Arm Mali G2-Ultra NX Puts Neural Accelerators Inside Shader Cores for AI-Native Mobile Graphics

On September 8, 2026, at the Arm Everywhere China annual event, Arm launched its second-generation mobile compute subsystem CSS for Mobile 2. Its centerpiece, the Mali G2-Ultra NX, is the first Mali GPU to integrate a dedicated neural network accelerator directly into the shader core, letting neural graphics workloads run alongside conventional graphics and compute workloads. This tight integration reuses the GPU's memory system, coherent caches, and control structures to reduce data movement across the pipeline. The accelerator block supports INT8 and INT16 processing with optical-flow acceleration.

Arm's framing: traditional rendering can deliver high image quality, but reaching desktop-class fidelity within mobile power, thermal, and bandwidth limits requires higher efficiency — neural graphics rebuilds detail, generates intermediate frames, and optimizes images while reducing GPU and system workload. Notably, two of Arm's three headline numbers are bandwidth figures, not compute figures — on mobile, the dominant cost is moving data between chip and memory, not compute speed.

Three Neural Techniques

• NSS (Neural Super Sampling): reconstructs high-resolution output from low-resolution renders, combining temporal data with built-in anti-aliasing to reduce conventional rendering workload.
• NFRU (Neural Frame Rate Up-scaling): generates intermediate frames using motion, depth, and rendered-frame data. Arm engineers noted NFRU currently inserts one AI frame between two rendered frames; multi-frame generation is planned, but latency control is the priority — within a 33 ms budget at 30 FPS, developers must cooperate on frame pacing.
• NSSD (Neural Super Sampling & Denoising): combines neural super-resolution with denoising for ray-traced scenes with complex lighting and shadows.

In the Arm × Sumo Digital demo game Neural Dawn, NFRU and NSSD delivered up to 4x performance efficiency and up to 70% lower external memory traffic versus native rendering — gains Arm says make techniques like Unreal Engine MegaLights viable on mobile.

Bandwidth Is the Main Battlefield

Arm's James McNiven described three layers of bandwidth optimization:

| Optimization layer | Figures | |---|---| | NSS adjustable modes | up to 70% bandwidth savings | | 3rd-gen ray tracing unit | up to 13% lower DRAM traffic on mainstream RT benchmarks | | Opacity Micromaps | +30% frame rate in Arm demos; up to 70% RT workload reduction | | Neural Dawn demo | up to 70% less external memory traffic; up to 4x performance efficiency |

Opacity Micromaps target transparent geometry (vegetation, fabric, layered surfaces) — historically a major ray-tracing cost.

Conventional Pipeline Upgrades

Beyond the neural side, Mali G2-Ultra NX features a new execution engine — Arm calls it the largest Mali GPU ISA upgrade in seven generations, with up to 2x registers per warp. Results vs. the previous generation: up to 24% baseline performance gain and up to 14% non-AI gaming performance gain. With NFRU, Arm claims support for up to 120 FPS mobile gaming sessions. Footnotes note comparisons are internal Arm tests from March–August 2026.

Beyond the GPU

• The C2 CPU cluster supports SME2, running lightweight AI models 1.7x faster than the previous generation, plus SI L2 interconnect and system IP — co-optimized for AI-agent workloads (latency-sensitive, bursty, memory-intensive).
• The Arm Neural Graphics Development Kit offers NSS/NFRU models, Vulkan ML extensions, Unreal Engine plugins, and SDK support for custom engines.

Ecosystem commitments: Tencent Games integrates Arm Neural Technology into MagicDawn and is co-developing an NSSD tech demo with *Arena Breakout: Infinite*; Unity China (Tuanjie Engine) ships native integration in October; NetEase's Messiah engine is integrating it; Papergames' *Infinity Nikki* adopts NSS. On silicon, Xiaomi confirmed its Xuanjie O3 chip uses Mali G2-Ultra NX.

Practical Implications

Per Arm's media briefing (as relayed by Chinese media), UE MegaLights-style dynamic global illumination — previously considered off-limits on mobile — can now support up to 1,400 dynamic lights without pre-baking. Arm senior edge-AI PM Deyan Lazarov positioned AI as a complement to, not replacement for, traditional rendering, with future hardware resources increasingly tilted toward neural acceleration.

Caveats

All performance figures come from Arm's internal tests (March–August 2026) against the previous generation. Mali G2-Ultra NX is IP, not shipping silicon — real-world results await end devices. Features must be enabled per-game; no universal retroactive switch exists. Arm targets 3–6 month integration for existing titles and under 18 months for new ones — unverified. NFRU's single-frame insertion shows the path is still constrained by frame-time budgets rather than algorithms; multi-frame generation timing is unspecified. Watch for: Xuanjie O3 device benchmarks, the October Tuanjie Engine delivery, and responses from MediaTek and Qualcomm, both of which invest in AI accelerators — Arm's differentiator is unifying graphics and neural matrix work within the shader cluster.

Sources

1. Arm Newsroom — "Inside the Arm Mali G2-Ultra NX GPU: Delivering desktop-class mobile gameplay with AI-native graphics" (2026-09-08, Deyan Lazarov) 2. Arm product pages and press release footnotes (INT8/INT16, optical flow, testing window) 3. Unity China announcement (2026-09-09) — October integration timeline 4. Tencent News coverage (2026-09-08) — NFRU details, bandwidth layers, MegaLights, Xiaomi Xuanjie O3 5. Arm × Sumo Digital Neural Dawn demo materials