A Samsung Galaxy S25 with Google AI silicon sounds weird, however is it?

Who doesn’t love wild rumor, particularly when it blends high-profile telephones just like the Samsung Galaxy S25 with sizzling matters like AI? In response to the newest whispers, Samsung’s subsequent flagship telephone can be powered by a brand new flagship Exynos 2500 processor (along with an as-yet-unannounced Qualcomm Snapdragon 8 Gen 4). The report requires a hearty dose of salt, however the juicy hypothesis is that the chip will characteristic a devoted AI accelerator constructed by Google, not not like the Google TPU discovered within the Pixel 8’s Tensor G3 processor.

For the previous two generations, Samsung has developed its personal AI accelerator for Exynos, consisting of massive and little AI-focused cores (NPUs) and a extra conventional number-crunching digital sign processor (DSP). Introducing a Google-developed processor half can be a tacit admission that Samsung’s AI silicon is behind the competitors, both in efficiency or improvement instruments.

Normally, we’d say completely no manner, however this isn’t an inconceivable improvement. Google and Samsung have labored intently on the Tensor chip sequence and AI software program like Galaxy AI. Licensing further expertise between the 2 can be enterprise as traditional. ASUS already makes use of Google TPUs for its Coral IoT platform, so why not Samsung?

The advantages would, presumably, be stronger AI number-crunching capabilities for the Galaxy S25 sequence. Whether or not or not the rumor seems to be true (we’re nonetheless removed from satisfied), it has drawn my consideration to a key inflection level within the present cellular AI arms race that’s value diving into.

Somebody will win the AI race, sooner or later

The AI panorama has modified rather a lot up to now yr, and so has the {hardware} necessities to run the newest fashions and instruments. “Fundamental” early inferencing and cloud-based instruments have shortly given solution to generative fashions that may produce swathes of textual content, pictures, and even video on machine, however require vastly extra specialist compute energy to take action. Cell processor designers have been investing an increasing number of transistors into NPU processing components to maintain up, with Google, Samsung, Qualcomm, MediaTek, and Apple all pursuing barely completely different approaches.

There are at all times winners and losers in an modern discipline, and there’s a component of luck in predicting silicon improvement instructions so a few years prematurely. A part of the issue is that AI workloads fluctuate a lot, and determining which directions and bit-depths to help in {hardware} is tough whereas the state of purposes stays in flux. The altering panorama could make improvement prices prohibitive, particularly for those who’re not promoting your developments far and broad.

Google’s benefit is having its hand in each AI software program and {hardware} improvement, maybe holding its TPU improvement slightly extra in tune with altering developments than some others. There are sure to be advantages to utilizing AI {hardware} from the identical firm bringing Gemini Nano fashions to varied smartphones, in spite of everything.

Trying simply barely additional forward, the way forward for AI purposes received’t restricted to some proprietary OEM apps that ship together with your telephone (sorry, Galaxy AI). Shoppers will need to take their AI purposes throughout gadgets and platforms like different apps, necessitating third-party developer entry to AI accelerator {hardware}.

Nevertheless, I examined the Exynos and Snapdragon Samsung Galaxy S24 sequence AI efficiency numbers utilizing GeekBench ML (admittedly nonetheless in improvement) and located the app couldn’t entry further NPU efficiency by way of the Android NN API. Operating immediately on the CPU and GPU was quicker, which shouldn’t be the case. Maybe it is a driver situation, or among the finest AI telephones in the marketplace don’t seem to provide third-party builders easy accessibility to leverage their AI smarts by way of a core Android API.

I ran the identical check on Google’s Pixel 8 Professional and located a notable uplift in some workloads when utilizing Android NN (besides in 32-bit floating level exams, which see no profit). In different phrases, builders can run TensorFlow AI workloads on the Pixel’s TPU for vital efficiency features.

Operating Ignatov Andrey’s AI Benchmark (outcomes under) showcased a really comparable development. Android NN API confirmed principally no uplift over CPU cores and likewise scored far under Qualcomm’s QNN frameworks on the Galaxy S24 Extremely. Curiously, QNN on the S24 Extremely truly outclassed the Pixel 8’s finest rating by a large margin, highlighting Qualcomm’s AI compute energy. As well as, Qualcomm has its AI Hub instruments for builders to entry its personal on-device AI fashions. Worryingly, Samsung’s ENN possibility couldn’t run on the Exynos Galaxy S24, leaving it caught with a horrible Android NN-based rating.

This isn’t to take situation with the Galaxy S24’s AI efficiency; it clearly works effectively sufficient for Galaxy AI. The purpose is that builders aren’t going to need to implement three or 4 completely different frameworks to attain the most effective AI app efficiency to everybody. This can be a broader drawback.

Hopefully, the above has helped spotlight the 2 halves of the AI silicon problem: first creating the {hardware} after which coding the libraries and instruments to open up that {hardware} in a manner that builders can and need to construct on. That’s a hefty funding, extra so when silicon is altering so quick.

If Google has already carried out the API legwork work with its TPUs, as seems to be the case, maybe that’s a adequate cause for Samsung to contemplate Google’s resolution over an in-house possibility for future Exynos chipsets, particularly if it believes in a future marketplace for third-party AI apps.

Right this moment’s cellular AI panorama is a Wild West, with manufacturers chasing their very own distinctive {hardware} and software program combos in a bit to get forward of the competitors. Ultimately, the market will decide on probably the most sensible, if not essentially probably the most highly effective, concepts, and we’ll nearly definitely see way more homogeny between gadgets. I wouldn’t wager in opposition to Google right here, even when this shaky AI silicon rumor doesn’t become true for the Samsung Galaxy S25 sequence.