It's nearly time for embedded world 2018 (yes, it likes to be trendy and put it all in lower-case), which takes place February 27 to March 1. It happens every year in Nuremberg, and is the biggest embedded conference in the world. Cadence will be there. It is a busy week, since it is also Mobile World Congress (in Barcelona) and DVCon (in San Jose). Cadence will be both of those events, too. Yes, I turned down the opportunity to go to Barcelona, which will be sunny (probably) and has wonderful food. Nuremberg is famous food-wise for Nuremberg Sausages, which are a sort of Bratwurst smaller than anywhere else (yeah, not much of a recommendation). The tomato in that picture is just a cherry tomato, they are the same size, although much better tasting, than the sausages you get in your Denny's Grand Slam. It is also famous, at least to me, as being the place where your plane gets diverted to when Munich airport is fogged in. You then get bused to Munich. If you are lucky, you get to your hotel in Munich only about three hours late. When I lived in France, I had to go to Munich every month for our operations review, so this happened to me more than once. embedded world has started to feel like automotive world, since ADAS is driving so much technology, especially in Germany (home to BMW, Mercedes, VW, and their other brands). We are no exception, and the Cadence booth will be showing several applications of Tensilica processors in automotive applications. So this is what you will see if you visit the booth, which is in Hall 4, booth 4-116. Autonomous Driving Video Fusion One of the challenges is autonomous driving is merging feeds from multiple cameras (and, perhaps, radar and lidar). This demo provides a synthetic birds-eye-view created by four cameras on a car, stitched together. It includes lens distortion correction, and perspective transformation, to create a seamless image that is displayed in an ADAS automotive application (such as park-assist). The hardware is an SoC containing four Tensilica Vison P6 DSP cores, manufactured (in Dresden) in GLOBALFOUNDRIES 22FDX process. This demo actually just uses one of the cores. On-Device AI for People Detection Obviously, a self-driving car needs to be able to detect people and distinguish them from other things that aren't likely to step off the sidewalk unexpectedly. The demo runs on a Tensilica Vision P6 DSP and a You Only Look Once (YOLO) algorithm based on a neural network specifically trained for people detection. YOLO is ideal for applications requiring fast, power-efficient object detection (with localization) for multiple object categories and can be retrained for up to 20 object classes. It uses custom network optimization, including pruning and quantization. The demo was developed in partnership with BDTi. THis uses the same SoC as the video fusion demo (see above), again only using one of the cores. On-Device AI for Image Classification One of the underrated factors advancing AI and neural networks was ImageNet. Despite the name, ImageNet is not a neural net, it is a database of millions of images tagged with what is in the picture (and not just at the level of "it's a dog" but "it's a labrador", too). This database has been used to train neural nets such as Inception V3 and Mobilenet. The training is done using 32-bit floating point, in the cloud. However, that is much too expensive in terms of power and area to run directly on the chip. Instead, the network is optimized for embedded inference by running it through XNNC, the Xtensa Neural Network Compiler. This converts the floating-point neural network into highly optimized fixed-point code that can then run on a Tensilica AI DSP. XNCC reduces the time to convert a trained neural network for embedded inference from several months to a few days. It does this without compromising the detection quality of the original floating point network while providing a low-power, area-efficient implementation that also significantly reduces the memory bandwidth required. Automotive Pedestrian Detection with Protium S1 When developing automotive applications, such as pedestrian detection, there needs to be a way to verify the performance of the software algorithms before the SoC is available, and when changes can still be made to the SoC if necessary. Cadence has two hardware-based solutions for doing this, the Palladium emulator and the Protium S1 FPGA-based prototyping platform. Palladium is not really practical to show in an exhibit hall. It is big, heavy, and produces a lot of heat—most users water cool it since that is the more efficient than air cooling. Palladium allows full visibility of the system and so is most appropriate when the hardware design is still in flux. Protium is faster, but the user needs to decide in advance which signals to monitor, so it is most appropriate for software development once the hardware is stable enough that running into hardware problems should be rare. I wrote a post about Protium: Next Generation FPGA Prototyping last year (on Mardi Gras it seems, which I believe is the end of Fasching, at least in Munich, not sure about Nuremberg) and about Pre-Silicon Software Development with Palladium and Protium . The demo at embedded world runs a pedestrian detection algorithm on an FPGA prototype of a Tensilica processor, running a neural network. So there are lots of things being shown either directly or indirectly: the Protium platform, the Tensilica DSP, and XNNC. It will not be running on a file of captured video data, it will be running on a live video feed. Booths Once again, we are at booth 4-116. Look, we're allready setting it up ready for your visit (well, this was 2017 of course, but it won't look much different). If you go to Mobile World Congress in Barcelona instead, some of these demos will be there, too. Our booth is hall 6, stand 6L3 (hint: it's at the back). Sign up for Sunday Brunch, the weekly Breakfast Bytes email.![]()
↧