Arm.com

Android's next generation of flagship chips now has its two competing answers to the same question: where inside a mobile GPU does AI hardware belong? Six days after Qualcomm disclosed its Adreno Neural Fusion architecture — Matrix Cores embedded per GPU slice alongside an 18MB high-bandwidth memory pool — Arm unveiled the Mali G2-Ultra NX on September 8, 2026, placing dedicated neural accelerators inside the shader cores themselves. The CSS for Mobile 2 platform announcement confirmed the two approaches share the same premise but execute it differently, and no independent benchmark has yet confirmed which delivers better real-world results. That test arrives when shipping devices can be compared — and MediaTek is expected to be the first to bring CSS for Mobile 2 silicon to market within weeks.

The practical stakes for any Android flagship buyer in the next product cycle: both Qualcomm and Arm are betting that AI-accelerated rendering — upscaling a lower-resolution render to full display resolution, and generating synthetic frames to smooth frame rates — will define what premium Android gaming looks and feels like. The Mali G2-Ultra NX is Arm's implementation of that bet, and it is now the direct competitor to Qualcomm's Adreno Neural Fusion for the non-Snapdragon Android market.

CSS, short for Compute Subsystem, is the platform Arm sells to chipmakers who want a pre-validated, pre-integrated reference design instead of licensing CPU cores, a GPU, and a system interconnect separately. The first CSS for Mobile debuted last year alongside the C1-Ultra CPU and Mali G1-Ultra GPU under the short-lived Lumex branding — Arm abandoned the Lumex name quietly after just one generation. CSS for Mobile 2 pairs the new C2-Ultra and C2-Pro CPU cores with the Mali G2-Ultra NX GPU, an updated SI L2 system interconnect, and a developer software stack built around the KleidiAI library and Arm's AI Portal.

What CSS for Mobile 2 Is and Why It Matters

The platform is modular: a chipmaker adopting CSS for Mobile 2 can use individual IP blocks, combine them with third-party designs, or configure the cluster for its own product requirements. The example flagship configuration Arm published pairs two C2-Ultra performance cores with six C2-Pro efficiency cores and two SME2 units — a layout that closely matches expectations for the Dimensity 9600 Pro chip from MediaTek. Samsung could substitute the Mali GPU for its in-house Radeon-derived design in the Exynos 2700, as it has done in prior generations. Google could do the same with Imagination Technologies IP. The CSS architecture is a reference, not a mandate.

Two Philosophies for Putting AI Inside a Mobile GPU

The central architectural decision in both Arm's and Qualcomm's neural GPU designs is not whether to add AI hardware to the graphics pipeline — both companies have committed to that direction — but where to place it and what memory it shares with the rendering workload.

Arm's Mali G2-Ultra NX puts dedicated neural accelerators inside each shader core. Those accelerators share existing caches and control structures with conventional graphics workloads. The result, Arm argues, is that neural inference workloads run "where the graphics already live" — reading from and writing to the same on-chip tile memory that the GPU uses for rendering, without crossing the chip's system interconnect bus to reach a separate processor.

Qualcomm's Adreno Neural Fusion takes a different path. The Matrix Cores sit inside each of the GPU's three processing slices as distinct compute units, alongside an 18MB Adreno High Performance Memory cache dedicated to keeping frame buffers and compute data local to the GPU subsystem. A full description of Adreno Neural Fusion architecture is available in TechTimes' prior Qualcomm coverage. The HPM eliminates most DRAM round-trips for upscaling data, but the inference hardware is at the slice level rather than at the per-shader-core level that Arm has chosen.

Both designs share the same underlying logic: the reason prior mobile AI rendering approaches were inefficient is that routing AI workloads from the GPU to a separate NPU requires writing data out to shared memory or DRAM, crossing the chip's system bus, waiting for inference to complete, and reading results back. Each of those steps adds latency and energy cost. Both companies are eliminating those bus crossings — they differ on whether the most efficient location to do so is inside each individual shader core (Arm) or at the level of a dedicated slice-level unit alongside a large local cache (Qualcomm).

Which approach delivers better results in practice — better output quality per watt, better frame generation responsiveness, better sustained gaming performance under real thermal conditions — will not be determinable until independent third-party benchmarks run on shipping devices. Arm's G2-Ultra NX performance claims compare its hardware to its own prior generation. Qualcomm's 40% power savings claim compares its own implementation to its own prior Snapdragon Game Super Resolution system. Neither set of numbers has yet been measured by an independent reviewer with access to production silicon.

