. 2020 Jul 31;27(Pt 5):1297–1306. doi: 10.1107/S1600577520008140

Table 2. Observed bandwidths (in Gb s⁻¹) from RNIC to CPU memory using RDMA.

The RoCE implementations use the RNIC capabilities efficiently. The SEND API is similar to a stream transfer. The WRITE API specifies, in addition, the destination memory address. The RAW API does not require a specific packet format.

Packet size (bytes)	RoCE-v2 SEND	RoCE-v2 WRITE	RDMA RAW
2048	94.5	95.2	33.3
4096	97.6	97.5	60.5
8246	N/A^†	N/A^†	99.1

^†

RoCE maximum payload (MTU) is 4096.

Table 2. Observed bandwidths (in Gb s−1) from RNIC to CPU memory using RDMA.

Table 2. Observed bandwidths (in Gb s⁻¹) from RNIC to CPU memory using RDMA.