ARM has announced its next generation of CPU and GPU cores, which will power the MediaTek Dimensity 9500 set to launch later this month. The company has introduced several changes this generation, starting with the naming itself.
ARM has retired its familiar Cortex-X and Cortex-A branding in favor of the new C1 family. At the top sits the C1-Ultra, followed by the C1-Premium (a smaller footprint version), the C1-Pro for mid-tier performance, and the efficiency-focused C1-Nano. All of these cores are based on the ARMv9.3 ISA architecture family.
The C1-Ultra is effectively the successor to the Cortex-X925, and ARM claims it delivers 12% IPC gain over last year’s core, scaling up to 25% with higher clock speeds on 3nm. Importantly, the C1-Ultra delivers this uplift while cutting power consumption by nearly 28% at equal performance levels.
Meanwhile, the C1-Pro replaces the Cortex-A725, which can achieve the same performance as its predecessor with 26% less power, or 11% more performance at the same power level.
Finally, the C1-Nano builds on the Cortex-A520 with a focus on branch prediction and cache improvements, yielding a 5–8% performance gain but a 26% efficiency improvement.
ARM C1 CPU Family – Key Specs
| Core | Successor to | Process | Clock (max) | IPC Gain | Power Efficiency | Notes |
| C1-Ultra | Cortex-X925 | 3nm | 4.1GHz | +12% | -28% @ same perf | High throughput, 2,000 in-flight instructions |
| C1-Premium | Cortex-X925* | 3nm | ~3.9GHz | Slightly lower | 35% smaller area | Cost-conscious version of Ultra |
| C1-Pro | Cortex-A725 | 3nm | ~3.5GHz | +11% | -26% @ same perf | Larger branch predictor, cache gains |
| C1-Nano | Cortex-A520 | 3nm | ~3.0GHz | +5–8% | +26% efficiency | Focused on low-power tasks |
SME2 instruction set
Perhaps the most notable CPU feature this year is SME2 (Scalable Matrix Extension 2). Unlike previous SIMD extensions such as NEON or SVE, SME2 operates as a shared external accelerator rather than being embedded inside each core. This modular design allows all cores to access SME2 units when needed, while shutting them down entirely to save power when idle.
ARM says SME2 handles AI tasks better with:
- 4.7× lower latency for speech recognition
- 2.8× faster Stable Audio generation
- 3.7× average gain across ML benchmarks
While SME2 won’t replace dedicated NPUs or GPUs for large models, it should enable smoother on-device AI inference for smaller models, particularly in scenarios where the NPU isn’t accessible.
Mali G1 GPU
On the graphics side, ARM is phasing out the Immortalis branding and introducing a new line of Mali GPUs. The first Mali G1 family has Ultra (10+ cores), Premium (6–9 cores), and Pro (1–5 cores) versions.
The flagship Mali G1-Ultra shows a 20% uplift in performance over last year’s Immortalis G925, along with 9% less energy per frame and up to 2× faster ray tracing thanks to dedicated hardware for BVH traversal. Improved cache and memory interconnect bandwidth further reduces bottlenecks and improves utilization.
That said, real-world ray tracing gains may be closer to 40%, as few mobile titles fully exploit the new hardware. Still, ARM’s decision to combine ray casting and intersection testing into a single structure should make the RTU more power-efficient than before.
Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!
For more daily updates, please visit our News Section.
(Via)
Comments