What Arm's Shader-Core Approach Actually Does

The Mali G2-Ultra NX uses this neural hardware to power three distinct graphics technologies. Neural Super Sampling, or NSS, reconstructs a higher-resolution output image from a lower-resolution render — the mobile equivalent of what NVIDIA's DLSS has done on PC GPUs since 2018, but running on dedicated in-shader hardware rather than general shader ALUs. NSS can produce 1080p output from 540p renders at 4ms per frame, according to Arm's published figures.

Neural Frame Rate Upscaling, or NFRU, generates intermediate frames from motion vectors and scene data, turning 30 FPS into 60 FPS output, or a 60 FPS render into 120 FPS output, without rendering each displayed frame natively. This is the technique that most directly affects perceived smoothness in games, and it is also the technique that carries the most important caveat: synthesized frames are computed after the rendered frames they sit between, which means a player's most recent inputs are not reflected in a synthesized frame. The delay from input to displayed output is higher with NFRU active than without it, by approximately one rendered-frame interval. For narrative games, action RPGs, and casual titles, this is imperceptible. For competitive mobile gaming — battle royale shooters, real-time fighting games — it is a real tradeoff, identical in character to the same limitation documented in PC frame generation implementations.

Neural Super Sampling and Denoising, or NSSD, combines upscaling with ray reconstruction and denoising for ray-traced scenes — the most demanding use case, targeting flagship visuals that approach desktop-class quality on a mobile power budget.

In Arm's Neural Dawn demo, built with Sumo Digital on Unreal Engine, a combination of NFRU and NSSD delivered 4x higher efficiency and reduced external memory traffic by as much as 70% compared with native rendering. Arm argues this headroom could make Unreal Engine 5 features such as MegaLights practical on a smartphone. The demo is scheduled for release later in 2026 on Android devices powered by the upcoming Mali hardware. Real games integrating the technology include NetEase's Where Winds Meet with NSS support and an NSSD demonstration in Tencent Games' Arena Breakout Infinite, alongside NSS integration in Infold Games' Infinity Nikki.

The neural accelerators support INT8 and INT16 precision — the numerical formats most common in deployed neural network inference, which require less memory and less computation than floating-point training-precision formats.

How Does Arm's Neural GPU Differ From Qualcomm's Adreno Neural Fusion?

The architectural distinction matters because mobile GPUs use tile-based deferred rendering, a technique that makes the on-chip tile memory the performance-critical resource for graphics work. Instead of holding the entire framebuffer in memory simultaneously (as desktop immediate-mode GPUs do), mobile GPUs subdivide the screen into small tiles and complete all rendering for each tile inside fast on-chip memory before writing the finished result to system RAM. This approach reduces memory bandwidth — the dominant power and latency bottleneck on battery-powered devices — by keeping as much work as possible on-chip.

Arm's neural accelerators share this on-chip tile memory with the shader workloads running in the same shader core. The inference work and the graphics work are physically co-located at the memory level. Qualcomm's Matrix Cores share the GPU slice's 18MB HPM with the rendering pipeline — also on-chip, also designed to minimize DRAM traffic — but they are separate compute units within the slice, not embedded inside the shader ALU pipeline itself.

Whether co-location at the shader-core level (Arm) produces measurably lower latency and higher efficiency than co-location at the slice level with dedicated HPM (Qualcomm) is the empirical question that reviewers will answer once silicon ships. Anshel Sag, a principal analyst at Moor Insights & Strategy, characterized the Qualcomm Adreno Matrix Core approach as "similar to what Apple did with Apple Silicon's GPU" — noting that Apple has used GPU-embedded matrix compute units across its A-series and M-series silicon as dedicated GEMM accelerators alongside general shader hardware. Both Arm and Qualcomm have now adopted a version of this architectural pattern.

C2-Ultra and C2-Pro: The CPU Cluster Powering Agentic AI

The C2-Ultra is Arm's new high-performance CPU core, accompanied by the C2-Pro as the mid-tier companion. On raw performance, Arm claims the cluster delivers up to 15% higher single-thread performance and 12% higher multi-thread performance versus the prior generation, translating to approximately 15% faster web browsing and 12% faster application launch times.

