苹果A20 Pro芯片在端侧AI性能上取得显著提升。iPhone 18 Pro可在本地运行270亿参数的AI模型Bonsai-27B，其处理速度相比iPhone 17 Pro提高了2倍1。这一性能飞跃得益于芯片架构的创新升级。
A20 Pro芯片首次搭载双16核神经网络引擎1，其峰值吞吐性能超过7核GPU1。该芯片配备12GB、96位位宽的LPDDR5X内存1，内存带宽达到115.2GB/s1，为大规模模型在手机端的本地运行提供了充足的数据传输支持。测试表明，Bonsai初代采用1比特量化可以完整存入本地内存，而Bonsai 2采用2比特量化但体积过大，无法完整存入本地内存1。
Apple's latest A20 Pro chip marks a significant advancement in on-device artificial intelligence capabilities for the iPhone 18 Pro. The processor features dual 16-core neural engines for the first time, enabling the device to locally run the Bonsai-27B model—a 27 billion-parameter AI model—at twice the speed of its predecessor, the iPhone 17 Pro. 1
The performance gains are underpinned by substantial hardware improvements. The iPhone 18 Pro is equipped with 12GB of 96-bit LPDDR5X memory offering a bandwidth of 115.2GB/s, significantly enhancing the phone's capacity for local AI processing. 1 The dual 16-core neural engine delivers peak throughput performance exceeding that of a 7-core GPU, positioning it as a powerful on-device compute resource. 1 The Bonsai-27B model uses 1-bit quantization to fit within the device's local memory constraints, though the newer Bonsai 2 model, which employs 2-bit quantization, proves too large to be fully stored locally. 1
评论
还没有评论，欢迎留下第一条。