The “Song” AI core is a highly configurable IP block that provides for efficient implementations of CNNs. Running at up to 1 GHz in 28-nm technology, each accelerator core contains 288 MAC units scalable to beyond 2 TFLOPS.
Performance, power, and size can be optimized efficiently supporting both low power embedded IoT devices, handsets, and edge processing designs. Accuracy can be optimized using built-in 16-bit integer or floating point operations. Support for compressed 8-bit data is provided. Extended precision intermediate results reduce errors in both regression and classification workloads.
Targeting inference applications, the AI accelerator supports multi-dimensional tensor processing. The highly optimized design provides direct support for convolution, pooling, dropout, padding, and programmable activation functions. A library of popular CNN networks is provided with generalized framework support (e.g., Tensorflow, Caffee, etc.) in development.
Earlier, hxGPT announced availability and licensing of the Changcheng out-of-order superscalar CPU. Validated in TSMC’s 28-nm process, the core is configurable providing for multiple performance levels based on customer requirements. Memories, caches, and out-of-order features (issue window, ROB, rename, etc.) can all be customized. Depending upon technology selection the core can run up to 3 GHz at very low power.
Building on hxGPT’s expertise in heterogeneous computing, the Song AI core easily integrates with hxGPT’s Changcheng CPU for ease of programming. Both are programmed using the same familiar processor programming environment.
See also: Google opens AI research center in China
See also: Bias in AI algorithms: How do we keep it out?
See also: AI, neural networks and the edge of the cloud
See also: AI brings indoor tracking, motion and vital sign detection to mesh routers