing Intel's optimal, recommended configuration: spanning CPU cores, tGPU, and NPU!

​

once Intel figures out better parallelism in their fpga design, the NPUs going to blow gpus away and much easier and 10x cheaper to scale.

Why do you keep talking about FPGAs? Nobody is using FPGAs for their integrated NPUs. Not Intel or AMD.

 

[bit_user]

ARM in a nutshell, they run great when the software is designed for them.

The "Hexagon" DSPs in Snapdragon SoCs are designed by Qualcomm, themselves. They have nothing to do with ARM.

It's a similar story with the Adreno iGPUs found in Snapdragons.

 

[bit_user]

The only thing I am finding when I Google "Fast Stable Diffusion" is this repo.

It mentions RunPod, Paperspace, and Colab which if I am not mistaken means it runs in cloud, not locally.

That repo has 1.2k forks!

Qualcomm's Hexagon-optimized port is quite likely in a private fork. Enterprise github customers can have private repos, and I assume also private forks of public repos.

 

[CmdrShepard]

That repo has 1.2k forks!

Qualcomm's Hexagon-optimized port is quite likely in a private fork. Enterprise github customers can have private repos, and I assume also private forks of public repos.

That isn't much of an assurance that their demo didn't run in the cloud unless you believe that repository having forks is somehow proof they ran it locally?

If they said they used custom tuned fork of AUTOMATIC1111 repo I would have given them the benefit of a doubt even though A1111 can also work by connecting to cloud.

But this repo?

This is suspect AF. There's almost no code in it, just Jupyter notebooks.

 

[bit_user]

That isn't much of an assurance that their demo didn't run in the cloud unless you believe that repository having forks is somehow proof they ran it locally?

I'm not sure why you're concerned with "proof". As long as it's a manufacturer-provided demo, we will never have proof they didn't fake anything. The only sure way is to wait for launch + 3rd party reviews.

But this repo?

This is suspect AF.

It does bother me that the video has no notes providing details of their testing. At least when Intel & AMD release slide decks with internal testing results, they have copious end notes that detail exactly what was tested & how. Speaking of which, if you want to see how they tipped things in their favor, it's often worth giving those a close read.

 

[CmdrShepard]

I'm not sure why you're concerned with "proof". As long as it's a manufacturer-provided demo, we will never have proof they didn't fake anything. The only sure way is to wait for launch + 3rd party reviews.

Then why even dignify them with publishing some PR fluff piece when you have no proof what they are showing you is valid?

It does bother me that the video has no notes providing details of their testing. At least when Intel & AMD release slide decks with internal testing results, they have copious end notes that detail exactly what was tested & how. Speaking of which, if you want to see how they tipped things in their favor, it's often worth giving those a close read.

I am not saying Intel and AMD don't do similar things, but with them at least you know what they did.

 

[bit_user]

Then why even dignify them with publishing some PR fluff piece when you have no proof what they are showing you is valid?

Because their customers are primarily the phone & laptop makers. They have more to lose than they stand to gain, if what their claims turn out to be wildly inflated. So, while I expect they're trying to position their product in the best light, I believe their demo shows what they claim it does.

Something you might not know about Qualcomm is that they were actually the first phone SoC to have a NPU. They're hardly Johnny-come-lately's to this domain. They were designing their own GPU, DSP, and ISP cores even before that, so I think they have the chops to deliver on these claims.

https://www.qualcomm.com/news/onq/2...mm-zeroth-processors-brain-inspired-computing
​

Also, Tom's publishes like a ton of PR put out by tech companies. Way more press releases and product announcements than actual reviews. If that really bothers you, perhaps you might want to find somewhere else to haunt.

 

[CmdrShepard]

Because their customers are primarily the phone & laptop makers. They have more to lose than they stand to gain, if what their claims turn out to be wildly inflated. So, while I expect they're trying to position their product in the best light, I believe their demo shows what they claim it does.

Too bad we will never know until it's too late, eh?

Something you might not know about Qualcomm is that they were actually the first phone SoC to have a NPU. They're hardly Johnny-come-lately's to this domain. They were designing their own GPU, DSP, and ISP cores even before that, so I think they have the chops to deliver on these claims.

I do know everything about Qualcomm including that they are a scummy bunch when it comes to patents and licensinc. I hate them enough already to have a healthy dose of suspicion for everything they say.

 

[bit_user]

I do know everything about Qualcomm including that they are a scummy bunch when it comes to patents and licensinc. I hate them enough already to have a healthy dose of suspicion for everything they say.

IMO, they're their own worst enemy. They sabotaged their own CPU-design efforts, in the past, and I'm hoping they don't repeat those mistakes with the group from Nuvia.

I think a lot of the blame can be followed back to their activist investors and the hostile takeover bid by Broadcom, which resulted in the death of their server CPU product line, just as the initial product launch was being staged.