GE Intelligent Platforms attended and exhibited again at NVIDIA’s GPU Technology Conference (GTC) in San Jose, Calif., March 24–27. This was our fourth straight year attending, and there were four reasons it stood out as a particularly memorable event.
- NVIDIA announced the Tegra K1, their latest mobile processor. It features a Kepler GPU with 192 CUDA cores, and a quad-core ARM Cortex-A15 processor. There is also hardware support for video encoding and decoding. This represents the first of the mobile-class Tegras to feature GPGPU programming via the same CUDA tools that have been available for years now on higher-end GPUs—like the EXK107 Kepler device featured on our IPN251 and GRA112 products. This offers the exciting prospect of over 300 GFLOPS peak at around 5 W of power—a compelling GFLOPS/watt metric for many power-constrained applications such as small unmanned vehicle payloads and man-wearable devices.
- As NVIDIA’s preferred provider of this new technology, we had early access to the evaluation platform, and we were able to demonstrate CUDA applications that were migrated from Kepler platforms on Tegra K1—right from our stand at the vendor exhibition. This featured optical flow algorithms running on video captured from a high-resolution optical camera, and range finding and tracking from a LiDAR sensor. We already have our first rugged K1 designs underway, so if you have a need for lots of GFLOPS but can’t afford many watts, contact your local GE Intelligent Platforms sales representative to see what we can do for you.
- NVIDIA introduced Pascal, the next generation of GPU devices that will follow the current Maxwell architecture. Forget that, to those of us of a certain age, Pascal is (as well as a 17th-century mathematician and physicist) a structural and procedural programming language. (Remember Turbo Pascal, anyone?) There are some interesting architectural features of this new device. It will feature memory stacked on the GPU die, allowing for a very wide memory bus and, therefore, a big increase in memory bandwidth. It will also introduce NVLINK, a high-speed successor to PCIe that may turn up as a GPU to CPU link, but will certainly feature as a GPU-to-GPU link, allowing for more efficient clustering of GPUs. Pascal is not slated to appear until 2016, but it serves to illustrate the continued innovation in the GPGPU space.
- Two members of our technical team presented at the event. Dustin Franklin talked about how GPUDirect RDMA can make GPUs more effective in applications such as radar, how Tegra K1 brings the compute power and ease of programming of CUDA to applications with stringent SWaP constraints, and took a look forward to Maxwell devices. See a stream of his presentation here.
Dave Tetley presented on how middleware libraries like VSIPL can help developers to write applications that take advantage of GPUs without having to write CUDA code, and can allow those applications to be migrated between different processor types and generations, thereby protecting software investments.
From demonstrations to product announcements to technical presentations, we experienced a huge amount of interest in the technology and discussed many interesting (and some strange) applications that we look forward to following up on.