NPU IP for Data Center and Automotive

VeriSilicon’s Vivante VIP9400 processing family offers programmable, scalable and extendable solutions for markets that demand real time and powerful AI devices. VIP9400 Series’ patented Neural Network engine and Tensor Processing Fabric deliver superb neural network inference performance with industry-leading power efficiency (TOPS/W) and area efficiency (mm2/W). The VIP9400’s scalable architecture can provide 80TOPS computing ability, enables AI for data center and automotive application.
 

In addition to neural network acceleration, VIP9400 Series are equipped with Parallel Processing Units (PPUs), which provide full programmability along with conformance to OpenCL 1.2 and OpenVX 1.2.

Programmable Engines (PPU) & NN Engine

• 128-bit vector processing unit (shader + ext)
• OpenCL 1.2 shader instruction set
• Enhanced vision instruction set (EVIS)
• INT 8/16/32b, Float 16/32b in PPU

Convolution layers

• INT8 with INT16, Float16, or BFloat16 variant in NN
• Tensor Processing Fabric
• Non-convolution layers
• Multi-lane processing for data shuffling, normalization, pooling/unpooling, LUT, etc.
• Network pruning support, zero skipping, compression
• On-chip SRAM for DDR BW saving
• Accepts INT 8/16b, BFloat16, Float16

Unified Programming Model

• OpenCL, OpenVX
• Parallel processing between PPU and NN HW accelerators with priority setting
• Supports popular vision and deep learning frameworks: OpenCV, Caffe, Caffe2, TensorFlow, TensorFlowLite, ONNX, PyTorch, MXnet, Cognitive Toolkit, PaddlePaddle, Keras

SW & Tools

• ACUITY: End-to-end Neural Network development tool
• Eclipse-based IDE for coding/debugging/Profiling
• Linux and Android NN API runtime support
• Task specific engines speed up for commonly used AI apps

Scalability

• Number of PPU and NN cores can be configured independently
• Same OpenVX/OpenCL code runs on all processor variants; scalable performance

Extendibility

• FLEXA API: Easy integration with other VSI IPs
• Reconfigurable EVIS for user to define own instructions
• VIP-ConnectTM: HW and SW I/F protocol to plug in customer HW accelerators and expose functionality via CL/VX custom kernels