Efficiently Deploying GPU Accelerated 5G CloudRAN for Edge AI Inferencing

Key Takeaways

the telecommunications industry is at the crossroads of immense possibility 5g with the rapidly growing number of connected devices is creating new capabilities and customer expectations the industry needs to deploy flexible and scalable networks to meet them nvidia egx makes this possible the edge computing platform enables an end-to-end 5g solution that can process real-time data and deliver exceptional customer experiences the o-rin alliance is driving the disaggregation of network hardware and software by defining open and standards-based interfaces for edge regional and core cloud architectures this is instrumental for delivering quality of service to the expected explosion of connected users here we'll demonstrate an oran front hall compliant converged cloud ram core network and multi-access edge compute solution accelerated by nvidia gpus you can see the end-to-end solution we've built in the nvidia 5g innovation lab we're using our aerial sdk to perform 5g physical layer signal processing along with solutions from other vendors to complete the 5g networking stack our deepstream sdk then performs streaming intelligent video analytics across the network this solution powered by nvidia egx brings the most advanced telecom transformation to market let's take a look at the details this is the rack containing all the components we're using a ue emulator from keysight a dell server with the aerial sdk running on an nvidia v100 gpu altran's l2 plus solution running on a cpu as well as nvidia mellanox smartnic a ptp grand master from quelsar and nvidia mellanox switch a core network from keysight and a dell server with the deepstream sdk running on an nvidia t4 gpu let's take a closer look at what's happening video streams from multiple cameras enter the ue emulator which enables the oran radio unit and user equipment such as cell phones and laptops to function within the emulator the user equipment encapsulates the video stream packets into the 5g protocol it then sends the physical layer data to the oran radio unit which performs the 5g node b lower layer functional split the radio unit uses the oran compliant front hall protocol to exchange the downlink and uplink waveforms with the 5g gnode b an nvidia melanox connectx 6dx smart nic enables the 5g genode b to perform essential front-haul processing tasks a single smartnic is used for both the front-haul and backhaul communications the smartnic embeds advanced 5t for 5g technology this enables ecpre packet-based ethernet rams to deliver highly accurate ptp time references to 5g front-hall and backhaul networks with this capability networks can efficiently handle time-sensitive network traffic scheduling smartnics provide gpu direct capability for better packet placing and pacing than traditional fpgas the end result is reduced processing time which means better performance and lower latency the egx 5g genodeb server contains the aerial sdk which provides an efficient 5g signal processing solution with in-line acceleration using nvidia's v100 gpu and cuda programming environment cuda baseband keeps all 5g physical layer signal processing such as forward error correction channel estimation and equalization on the nvidia gpu and cuda virtualized network function optimizes the transport of packets between the gpu and nick in this demonstration the aerial sdk is performing the 5g physical layer signal processing it and the l2 plus software stack from altran are both running within a docker container the l2 stack is shared across the distributed and centralized units in the 5g genome b server lastly timing synchronization is achieved using ptp protocol ptp 4l and phc to assist routines the 5g genode b software stack sends and receives data packets from the core network where the gtpu tunneling protocol is used to add and remove packet headers the core network also interfaces to mac applications such as nvidia deepstream after the video stream is sent through the 5g stack deepstream ingests the stream and performs streaming analytics using tensorrt for ai inferencing [Music] in this case we are deploying a sample application running a resnet model that identifies and classifies cars and pedestrians from the nine simultaneous video streams and applies bounding boxes to the identified objects the processed video streams are restreamed out of deep stream and near real time some key performance indicators were captured to visualize aerial 5g network performance these include uplink data rate throughput measured in megabits per second which confirms the 5g bit rate time slot duration accuracy measured in microseconds which shows the time slot accuracy derived by ptp-4l and phc2 sys and locally derived timing error rate measured in nanoseconds which shows the low timing error within the connectx 6dx smartnic this 5g network dashboard is updated once per second this demonstration of a converged cloud ram core network and mech solution showcases the powerful capabilities of the nvidia egx platform and the aerial and deep stream sdks