Future Tech

Intel, AMD take a back seat as Qualcomm takes center stage in Microsoft's AI PC push

Tan KW
Publish date: Tue, 21 May 2024, 10:35 AM
Tan KW
0 447,391
Future Tech

Build Microsoft isn't waiting around for Intel and AMD to get their neural processing units (NPUs) up to snuff and is pushing its AI PC agenda forward with Qualcomm's Snapdragon X Elite and X Plus system-on-chips (SoCs).

During a special event ahead of Tuesday's Build conference, Microsoft revealed these latest Arm-compatible notebook processors from Qualcomm will power the Windows giant's next-gen Surface Laptop and Pro tablets. Microsoft's new Surfaces are, we're told, just two of 20 "Copilot+ PCs" Qualcomm says it will launch with X Elite and Plus chips starting June 18.

First teased last autumn, the 4nm TSMC-made X processor pair are among Qualcomm's first to use a Nuvia-derived Arm-compatible CPU core design. If you don't recall, Qualcomm acquired Nuvia back in 2021 for $1.4 billion, kicking off a still-ongoing lawsuit with Arm along the way.

We're told the X Plus has ten CPU cores, and the X Elite will sport 12. The Elite can clock up to 4 to 4.2 GHz for two cores, or up to 3.4 to 3.8 GHz for all cores; the Plus can go up to 3.4 GHz for all cores. If Microsoft is to be believed, Qualy's chips are more than ready to give Apple's homegrown silicon a run for its money.

Citing Geekbench 2024, Redmond says its X Elite-equipped Surface Laptop delivers higher multi-threaded performance than Apple's M3 MacBook Air. It doesn't appear Microsoft has published those results just yet, but similar submissions from Acer and others would seem to back those claims up.

Having said that, single core performance does appear to be higher on the M3. So that would mean Apple's eight-core M3 in the MacBook Air is faster per CPU core, though Qualcomm wins out by throwing more CPU cores and potentially more power at the benchmark.

It isn't clear how much power the X Elite needs to hit that mark. When Qualcomm unveiled the X Elite last year, we were told the chip could be configured up to 50 watts. For reference, the M3 is widely believed to have a TDP of around 20 watts.

We've reached out to Microsoft for comment on how hot it's letting Qualcomm's chips run, and the IT giant declined to say anything on the subject. As usual and as such, we recommend taking any and all vendor benchmark claims with a decent pinch of salt.

In any case, it appears that, equipped with a 12-core X Elite, 16GB of RAM, and 512GB of flash storage, Microsoft's Surface Laptop will at least be cheaper than Apple's latest MacBook Air. So, in terms of performance per dollar, Microsoft may have an edge.

Alongside Qualcomm's new Snapdragon duo, Microsoft is also rolling out an emulation layer called Prism, presumably for running legacy x86 apps on the laptop's Arm cores. The software giant claims that, using this emulation layer, its latest-gen Surfaces are as much as two times faster than the older Arm-based Surface Pro 9.

The good news is you may not need emulation much longer. Microsoft says there are already native versions of Office, Chrome, Spotify, Zoom, WhatsApp, Blender, the Affinity suite, and Davinci Resolve ready to run on the machines. Meanwhile, Adobe has committed to bringing its apps to Windows on Arm.

It probably doesn't hurt that Apple started its transition to Arm four years ago, and many of these apps already have Arm64 compatible binaries.

All about them TOPS

However, building a performance-competitive Arm notebook SoC is arguably secondary to the NPU found at the heart of the X pair. Capable of pushing 45 TOPS of ML performance at INT8, the X Elite and Plus are claimed to be the only part available today that meets Microsoft's 40 TOPS target - something, we'll note, Qualcomm was keen to highlight in its own launch announcement.

As it stands, Intel and AMD's rival parts - the Intel 14th-gen (Meteor Lake) and AMD Ryzen 8040-series - top out at 10 and 16 TOPS respectively. Qualcomm's lead won't last forever. While Intel had previously teased that its upcoming Lunar Lake Core processors would deliver 45 NPU TOPS, it revealed on Monday that the chip's integrated GPU was capable of pushing another 60 TOPS, bringing the chip's total output to more than 100 TOPS when it hits the market in Q3.

We expect similar gains from AMD when its next-gen Ryzen processors arrive later this year. Until those chips hit the market it seems that Qualcomm has the AI PC market cornered.

Compared to Apple's M3, Qualcomm boasts its silicon has up to 2.6x higher NPU performance. However, while the Neural Engine in the M3 maxes out at 18 TOPS, Apple's newly announced M4 SoCs are capable of pushing this to 38 or maybe higher. Unfortunately, that chip is only available in the iGiant's pricey iPad Pros.

As for what you'll do with all those NPU TOPS, Microsoft says they'll enable a variety of new AI features coming to Windows. This includes Cocreator for on-device art-generation in Paint; and live captions in things like video calls. And then there's Recall...

Microsoft also touted several apps which have added support for NPU acceleration, including Davinci Resolve, Cephable, and CapCut.

Going against the grain

Qualcomm's X Plus and X Elite promise impressive performance and the NPU grunt Microsoft needs to advance its AI PC agenda, but the chips are also at odds with Arm's trajectory.

As we mentioned before, Arm is none too happy that Qualcomm acquired Nuvia and is now designing its own Arm-compatible CPU cores; both Qualcomm and Nuvia are Arm licensees, and Arm isn't thrilled at the way Qualcomm ingested Nuvia's Arm-compatible technology. So unhappy, in fact, that Arm sued Qualcomm over the whole debacle, arguing the Snapdragon maker was in breach of its architectural license with Arm and needed to negotiate a new one to use Nuvia's designs, no doubt a higher price.

Nuvia's cores are also a step backward as far as some in the Arm world are concerned. As we understand it, the Nuvia-derived CPU cores at the heart of Qualcomm's X-series SoCs are based on the older Armv8 architecture and rely on a combination of GPU and NPU acceleration for speeding up AI workloads along with Arm's usual NEON CPU-level instructions.

By contrast, Arm has spent the past few years baking AI-accelerating features - such as Scalable Matrix Extensions 2, or SME2 - into its newer Armv9 CPU architecture. These are already being employed by Apple in its M4 Armv9 SoCs.

Qualcomm's decision to move away from Armv9 - the tech giant has previously employed the architecture in its smartphone SoCs - and its CPU-level improvements in favor of offloading AI work to NPUs means the Arm world is facing a potential divide: On one side, you've got Microsoft and Qualcomm today pushing Armv8 processors with custom units to specifically handle ML code; and on the other, you've got Arm trying to get people onto Armv9 where AI-accelerating instructions are baked into the CPU architecture as standard.

Considering Microsoft has thrown its weight behind Qualcomm's Armv8 chips in its quest to drive AI features out of the datacenter and out to the client; Apple being Apple; Google in love with its own on-device ML accelerators; and Samsung and MediaTek floating in the middle, we feel for developers - from OS and framework makers to app builders - trying to achieve maximum portability and maximum performance on Arm. ®



Be the first to like this. Showing 0 of 0 comments

Post a Comment