1.2 Tb/S Packet Processing of Radio Telescope Data on a GPU
summary
System successfully streams 1196 Gb/s from three 400GbE NICs into an NVIDIA Grace Hopper GPU using DPDK GPU support, splitting packets so headers are processed on CPU and 8K payloads DMA straight into GPU memory for zero-copy DMA into GPU memory. A circular pktmbuf ring of about 9.5 million packets buffers 130 ms of data; a 130 ms slice takes ~100 ms to channelize and correlate (77 ms for kernels). GPU and CPU clocks drop (GPU 1.98→1.50 GHz) to save ~100 W while maintaining real‑time performance.
AI-written summary · about
chapters (7)
- 0:00Radio astronomy overview
- 2:47Data processing pipeline
- 8:18Hardware advancements
- 11:22Memory bandwidth challenges
- 13:59Zero-copy implementation
- 17:45Performance and energy
- 19:44Conclusion and Q&A
related talks
36mGNU Radio at the Allen Telescope Array
30mBeyond the DC Walls: Building a Nationwide GPU Fabric With KubeVirt and Wide-Area…
24mBreaking the TCP Barrier: Accelerated I/O for S3 With RDMA
14mServing Frontier MoE Models at the Lowest Token Cost
36mWhy Are GPUs (Not) Fast - A Trip Through the Driver Stack
15mNixie: Efficient, Transparent Temporal Multiplexing for Consumer GPUs
26mAre You Sure About Your Mesh Performance?
39mSuper Fast Real-time Data Processing on Cloud-Native Architecture