It’s been 8 years of phone AI chips — and they’re still wasting their potential

It’s been somewhat over eight years since we first began speaking about Neural Processing Models (NPUs) inside our smartphones and the early prospects of on-device AI. Massive factors if you happen to keep in mind that the HUAWEI Mate 10’s Kirin 970 processor was the primary, although related concepts had been floating round, notably in imaging, earlier than then.

In fact, quite a bit has modified within the final eight years — Apple has lastly embraced AI, albeit with combined outcomes, and Google has clearly leaned closely into its Tensor Processor Unit for the whole lot from imaging to on-device language translation. Ask any of the massive tech firms, from Arm and Qualcomm to Apple and Samsung, and so they’ll all inform you that AI is the way forward for smartphone {hardware} and software program.

And but the panorama for cell AI nonetheless feels fairly confined; we’re restricted to a small however rising pool of on-device AI options, curated principally by Google, with little or no in the way in which of a inventive developer panorama, and NPUs are partly in charge — not as a result of they’re ineffective, however as a result of they’ve by no means been uncovered as an actual platform. Which begs the query, what precisely is that this silicon sitting in our telephones actually good for?

Earlier than we will decisively reply whether or not telephones actually “want” an NPU, we must always most likely acquaint ourselves with what it really does.

Identical to your telephone’s general-purpose CPU for working apps, GPU for rendering video games, or its ISP devoted to crunching picture and video information, an NPU is a purpose-built processor for working AI workloads as shortly and effectively as attainable. Easy sufficient.

Particularly, an NPU is designed to deal with smaller information sizes (similar to tiny 4-bit and even 2-bit fashions), particular reminiscence patterns, and extremely parallel mathematical operations, similar to fused multiply-add and fused multiply–accumulate.

Now, as I mentioned again in 2017, you don’t strictly want an NPU to run machine studying workloads; numerous smaller algorithms can run on even a modest CPU, whereas the information facilities powering numerous Massive Language Fashions run on {hardware} that’s nearer to an NVIDIA graphics card than the NPU in your telephone.

Nonetheless, a devoted NPU will help you run fashions that your CPU or GPU can’t deal with at tempo, and it will probably usually carry out duties extra effectively. What this heterogeneous method to computing can price when it comes to complexity and silicon space, it will probably achieve again in energy and efficiency, that are clearly key for smartphones. Nobody desires their telephone’s AI instruments to eat up their battery.

In the event you’ve been following the continuing RAM value disaster, you’ll know that AI information facilities and the demand for highly effective AI and GPU accelerators, notably these from NVIDIA, are driving the shortages.

What makes NVIDIA’s CUDA structure so efficient for AI workloads (in addition to graphics) is that it’s massively parallelized, with tensor cores that deal with extremely fused multiply–accumulate (MMA) operations throughout a variety of matrix and information codecs, together with the tiny bit-depths used for contemporary quantized fashions.

Whereas fashionable cell GPUs, like Arm’s Mali and Qualcomm’s Adreno lineup, can help 16-bit and more and more 8-bit information sorts with extremely parallel math, they don’t execute very small, closely quantized fashions — similar to INT4 or decrease — with anyplace close to the identical effectivity. Likewise, regardless of supporting these codecs on paper and providing substantial parallelism, they aren’t optimized for AI as a main workload.

Not like beefy desktop graphics chips, cell GPU architectures are designed at the start for energy effectivity, utilizing ideas similar to tile-based rendering pipelines and sliced execution items that aren’t completely conducive to sustained, compute-intensive workloads. Cell GPUs can undoubtedly carry out AI compute and are fairly good in some conditions, however for extremely specialised operations, there are sometimes extra power-efficient choices.

Software program growth is the opposite equally essential half of the equation. NVIDIA’s CUDA exposes key architectural attributes to builders, permitting for deep, kernel-level optimizations when working AI workloads. Cell platforms lack comparable low-level entry for builders and system producers, as a substitute counting on higher-level and infrequently vendor-specific abstractions similar to Qualcomm’s Neural Processing SDK or Arm’s Compute Library.

This highlights a big ache level for the cell AI growth setting. Whereas desktop growth has principally settled on CUDA (although AMD’s ROCm is gaining traction), smartphones run a wide range of NPU architectures. There’s Google’s proprietary Tensor, Snapdragon Hexagon, Apple’s Neural Engine, and extra, every with its personal capabilities and growth platforms.

