As the new year dawns, there’s one topic in the tech world that seems hard to avoid: artificial intelligence. Companies are talking so much about artificial intelligence, and the popularity is only growing. In the semiconductor field, most of the attention on AI is focused on GPUs or dedicated AI accelerator chips (such as NPU and TPU). But it turns out there are quite a few components that can directly impact or even run AI workloads. FPGA is one of them.
This is not surprising to those who understand FPGA flexibility and programmability, but to many others the connection between the two may not be obvious. The key to the problem is to use software to allow some classic AI development tools (such as convolutional neural networks (CNN)) to be optimized for customizable circuit designs supported by FPGAs. FPGAs can also create multiple parallel computing pipelines, which is useful for the type of matrix multiplication calculations that are at the heart of many AI algorithms. In addition, the flexibility of FPGA architecture design can be used to allocate memory blocks on the chip, thereby optimizing data transfer, another key requirement for AI software.
Lattice Semiconductor has been developing software tools that enable these types of functions for many years and has a comprehensive suite of products. From adapting existing or newly built AI models into formats that run most efficiently on their low-power designs, to creating circuit and chip designs that work best for those models, these applications can do just about anything. This complete closed-loop system will greatly help enterprises integrate artificial intelligence capabilities into their devices and other hardware.
In terms of AI models, Lattice’s sensAI solution can use models trained in industry-standard AI frameworks such as TensorFlow, Caffe, and Keras, and leverages techniques such as model quantization, pruning, and sparsity utilization to make it more efficient on FPGA resources. run on. The Lattice Neural Network Compiler can then analyze the model and make recommendations based on the type of circuits and on-chip networks to run most efficiently. On the software side, Lattice's Propel and Radiant chip design software can be used to create the right combination of circuits to accelerate the operation of these models in the most energy-efficient manner possible.
Instead of starting from scratch when creating these chip designs, companies can leverage Lattice's purpose-built key IP blocks, such as its family of CNN accelerators. These pre-built circuit sets provide the core foundation for a variety of applications, including person and object detection, object classification, keyword recognition, and more. Additionally, due to the programmable nature of FPGAs, these IP blocks can be edited and added to meet application-specific requirements.
An overlooked but important benefit of this combination of pre-built IP blocks is that it enables more developers to create custom FPGAs. This is critical because many acknowledge that FPGAs, despite their powerful and flexible features, are difficult to program. Developing dedicated RTL code for the core of an FPGA design has always been a specialized task that can only be completed by a few people, so it is necessary to provide chip designers with the right tools to connect pre-built components together in a Lego-like way to make development more efficient. Simple.
Likewise, Lattice is empowering traditional AI frameworks that many software developers are already familiar with, such as TensorFlow, to help a broader group of people create AI models that run on FPGAs. In fact, it is precisely thanks to this simplification that the application potential of FPGAs in AI applications is so diverse. As companies across a wide range of industries scramble to figure out how best to apply artificial intelligence in areas ranging from automotive to medical to consumer electronics to industrial and beyond, there will be a broader base of potential customers looking to semiconductor solutions to enable these capabilities. While some of these people may have been aware of or considered FPGAs as a potential option in the past, Lattice Semiconductor offers products that can make FPGAs a strong choice that even more groups can rely on.