New Qualcomm silicon for new Surface May 2024

wco81

Legend
Microsoft has been one of the biggest beneficiaries of the excitement around AI.

In any event, the 4/2/24 Vergecast has some more details about what MS might be doing with AI. First, apparently they tried Bing branding and most people assumed Bing was just a rebranded ChatGPT. But in fact MS has hired some people from lesser-known AI companies, including Inflection, for their in-house AI unit. Probably hedging their bets.

But they're going to have a big Surface event in May and they expect to unveil devices with OLED. However, it will also be the unveiling of the Qualcomm silicon developed probably from their acquisition of former Apple engineers.

Apparently MS has expressed confidence to The Verge's reporters that not only would this new Qualcomm chip get them near Apple Silicon, they expect this first chip to have greater performance than the M3.

It will also have some kind of AI chip which isn't new but they will tout it for Windows AI PCs, which have to have top Intel chips and this other AI processor -- which would do things like noise cancellation.

The new Surface and AI PCs will have a Copilot button and the "killer" feature will be something to let you control the timeline of your usage. So for instance, you were searching for a trip to Europe a week or two ago. You'd tell the AI to take you to where you were and it would open all the tabs you had opened when you were doing this trip research.

Makes sense to strike while the AI iron is hot. MS among the tech giants is the one probably benefitting the most from AI -- well now that Nvidia is worth over $2 trillion, because its stock has risen so fast, maybe it's benefitted the second most.

If MS could announce by the end of the year that they have x million users regularly using Copilot and that they've shipped x million AI PCs, it would stoke MSFT even further upwards.

At least until the AI bubble bursts.;)
 
I get so confused about if they're using "AI" as a pr term or an actual descriptor. Looked up the whole NPU thing and started asking myself why everyone needs one so badly and how will I survive without one?!?!

The futar is scary.
 
DirectML still does not support NPUs.

Windows Studio Effects are available on Surface Pro 9 5G (SQ3), laptops with Ryzen AI APUs and Core Ultra, but not on my Surface Pro X (eye contact is available, but not background blur).


If they bring affordable fanless device, it would replace my Surface Pro X.


----
 
AMD and Intel's 9W 4nm chips could probably do fanless equally well with less compatibility headaches.

PS. NPU's are silly on Windows, only Apple has any real interest in local models. All that effort for slightly more efficient background removal in teams is kinda silly, the GPU can handle it well enough.
 
Last edited:
DirectML still does not support NPUs.
There has been dev preview of DirectML for couple months now with Intel NPU support. Pretty sure they'll try to release the proper version with support for AMD, Intel and Qualcomm NPUs as close to SD X Elite computers launch as possible.
 
~1900-2500 Points for Snapdragon X Plus
And it looks like MS will offer SDX Plus (base model?)
 
For comparison the Apple M2 Max SoC gets 3497 (Core ML) and the Apple M1 SoC gets 2998 (Core ML).

It clearly shows the Qualcomm Oryon cores are really good at certain workloads (I8).
 
Last edited:
For comparison the Apple M2 Max SoC gets 3497 (Core ML) and the Apple M1 SoC gets 2998 (Core ML).

It clearly shows the Qualcomm Oryon cores are really good at certain workloads (I8).
I'm not familiar at all with that benchmark. But I'd say that Oryon does more poorly on I8 than on F16/F32: https://browser.geekbench.com/ml/v0/inference/compare/367614?baseline=369007
It gets destroyed on Text but wins on Image Segmentation.

Edit: I had picked the one with 1900 as base... Picking the one with 2400 draws a very different picture.
That's odd. Is anyone familiar with that benchmark and could explain that difference?
 
Last edited:
I'm not familiar at all with that benchmark. But I'd say that Oryon does more poorly on I8 than on F16/F32: https://browser.geekbench.com/ml/v0/inference/compare/367614?baseline=369007
It gets destroyed on Text but wins on Image Segmentation.

Edit: I had picked the one with 1900 as base... Picking the one with 2400 draws a very different picture.
That's odd. Is anyone familiar with that benchmark and could explain that difference?
I didn't even notice the difference.

But they are the same device, one is using the DirectML backend and the other the CPU backend. Is the DirectML backend testing the NPU or GPU? It does list the GPU under Inference Information, so we can assume it is the GPU?

Screenshot 2024-04-23 at 10.18.54.png

Initially I only tested the CPU backend with the two Macs above but here are the scores for the M2 Max (9751 / Neural Engine), M2 Max (8239 / GPU) and the M1 (6821 / Neural Engine), M1 (3390 / GPU) for comparison.
 
I didn't even notice the difference.

But they are the same device, one is using the DirectML backend and the other the CPU backend. Is the DirectML backend testing the NPU or GPU? It does list the GPU under Inference Information, so we can assume it is the GPU?

View attachment 11182

Initially I only tested the CPU backend with the two Macs above but here are the scores for the M2 Max (9751 / Neural Engine), M2 Max (8239 / GPU) and the M1 (6821 / Neural Engine), M1 (3390 / GPU) for comparison.
I failed to notice one was CPU and the other was DirectML. I'll blame the lack of coffee :eek:

It's strange that the CPU is faster than the supposedly accelerated DirectML. This make me think it isn't using the NPU.
 
I missed this, too. But bad GPU drivers, who knows.

GeekbenchML cannot use the NPU under Windows ATM, DirectML doen't support NPUs, GPUs only.

Here some SQ3 results: https://browser.geekbench.com/ml/v0/inference/364108
Support for Intel NPUs has been out since February with DirectML 1.13.1, pretty sure there are already builds with Qualcomm & AMD support too even if they're not public.
 
The Snapdragon X Elite was previously reported to operate on two power limits; 23W and 80W. However, according to the latest findings, Qualcomm’s newest ARM chipset can touch almost 100W, and that is for the CPU alone, making it a power guzzler compared to its closest competition, Apple’s M3 Pro. However, there are several variables to keep in mind before criticizing the massive power draw of the chipset, and we will be talking about those details here.
 
"The Snapdragon X Elite was previously reported to operate on two power limits; 23W and 80W."

Wrong, it's device TDP.


Even Intel Core M3 with 4W TDP could consume upto 30 W to reach its turboclocks. So what's the point?

There’s a lot of variance in silicon devices, so Qualcomm lists the values as they should be able to be achieved by 95% and 50% of the manufactured chips, respectively. In practice, the better parts will simply get binned as the higher-end SKUs, though, so the distinction doesn’t have any significant implications in reality. However, it’s still important to keep in mind.

But why is it important and for whom?
 
Android Authority seems to have some power figures as well as performance figures here showing the different SKUs compared to the Intel Ultra 7 155H and the Apple M3 Pro.

Performance:
Screenshot 2024-04-26 at 08.37.51.png

Power:
Screenshot 2024-04-26 at 08.38.05.png

The Intel Ultra 7 155H used around 80 Watts while the Apple M3 Pro used around 42 Watts.
 
Back
Top