Episode 28 dives into Remote Direct Memory Access (RDMA), the networking technology that lets servers move data directly between their memory without involving the CPU or operating system. Lucas explains how RDMA has become essential for high-performance computing and storage in modern data centers, citing a specific example: a large cloud provider cut latency from 100 microseconds to under 10 for storage operations by switching to RDMA over Converged Ethernet (RoCEv2). Luna asks about the main trade-off—RDMA requires lossless networks, so data centers must handle flow control carefully. They discuss how InfiniBand, the original RDMA fabric, competes with RoCEv2, and why NVIDIA's acquisition of Mellanox tilted the balance. A quick look at how RDMA enables GPU-direct communication for AI training clusters, moving data between GPUs at 200 gigabits per second. Donation segment near the end: 'If this tech deep-dive was worth a coffee to you...'
#RDMA #RemoteDirectMemoryAccess #DataCenterNetworking #RoCEv2 #InfiniBand #NVIDIA #Mellanox #GPUDirect #HighPerformanceComputing #StorageNetworking #LosslessNetworking #AIInfrastructure #Technology #NetworkingTech #FexingoBusiness #BusinessPodcast #LucasAndLuna #NetworkEngineering
Keep every episode free: buymeacoffee.com/fexingo
