Compare Microsoft Cognitive Toolkit vs. Neural Magic

The Microsoft Cognitive Toolkit (CNTK) is an open-source framework designed for high-performance distributed deep learning applications. It represents neural networks through a sequence of computational operations organized in a directed graph structure. Users can effortlessly implement and integrate various popular model architectures, including feed-forward deep neural networks (DNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs/LSTMs). CNTK employs stochastic gradient descent (SGD) along with error backpropagation learning, enabling automatic differentiation and parallel processing across multiple GPUs and servers. It can be utilized as a library within Python, C#, or C++ applications, or operated as an independent machine-learning tool utilizing its own model description language, BrainScript. Additionally, CNTK's model evaluation capabilities can be accessed from Java applications, broadening its usability. The toolkit is compatible with 64-bit Linux as well as 64-bit Windows operating systems. For installation, users have the option of downloading pre-compiled binary packages or building the toolkit from source code available on GitHub, which provides flexibility depending on user preferences and technical expertise. This versatility makes CNTK a powerful tool for developers looking to harness deep learning in their projects.

GPUs excel at swiftly transferring data but suffer from limited locality of reference due to their relatively small caches, which makes them better suited for scenarios that involve heavy computation on small datasets rather than light computation on large ones. Consequently, the networks optimized for GPU architecture tend to run in layers sequentially to maximize the throughput of their computational pipelines (as illustrated in Figure 1 below). To accommodate larger models, given the GPUs' restricted memory capacity of only tens of gigabytes, multiple GPUs are often pooled together, leading to the distribution of models across these units and resulting in a convoluted software framework that must navigate the intricacies of communication and synchronization between different machines. In contrast, CPUs possess significantly larger and faster caches, along with access to extensive memory resources that can reach terabytes, allowing a typical CPU server to hold memory equivalent to that of dozens or even hundreds of GPUs. This makes CPUs particularly well-suited for a brain-like machine learning environment, where only specific portions of a vast network are activated as needed, offering a more flexible and efficient approach to processing. By leveraging the strengths of CPUs, machine learning systems can operate more smoothly, accommodating the demands of complex models while minimizing overhead.