The more substantial changes arrive in AI capability. The reference CSS for Mobile 2 configuration includes two SME2 (Scalable Matrix Extension 2) units — double the SME2 capability of the prior-generation C1-Ultra configuration. SME2 is an Arm matrix compute extension that enables the processor to run GEMM (General Matrix Multiply) operations natively — the same mathematical primitive that underlies neural network inference — without routing every AI task to a dedicated NPU. With doubled SME2 capability, Arm reports the new CPU cluster can deliver up to 1.7x AI performance across its tested models and achieves a 70% speedup on the latest Small Language Models compared with the prior cluster.

A two-C2-Ultra, two-SME2 configuration finished agentic workflows 24% faster than the equivalent C1-Ultra setup in Arm's testing. Arm is framing this additional headroom around the tasks that are quietly becoming the most demanding part of a modern smartphone workload: local speech processing, on-device retrieval, reasoning chains, and the orchestration layer required for agentic applications — agents that maintain context, run applications, coordinate models and services, and act on a user's behalf, all within a smartphone's power budget.

The C2-Ultra also uses 38% less power at equivalent performance levels compared with the C1-Ultra, according to Arm's published figures — a number that matters for sustained, thermally limited workloads.

Underpinning the cluster is Arm's updated SI L2 system interconnect, which provides coherency, quality-of-service controls, and lower-latency, higher-bandwidth links between the platform's heterogeneous compute blocks.

Ray Tracing Advances: Opacity Micromaps Come to Mobile

The G2-Ultra NX's Ray Tracing Unit is now in its third generation, and the headline addition is hardware support for Opacity Micromaps — a technique that has improved ray tracing performance significantly on PC GPUs and now arrives on mobile for the first time.

The problem Opacity Micromaps solve: when a ray tracing algorithm encounters complex transparent geometry — foliage, fences, hair, chain-link, alpha-tested surfaces — it traditionally has to launch an "any-hit" shader to evaluate each intersection point's material transparency at runtime. These shader invocations are expensive, particularly in outdoor scenes dense with vegetation. Opacity Micromaps pre-compute transparency data during scene setup by subdividing each triangle into a grid of micro-triangles and classifying each as permanently opaque, permanently transparent, or unknown. At ray traversal time, the ray tracing hardware reads this pre-computed classification and handles opaque and transparent intersections directly — without launching a shader invocation at all for the common cases. Only the genuinely ambiguous micro-triangles trigger shader evaluation.

Arm says third-generation RTU Opacity Micromaps reduce ray tracing workload overhead by up to 70% versus the prior generation on scene types that benefit from the technique. On PC hardware, NVIDIA's implementation in Indiana Jones: The Great Circle reduced GPU time spent ray tracing alpha-tested vegetation from 7.90ms to 3.58ms — a 55% reduction on a GeForce RTX 5080, according to NVIDIA's published analysis. The scale of improvement on mobile will depend on scene complexity and implementation, but the architectural capability now exists at the hardware level for the first time.

Alongside Opacity Micromaps, Arm's execution engine overhaul — described as Mali's biggest ISA upgrade in seven generations — doubles the number of registers available per warp. Registers are the fastest storage on a GPU — on-chip, directly accessible by shader execution units. When a shader exhausts its register allocation, it "spills" data to slower off-chip memory, degrading throughput. More registers per warp means complex shaders — ray tracing, advanced lighting, post-processing passes — spend less time stalling on spilled data. Arm cites 24% higher benchmark performance and 14% better non-AI gaming performance versus the Mali G1-Ultra as the composite result of these changes.

When Will Phones With Mali G2-Ultra NX Arrive?

Arm says devices built on the Mali G2-Ultra NX platform are expected to start appearing within weeks. MediaTek is widely anticipated to be first to market, with the Dimensity 9600 as the most probable vehicle — benchmark leaks from Notebookcheck show the Dimensity 9600 Pro on Geekbench on a Vivo X500 Pro Max test device, with prime cores clocked at 4.55 GHz.

Samsung's Exynos 2700 — which has appeared in leaked Exynos 2700 die shots ahead of its expected use in the Galaxy S27 series — is not anticipated until early 2028. Samsung may substitute its own Radeon-derived GPU for the Mali in the Exynos 2700, as it has done previously, which would mean CSS for Mobile 2's GPU would not necessarily appear in Samsung's flagship line even when the CPU cluster does.

Qualcomm's competing Snapdragon platform carrying Adreno Neural Fusion is expected to be officially unveiled at Snapdragon Summit 2026, scheduled for September 22 to 24 in Maui, Hawaii — two weeks after today's Arm announcement. Commercial devices based on that chip are widely anticipated to follow in the second half of 2026 and into early 2027. Apple's A20 Pro, expected alongside the iPhone 18 Pro and a rumored iPhone Ultra, will not use Arm's Mali GPU but will compete in the broader premium smartphone market that these announcements address.

