Organisation Cloud Software Execution: Difference between revisions
Content deleted Content added
No edit summary |
|||
Line 81:
== Why is my run on the cloud slower than I expected based on the specs? ==
Although cloud hardware may be faster for some use cases, and certainly a lot more expensive to purchase, it may not be guaranteed to run your TUFLOW model faster. This mostly depends on how modern the NVIDIA hardware architecture is, how many CUDA cores it has available and specific metrics of the hardware like the amount of memory, the clock speed of the memory, the clock speed of the cores, and how the GPU is connected to the rest of the hardware. For a good assessment of whether you should expect better performance, refer to <u>[[Hardware Benchmarking
If you're wondering why TUFLOW software doesn't benefit from these supposedly faster and more expensive GPUs, consider that a GPU has many different features, and TUFLOW only makes use of an important subset of these. Also, most TUFLOW models are executed using the single-precision floating-point executable, which is faster than the double-precision executable. Desktop GPUs are highly optimised for single-precision compute, because this is what benefits gaming, and as it happens, TUFLOW runs. Data centre GPUs are more optimised for double-precision compute, but most TUFLOW simulations don't benefit in result quality from using this.<br>
Even when the hardware should be faster according to benchmarks, it's possible that you have some other restrictions. For one, if your cloud environment shares GPUs between many users, the part of the GPU available to your model run may only see a small percentage of the performance it would show with exclusive access to the GPU. This is particularly true in Virtual Desktop Infrastructure (VDI) setups. The way TUFLOW uses the GPU is very different from normal graphics processing, and VDI solutions are often not good for model running.<br>
| |||