Study Results: Combustor Model
(~24 Million Elements)
In the graphs above, one can see vast speed improve-
ments using the segregated solver as compared to the
coupled solver with GPUs. In some cases, that resulted
in approximately 50% more equivalent CPU cores just
by changing the solver type and holds true when using
GPUs in parallel. As in the mixing tank benchmark model,
the speedup benefits of GPUs far exceed those of the
comparably priced 128 CPU core machine.
©2024 Rand Simulation 8
The Impact of GPU Solving on Simulation Performance
Continued...
In the case of the combustor model using a coupled
solver, two A100 GPUs have the equivalent of approxi-
mately 475 CPU cores compared to the 128 CPU core
baseline case, a 3.7x speedup for the same cost. When
leveraging the segregated solver, that same comparison
shows two A100 GPUs having the equivalent of approxi-
mately 800 CPU cores compared to the same baseline
case. This is a 6.3x speedup for the same money when
using the segregated solver on this larger model. A big
reason for a lot of these performance variations stem
from the size of model, and each model has certain
memory requirements to operate.