What Still Needs to Be Proven

Every performance figure in this article comes from Arm's own testing against Arm's own prior generation, or from Qualcomm's own testing against Qualcomm's own prior implementation. No independent reviewer has run tests on production G2-Ultra NX hardware, because that hardware has not yet shipped in a consumer device. The 4x efficiency gain, the 70% memory traffic reduction, the 24% benchmark improvement, the 14% gaming uplift, the 1.7x AI performance gain — these are all Arm's claims, published by Arm.

The same applies to Qualcomm's 40% power savings claim for Adreno Neural Fusion, which compares Qualcomm's new implementation against Qualcomm's prior Snapdragon Game Super Resolution system. Until a third-party reviewer can run identical workloads on both platforms — a comparison that is not possible until MediaTek's devices and Qualcomm's devices both ship — the competitive claim at the center of this article remains architecturally argued rather than empirically settled.

What this announcement establishes is the architectural battle lines: Arm has chosen shader-core-level co-location, shared coherent cache, and the GPU's existing memory hierarchy as the foundation for neural graphics. Qualcomm has chosen slice-level dedicated Matrix Cores, a large local HPM cache, and Day 1 engine integration as its foundation. Both are real architectural commitments, both are shipping within months, and both will be tested by independent reviewers when devices arrive.

Frequently Asked Questions

What is the Mali G2-Ultra NX and why is it significant?

The Mali G2-Ultra NX is the first Arm GPU to embed dedicated neural accelerators directly inside its shader cores, rather than routing AI inference to a separate NPU or running it on general shader hardware. This makes it — alongside Qualcomm's Adreno Neural Fusion — one of the first mobile GPU architectures to treat neural rendering as a hardware-native capability rather than a software feature. The practical consequence is that upscaling, frame generation, and AI-accelerated rendering can run within the graphics pipeline itself, at lower power and with less data movement overhead than prior approaches. The CSS for Mobile 2 launch announcement confirms Arm introduced the GPU on September 8, 2026, as part of its CSS for Mobile 2 compute platform.

How does Arm's approach differ from Qualcomm's Adreno Neural Fusion?

Both Arm and Qualcomm have embedded AI hardware inside the mobile GPU rather than relying on a separate NPU — but they have done so at different levels of the chip's architecture. Arm's neural accelerators sit inside each individual shader core and share that core's existing memory, cache, and control structures with conventional graphics work. Qualcomm's Matrix Cores are dedicated compute units within each of the GPU's three processing slices, alongside an 18MB high-performance cache (HPM) that keeps frame buffers and AI compute data local to the GPU subsystem. Both approaches reduce the data movement overhead that made prior GPU-to-NPU AI routing inefficient; which delivers better real-world results at equivalent power levels is a question only independent benchmark testing of shipping devices can answer. For more on Qualcomm Adreno Neural Fusion, see TechTimes' September 3, 2026 coverage.

Is AI frame generation good for mobile gaming performance?

For most mobile gaming use cases — narrative games, action RPGs, open-world titles, casual games — frame generation delivers noticeably smoother visual output with lower power draw, because the GPU renders fewer frames natively while the AI synthesizes the frames displayed in between. For competitive mobile gaming — battle royale shooters, fighting games, any title where millisecond-level input response affects outcomes — frame generation adds display latency, because synthesized frames are computed after the rendered frames they follow, without incorporating the player's most recent inputs. This frame generation latency tradeoff is consistent across PC and mobile implementations and is not specific to Arm or Qualcomm's design; both NFRU and Adreno Neural Fusion's frame generation carry the same underlying physics.

When can I buy a phone with Mali G2-Ultra NX?

Arm has confirmed that devices built on CSS for Mobile 2 — which includes the Mali G2-Ultra NX — are expected to begin shipping within weeks of the September 8, 2026 announcement. MediaTek is widely anticipated to be the first chipmaker to ship a CSS for Mobile 2-based SoC, most likely in the form of the Dimensity 9600, which has already appeared in early benchmark runs. Samsung's Exynos 2700, expected to use a CSS for Mobile 2 CPU cluster, is not anticipated until early 2028. Specific device announcements from phone manufacturers will follow as chipmakers finalize their implementations. See Notebookcheck's CSS for Mobile 2 coverage for the latest device timing details.