The latest Broadcom Tomahawk Ultra switch offers support for minimum packet size, enabling efficient communication in scientific and computational workloads, especially in scale-up networks. This capability allows for streaming of packets at full bandwidth, crucial for GPU-to-switch-to-GPU communication in a single hop scenario.
Lossless Ethernet receives a significant boost with the Tomahawk Ultra, incorporating features defined in the Ultra Ethernet Consortium’s 1.0 specification. Link Layer Retry (LLR) automatically detects transmission errors using Forward Error Correction (FEC) and requests retransmission, while Credit-Based Flow Control (CBFC) prevents packet drops due to buffer overflow. These enhancements ensure reliable and efficient data transmission.
In-network collectives (INC) further enhance the speed of HPC and AI operations by reducing network operations. With the Tomahawk Ultra, compute units like GPUs can share and combine computational results more efficiently. The switch’s hardware can perform operations like averaging directly in the network, reducing bandwidth usage and offloading computation tasks. This streamlined approach benefits overall performance in high-performance computing and artificial intelligence applications.
