To make our table useful, we want to balance coverage of options with focus on common choices -- to cover all of the configurations that fit the workloads that most people are interested in.
The typical purpose of multi-GPU instances is to handle workloads that are infeasible when parallelized across multiple single GPU instances.
In that case, the primary benefits come from efficient inter-GPU communication and secondarily from balanced divisibility of a unit of work across the GPUs. Powers of two (2x, 4x, 8x) are common for improving divisibility, which leads to gains all the way down to the driver level, and they support standard interconnect topologies.
Notably, almost all of the information I can find for 7x GPU setups is for cryptocurrency mining, which is a very different kind of workload and out of scope for our table.
Explain why the information needs to be updated.
To make our table useful, we want to balance coverage of options with focus on common choices -- to cover all of the configurations that fit the workloads that most people are interested in.
The typical purpose of multi-GPU instances is to handle workloads that are infeasible when parallelized across multiple single GPU instances.
In that case, the primary benefits come from efficient inter-GPU communication and secondarily from balanced divisibility of a unit of work across the GPUs. Powers of two (2x, 4x, 8x) are common for improving divisibility, which leads to gains all the way down to the driver level, and they support standard interconnect topologies.
Notably, almost all of the information I can find for 7x GPU setups is for cryptocurrency mining, which is a very different kind of workload and out of scope for our table.