Smartphone chipsets that boast NPU capabilities (which is actually all of them) are constructed to resolve one drawback — supporting smaller information values, complicated math, and difficult reminiscence patterns in an environment friendly method with out having to retool GPU architectures. Nonetheless, discrete NPUs introduce new challenges, particularly with regards to third-party growth.

Whereas APIs and SDKs can be found for Apple, Snapdragon, and MediaTek chips, builders historically needed to construct and optimize their functions individually for every platform. Even Google doesn’t but present simple, basic developer entry for its AI showcase Pixels: the Tensor ML SDK stays in experimental entry, with no assure of basic launch. Builders can experiment with higher-level Gemini Nano options by way of Google’s ML Package, however that stops properly in need of true, low-level entry to the underlying {hardware}.

Worse, Samsung withdrew help for its Neural SDK altogether, and Google’s extra common Android NNAPI has since been deprecated. The result’s a labyrinth of specs and deserted APIs that make environment friendly third-party cell AI growth exceedingly tough. Vendor-specific optimizations have been by no means going to scale, leaving us caught with cloud-based and in-house compact fashions managed by a number of main distributors, similar to Google.

Fortunately, Google launched LiteRT in 2024 — successfully repositioning TensorFlow Lite — as a single on-device runtime that helps CPU, GPU, and vendor NPUs (at present Qualcomm and MediaTek). It was particularly designed to maximise {hardware} acceleration at runtime, leaving the software program to decide on essentially the most appropriate technique, addressing NNAPI’s greatest flaw. Whereas NNAPI was supposed to summary away vendor-specific {hardware}, it in the end standardized the interface fairly than the conduct, leaving efficiency and reliability to vendor drivers — a niche LiteRT makes an attempt to shut by proudly owning the runtime itself.

Curiously, LiteRT is designed to run inference completely on-device throughout Android, iOS, embedded techniques, and even desktop-class environments, signaling Google’s ambition to make it a really cross-platform runtime for compact fashions. Nonetheless, not like desktop AI frameworks or diffusion pipelines that expose dozens of runtime tuning parameters, a TensorFlow Lite mannequin represents a totally specified mannequin, with precision, quantization, and execution constraints determined forward of time so it will probably run predictably on constrained cell {hardware}.

Whereas abstracting away the vendor-NPU drawback is a serious perk of LiteRT, it’s nonetheless price contemplating whether or not NPUs will stay as central as they as soon as have been in gentle of different fashionable developments.

As an example, Arm’s new SME2 exterior extension for its newest C1 collection of CPUs offers as much as 4x CPU-side AI acceleration for some workloads, with extensive framework help and no want for devoted SDKs. It’s additionally attainable that cell GPU architectures will shift to raised help superior machine studying workloads, probably lowering the necessity for devoted NPUs altogether. Samsung is reportedly exploring its personal GPU structure particularly to raised leverage on-device AI, which might debut as early because the Galaxy S28 collection. Likewise, Immagination’s E-series is particularly constructed for AI acceleration, debuting help for FP8 and INT8. Possibly Pixel will undertake this chip, finally.

LiteRT enhances these developments, liberating builders to fret much less about precisely how the {hardware} market shakes out. The advance of complicated instruction help on CPUs could make them more and more environment friendly instruments for working machine studying workloads fairly than a fallback. In the meantime, GPUs with superior quantization help may finally transfer to turn out to be the default accelerators as a substitute of NPUs, and LiteRT can deal with the transition. That makes LiteRT really feel nearer to the mobile-side equal of CUDA we’ve been lacking — not as a result of it exposes {hardware}, however as a result of it lastly abstracts it correctly.

Devoted cell NPUs are unlikely to vanish any time quickly, however the NPU-centric, vendor-locked method that outlined the primary wave of on-device AI clearly isn’t the endgame. For many third-party functions, CPUs and GPUs will proceed to shoulder a lot of the sensible workload, notably as they achieve extra environment friendly help for contemporary machine studying operations. What issues greater than any single block of silicon is the software program layer that decides how — and if — that {hardware} is used.

If LiteRT succeeds, NPUs turn out to be accelerators fairly than gatekeepers, and on-device cell AI lastly turns into one thing builders can goal with out betting on a selected chip vendor’s roadmap. With that in thoughts, there’s most likely nonetheless some option to go earlier than on-device AI has a vibrant ecosystem of third-party options to get pleasure from, however we’re lastly inching somewhat bit nearer.