Today General Processor Technologies, China’s licensor of customizable processor IP cores, announced two innovative new additions to its IP portfolio: the ‘GPT’ AI accelerator and the VLVm1 Digital Signal Processor (DSP) for digital and image processing. With the new cores, designers can reduce chip cost, area, and power consumption in SoCs targeting applications across autonomous vehicles and intelligent transportation, smart cities, industrial automation, robotics, and machine vision.
The configurable GPT AI accelerator is designed for Convolutional Neural Network (CNN) inference. The highly-optimized design provides for convolution, pooling, dropout, padding, and programmable activation functions. Running at up to 1GHz in 28nm technology, each CNN building block contains 288 MAC units, scalable to 8+ units. Performance, power, and accuracy can be optimized using built-in integer or 16-bit floating point operations. GPT provides a library of popular CNN networks; support for frameworks such as Tensorflow and Caffe are in development.
The VLVm1 is the first of GPT’s optimized Variable Length Vector (VLV) DSP execution unit targeted for digital and image processing. VLVs allow forward and backward code compatibility by architecturally specifying a vector length. Unlike the SIMD architecture which has a fixed width, VLV specifies an arbitrary and dynamic length for vectors. An implementation can be customized for performance/power/price based on the number of vector elements to be executed simultaneously. Code does not need to be recompiled when the vector length changes. Non-power-of-two vectors can be easily specified.
Combining GPT’s superscalar out-of-order ‘Unity’ CPU with a VLV execution unit, the VLVm1 processors run at up to 3GHz. All instructions are part of the Unity 1.0 Instruction Set Architecture (ISA) which provides a common programming view of heterogeneous processors through its ability to efficiently map Heterogeneous System Architecture Intermediate Language (HSAIL) instructions.
The new cores add to GPT’s IP portfolio, which also includes the high-performance ‘Unity’ CPU that has been validated in TSMC’s 28nm process and is configurable for multiple performance levels.
“GPT is a leading provider of processor IP to the China market,” said Dr. John Glossner, CEO of GPT. “We have processor blocks needed to create a heterogeneous system for a wide range of applications, including CPUs, CNN/AI accelerators, communications processors, DSPs and soon a GPU that will extend the Unity instruction set to support all major types of compute capability.”
All GPT IP cores are optimized for Heterogeneous Systems Architecture (HSA), providing a common programming view and object code compatibility. The GPT AI accelerator and VLVm1 DSP will be validated in silicon by the third quarter of 2018, implemented in the GP8300 SoC from Optimum Semiconductor.
Leave a Reply