Aisin Develops 16nm STT-MRAM AI SoC for Object Detection
爱信开发全球首款 16nm STT-MRAM AI SoC，实现 50 倍目标检测能效
On-chip MRAM removes the DRAM/SSD bottleneck that normally limits edge AI boot times and power budgets, pointing to always-on vision systems in cars and IoT devices. Automotive engineers and embedded developers designing low-power perception pipelines can use this as a reference architecture for MRAM-based inference. As a Japanese automotive supplier, Aisin's work sits alongside efforts by ASEAN auto manufacturing hubs like Thailand that depend on such component suppliers for next-generation vehicle electronics.
Aisin, a Toyota-affiliated auto parts supplier, announced on 2 Nov with Tohoku University the world's first AI SoC in 16nm process with embedded STT-MRAM. The chip integrates 32MB of on-chip MRAM under a near-memory computing design, storing AI model weights and OS boot data locally without external DRAM or SSD. Using an 8/16-bit hybrid multiply architecture and weight sharing, it runs Tiny YOLOv3 object detection at 50.7mJ per frame, about 1/50 the energy of conventional solutions. Boot-to-OS time drops to 65.7ms, a 38.59x speedup, while AI processing speeds up 10.16x and overall runtime 25.2x.
爱信开发全球首款 16nm STT-MRAM AI SoC，实现 50 倍目标检测能效
IT之家 11 月 3 日消息，丰田系汽车零组件供应商 Aisin（爱信）当地时间昨日宣布与日本东北大学合作成功开发了 全球首款集成 STT-MRAM （IT之家注：自旋转移矩-磁随机存储器） 的 16nm 工艺 AI SoC 。 这一芯片 采用近内存计算 (NMC) 设计概念 ，直接集成 32MB 嵌入式 MRAM，可直接在本地存储 AI 模型权重和操作系统启动数据，开机时无需调用外部 DRAM 缓存或 SSD 存储。 该 SoC 通过 8/16 位混合乘法架构和权重共享技术，利用 Tiny YOLOv3 在有限的片上 STT-MRAM 容量内实现了目标检测 。 其从开机到操作系统启动的时间缩短至 65.7ms，速度提升 38.59 倍；AI 处理速度提升 10.16 倍；整体运行速度提升 25.2 倍。其进行目标检测所需的能量也仅有 50.7mJ / 帧， 是传统解决方案的 1/50 。
