Arm Mali Update Adds Matrix Operations for Unified Edge AI
A 36Kr report says Arm updated the Mali GPU architecture with upgraded matrix-operation resources and unified scheduling for graphics and AI workloads across automotive, mobile and embedded devices. It also covers Imagination’s E-series GPU IP, which the company says delivers 4.4 times the Conv2D performance, 4.8 times the GEMM performance and 4.7 times the prefill performance of the preceding D-series. Imagination is also working to submit an optimized Llama.cpp backend upstream and plans native support for PyTorch and ONNX Runtime.
The report identifies unified scheduling and resource reuse across graphics, general computing and AI inference as a way to reduce system fragmentation and data movement in edge devices.
近日，Arm公布新一代Mali GPU架构更新，在图形管线内升级矩阵运算单元，强化端侧大模型推理与神经渲染协同，面向车载、移动及嵌入式终端推出统一软件调度方案。
这种把AI算力融入GPU、以统一硬件承载多类负载的方式，正成为行业共同选择。类似的融合尝试也在不同市场层次展开，从移动端到桌面端，再到IP授权市场，厂商路径各异，指向的却是同一个方向。
01 本地AI为何需要GPU
随着边缘端应用场景持续丰富，设备需要同时承载图形渲染、机器视觉、信号处理与大模型推理等多类任务。
以端侧大模型为例，从单一对话到多模态协同，从短上下文到长上下文，负载特征持续变化，对硬件灵活性提出更高要求。过去一段时间，芯片设计的主流思路是针对不同负载单独配置专用硬件模块，图形单元、AI加速器、视频编解码单元各自拥有独立的数据通路与软件栈。这种方案可以针对单一指标做到性能最优，但短板同样突出。多加速器并存会显著提升系统集成难度，硬件资源难以跨任务调度。一旦终端需求迭代，部分专用固定功能硬件就会出现利用率不足的问题，推高芯片整体成本与功耗。
#Arm#Imagination Technologies#Imagination E-series#Imagination D-series#Qwen 3.5 4B#Llama.cpp#PyTorch#ONNX Runtime#PowerVR SDK