Machine Learning Accelerator
Seamless Deployment, Broad Network Support, Power Efficient
solution brief
Seamless Deployment:
- Same software environment as CPU/GPU
- No application software changes required
- No new training required
- U.2 Small Form Factor (SFF8639)
Broad Network Support:
- Works under Caffe, Caffe2, TensorFlow, MXNET
- Same CPU/GPU-based trained neural network with similar accuracy without any change
Power Efficient:
- Scalable and Lower Power than GPUs
- Ideal for Edge or Data Center Applications
Mipsology
| Network | Performance (Frames/Sec) |
| Caffenet | 1715 |
| Googlenet | 561 |
| Inception v3 | 143 |
| Inception v4 | 76 |
| Resnet50 | 239 |
| Resnet152 | 120 |
| Yolo v1 | 24 |
| Yolo v2 | 44 |
Seamless Deployment, Broad Network Support, Power Efficient
No longer does the CPU have to be the center of a system. Data centric solutions are quickly emerging to unlock the value in Big Data and Fast Data by using purpose-built architectures. One important technology which can extract insights from data is Machine Learning.
Western Digital partnered with Mipsology to provide a complete, easy to adopt and power efficient Machine Learning Accelerator based on the Zebra platform
Seamless Deployment
Broad Support
Neural Networks
- Tested with many pretrained networks: AlexNet, CaffeNet, GoogLeNet V1, Inception V3, Inception V4, VGG16, VGG19, ResNet50, ResNet152, Nin, Yolo, SSD…etc.
- Supports custom CNN without modification
- Supported layers: Convolutions, Fully Connected, Max/Average Pooling, Concat, LRN, Relu, Softmax, Batch Norm, Scale, Eltwise, etc
- Up to 1 billion weights in a single network
- Up to 1 million layers
- Up to 200,000 filters per convolution
Supported Frameworks
- Caffe, Caffe2, MXNet or TensorFlow
- No code change required
Computation
- 8-bit or 16-bit integers with automatic quantization
- No mandatory pruning
Migration from GPU or CPU
- Trained parameters from GPU training without changes
- No proprietary training or re-training needed
- Similar accuracy as GPU or CPU
- Switch networks instantaneously without reprogramming FPGA
- Supports multiple users, multiple networks in parallel, multiple boards
Hardware format, power & cooling
- U.2 Small Form Factor (SFF8639)
- Typical power less than 20W, passive cooling
- Single BGA package planned
For further details contact MLA@wdc.com
© 2018 Western Digital Corporation or its affiliates. All rights reserved.