Pimchip provides dedicated acceleration cores for machine learning and artificial intelligence, specially optimized for the large number of neural network models involved in AI computing. It can process machine learning algorithms and deep learning models such as artificial neural networks with higher efficiency and lower energy consumption, facilitating the implementation of AI in multiple scenarios.
The N300 provides computility of 0.5 TOPS in single-core and a Cluster architecture composed of multiple NPU cores.
The N300 supports mixed-precision computing, including integer computation at 4-bit, 8-bit, and floating-point computation at 16-bit, which enables a better balance between energy consumption, computility density, and computation accuracy.
The high-speed task scheduling acceleration unit supports real-time task scheduling for multi-core or multiple computing units.
The N300 supports custom operators to meet the deployment needs of various models, and provides configuration schemes and specific optimizations for application scenarios such as voice activity detection, eye-tracking, active noise control, and environmental perception.
To meet customers' needs for more autonomous and flexible algorithm migration, the N300 have opened up NPU intermediate representation layer specifications, model parsers, model optimizers, drivers, etc., and provided a free software toolchain, including software simulators, debuggers, and C compilers.
MobileNet, ResNet, Yolo-v2/v3, UNet, ShuffleNet, SqueezeNet, EfficientNet, LSTM, and networks that can be decomposed into combinations of the above operators.
|
Operator name |
Parameter |
|---|---|
Conv |
Kernel size: Any shape <=7x7 Stride: 1,2,3 Zero padding: 0~kernel size-1 |
Depthwise |
Kernel size:Any shape <=5x5 Stride: 1,2,3 Zero padding: 0~kernel size-1 |
Transposed convolution |
Kernel size: Any shape <=7x7 Stride: 1,2,3 Zero padding: kernel size-1 |
FC |
Channel size: 1~2048 |
Pooling |
Kernel size:1~15 Type: max, avg, min |
Concat |
Only supports channel direction |
|
Binary operator |
Arithmetic operation：add,sub, mul, div, mod, avg Logical operation：or,and Comparison operation：min,max,cmp,logic,sel,compsel |
|
Unary operator |
Log,exp,abs,not,sqrt,tanh |
Reorg |
HWC<->CHW |
Upsampling |
Nearest interpolation, Bilinear interpolation |
|
Nonlinear |
Relu,p-relu,leaky-relu,relu6,softmax |
MobileNet, ResNet, Yolo-v2/v3, UNet, ShuffleNet, SqueezeNet, EfficientNet, LSTM, and networks that can be decomposed into combinations of the above operators.
