At the Arm Everywhere China annual event on September 8, 2026, Arm announced its second-generation mobile compute subsystem (CSS for Mobile 2), featuring the Mali G2-Ultra NX — the first Mali GPU to integrate a dedicated neural accelerator directly into the shader core. Three headline figures from Arm cover bandwidth rather than raw compute, reflecting the reality that on mobile the most expensive resource is not arithmetic but the cost of moving data between chip and memory.
One Frame, 33 Milliseconds
At 30 FPS, a frame's budget is 33 ms; at 120 FPS it shrinks to 8.3 ms. Within that window the GPU must handle geometry, rasterization, shading, post-processing, and memory round-trips. Mobile adds hard constraints: power and thermals. Arm's positioning is that traditional rendering can achieve high image quality, but reaching desktop-level fidelity within mobile power, thermal, and bandwidth limits requires higher efficiency — and neural graphics offers another path: using AI to reconstruct detail, generate intermediate frames, and optimize images while reducing GPU and system workload.
Mali G2-Ultra NX tightly couples the neural accelerator into the shader core so neural graphics workloads run alongside regular graphics and compute. This tight integration reuses the GPU's memory system, coherent caches, and control structures, reducing data movement through the pipeline. The accelerator block supports INT8 and INT16 processing and includes optical flow acceleration.
Three Technologies, Three Stages
- NSS (Neural Super Sampling): reconstructs high-resolution frames from low-resolution renders, reducing conventional rendering workload; combines temporal data with built-in anti-aliasing.
- NFRU (Neural Frame Rate Upscaling): generates intermediate frames using motion, depth, and rendered-frame data. Arm engineers said in media Q&A that NFRU currently inserts a single AI frame between two rendered frames; multi-frame generation is planned, but latency control is the priority — fitting both the original and AI frames into a 33 ms budget requires developer-supported frame pacing.
- NSSD (Neural Super Sampling + Denoising): combines neural upscaling and denoising for ray-traced scenes with complex lighting and shadows.
- Tencent Games: integrating Arm Neural Technology into MagicDawn, co-developing an NSSD tech demo with *Arena Breakout: Infinite*
- Unity China: bringing the integration to the Unity (Tuanjie) engine developer workflow in October
- NetEase: the Messiah engine behind *Where Winds Meet* is integrating; engine lead Zhou Kaili says the title will be among the first to bring Arm neural tech to players
- Papergames: *Infinity Nikki* adopts NSS
- Xiaomi: the Xring O3 SoC will use Mali G2-Ultra NX
In the demo game *Neural Dawn*, built with Sumo Digital, NFRU and NSSD delivered up to 4x performance efficiency and up to 70% lower external memory traffic versus native rendering. Arm says these gains make technologies like Unreal Engine MegaLights viable on mobile hardware.
Bandwidth Is the Main Battlefield
Arm's James McNiven described three layers of bandwidth optimization:
| Layer | Claim | |---|---| | 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, up to 70% RT workload reduction in Arm demos | | Neural Dawn demo | Up to 70% lower external memory traffic, up to 4x performance efficiency |
Opacity Micromaps hardware targets transparent geometry — vegetation, fabric, layered surfaces — historically a costly category in ray tracing.
Desktop Already Walked This Road; Mobile's Constraints Differ
Neural super sampling is not new on desktop, where GPUs hit the compute wall first and power/thermal headroom is generous. Phones hit the bandwidth and power walls first: reading memory once costs far more energy than an extra arithmetic operation, and battery and skin temperature offer no negotiation room. That explains why Arm's headline figure is a 70% reduction in external memory traffic rather than doubled frame rates.
Traditional Pipeline Changes Too
Beyond the neural blocks, Mali G2-Ultra NX features a new execution engine that Arm calls the largest Mali GPU ISA upgrade in seven generations, with per-warp registers up to 2x. Compared to the previous generation: baseline performance up to +24%, non-AI game performance up to +14%, and with NFRU, support for up to 120 FPS mobile gaming. The third-generation ray tracing unit supports more complex lighting, shadows, reflections, and geometry. Product page footnotes note comparisons are against the previous generation, based on internal testing from March to August 2026.
Beyond the GPU: The Whole Subsystem
Mali G2-Ultra NX is part of CSS for Mobile 2. The platform's C2 CPU cluster supports SME2, running lightweight AI models up to 1.7x faster than the previous generation, plus SI L2 interconnect and system IP. Arm says the combination is co-optimized for agentic AI workloads characterized by latency sensitivity, burstiness, and memory intensity. The consistent direction: neural inference is spreading from a standalone accelerator into a layer distributed across CPU, GPU, and memory paths.
Ecosystem Laid Two Years Ago
Arm started Arm Neural Technology and an open neural graphics development kit two years before the GPU launch. The Arm Neural Graphics Development Kit provides NSS/NFRU models, Vulkan ML extensions, and plugins for major engines including Unreal Engine, plus an SDK for proprietary engines with profiling, optimization, training, and model optimization tools.
Adopters already announced:
What It Means in Practice
For mobile gaming, the most tangible implication is that dynamic global illumination previously considered off-limits on phones — such as UE MegaLights — can now support up to 1,400 dynamic lights without pre-baking, per Arm's media Q&A as relayed by Chinese media. Arm's senior edge AI product manager Deyan Lazarov framed AI as a complement and enhancement to traditional rendering, not a replacement, with hardware resources increasingly tilting toward neural acceleration while the traditional pipeline continues to evolve.
Caveats
All performance numbers come from Arm internal testing (March–August 2026) against the previous generation. Mali G2-Ultra NX is IP, not a shipping chip — real-world results await actual devices. Adoption is per-game with no universal switch retroactively enabling features on installed titles; Arm targets 3–6 month integration cycles for existing games and under 18 months for new titles — targets not yet validated. NFRU's single-frame insertion and latency-first design suggest the current limit is the time budget, not the algorithm; multi-frame generation has no confirmed timeline. Watch: Xring O3 device benchmarks, October's Unity (Tuanjie) integration quality, and responses from MediaTek and Qualcomm, both of which invest in AI accelerators — Arm's differentiator is unifying graphics and neural work in 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, benchmark windows) 3. Unity China announcement on Tuanjie engine integration (2026-09-09) 4. Tencent News coverage of Mali G2-Ultra NX (2026-09-08) 5. Arm and Sumo Digital Neural Dawn demo materials