Difference between revisions of "Hardware Selection Advice"
(Created page with "This page summarises general hardware advice for running TUFLOW models on GPU or CPU hardware. <br> right =Introduction= We o...")
Revision as of 10:44, 25 May 2020
This page summarises general hardware advice for running TUFLOW models on GPU or CPU hardware.
We often get asked about the optimum computing setup to run TUFLOW models. While every model is different and will interact differently with your hardware there is some general advice that we can offer. In the sections below you will find more detailed advice on GPU and CPU but generally:
- The amount of RAM in your computer will be the limiter for the size of model you can to run. This applies to CPU RAM (TUFLOW Classic, TUFLOW FV and TUFLOW HPC with Hardware == CPU) and also GPU RAM (TUFLOW HPC with Hardware == GPU).
- The processing speed of your CPU, the architecture, cache size, speed and number of processors play a role.
- For GPU simulations, the number of CUDA cores, the core speed, GPU card architecture, memory speed and interfacing with the motherboard PCI lanes and CPU are all important.
- The system must be well cooled to avoid throttling.
The TUFLOW Software Suite
The TUFLOW Software suite has a range of solvers. Each interact differently with your hardware so pairing the correct solver (or the range of solvers you want to run) and hardware is an important consideration. A brief summary of each solvers's needs is provided as follows:
- TUFLOW Classic: A single model run can only use the CPU and cannot be run across multiple CPU cores or GPU hardware. In general terms: The maximum model size is dependent on the available CPU RAM and the runtime is driven by the CPU speed, architecture and cache size.
- TUFLOW HPC - Run on CPU Hardware: A single model run uses the CPU and is parallelised to run across multiple cores. In general terms: The maximum model size is dependent on the available CPU RAM and the runtime is driven by the CPU speed, the number of cores available to be run in parallel, architecture and cache size.
- TUFLOW HPC - Run on GPU Hardware: A single model run uses the GPU(s) for computation. In general terms: The maximum model size is dependent on the available GPU and CPU RAM and the runtime is driven by the CUDA core speed, the number of CUDA cores available and the GPU architecture. GPU performance is complex and is not easily inferred from GPU clock speed and number of cores, it is also very dependent on the ‘generation’ or architecture of GPU. As TUFLOW HPC requires some data exchange between GPU and CPU, the motherboard bus speeds and CPU speeds also play a role but typically a much lesser role compared to the GPU CUDA compute.
- TUFLOW FV: A single model run uses CPU and is parallelised to run across multiple cores. GPU compute is not yet available. In general terms: The maximum model size is dependent on the available CPU RAM and the runtime is determined by the CPU speed, the number of cores available to be run in parallel, chip architecture and cache size.
On our Hardware Benchmarking page you can compare recently run combinations of GPU, CPU and RAM with the system you are planning to purchase. We recommend that if building a computer that you seek advise from an appropriate computer hardware vendor who can advise on the compatibility and optimisation of your setup.
TUFLOW HPC on GPU Hardware is typically our fastest solver for 1D/2D pipe and flood plain simulations.
- TUFLOW HPC supports CUDA enabled NVIDIA GPU cards. For list of supported CUDA enabled graphics cards please visit the NVIDIA website.
- TUFLOW HPC on GPU Hardware can be run in either single or double precision. However, for the vast majority of flood applications single precision is sufficient. We typically run our models on single precision. If you are unsure we recommend running with both the single and double precision solvers and comparing your results.
The precision solver you require will determine the type of GPU card that is best suited for your compute. For any given generation/architecture of cards, the “gaming” cards such as the GTX GeForce and RTX provide excellent single precision performance – typically comparable to that of the “scientific” cards such as the Tesla series. If double precision is required then the scientific cards are substantially faster, but these are also significantly more expensive. The Quadro series cards sit in between for both double precision performance and cost. When checking the specifications of the card it should provide you with a breakdown of the single and double precision throughput in flops.
RAM is the computer memory required to store all of the model data used during the computation. A computer has CPU RAM which is located on the motherboard and accessed from the CPU, and it has GPU RAM which is located on the GPU device and accessed from the GPU. The two memory storage systems are physically separate. The amount of GPU RAM is one of two factors that will determine the size of the model that can be run (the other being CPU RAM). As a rule of thumb, approximately 5 million cells can be run per gigabyte (GB) of GPU RAM depending on the model features, e.g. a model with infiltration requires more memory due to the extra variables needed for the infiltration calculation.
TUFLOW HPC on GPU hardware still uses the CPU to compute and store data (in CPU RAM) during model initialisation and for all 1D calculations. While we are working on improving our CPU RAM usage, currently we tend to find that CPU RAM is often the limiter to the size of the model domain you can run, particularly if using running over multiple GPU cards. During initialisation and simulation a model will typically require 4-6 times the amount of CPU RAM relative to GPU RAM. As an example, a model that utilises 11GB of GPU RAM (typical memory for high-end gaming card, and corresponds to about a 50 million cell model) the CPU RAM required during initialisation will typically be in range 44GB to 66GB. A model that fully utilises two 11 GB GPUs (i.e. a 100 million cell model) may require as much as 128GB of CPU RAM during initialisation.
CUDA Cores, GPU Clock speed, and FLOPs
One way of reporting a GPU card's throughput is in Floating Point Operations per second (FLOPs). The more FLOPs, the more calculations that can get crunched per second and the faster the model should run. For any given generation of GPU, FLOPs are approximately proportional to number of CUDA cores times the GPU clock speed. However, there have been significant improvements in GPU architecture since the inception of CUDA, and this has contributed to increases in overall FLOPs performance beyond just the increases in cores and clock speed that have occurred over this time.
TUFLOW can use multiple GPU cards on a machine to run a single model. This is useful for models that are too large for a single GPU, or for running a model as quickly as possible. In general terms the run time benefit of using multiple cards increases with model size.
- TUFLOW HPC-GPU does not support SLI for inter-GPU communications. It does (as of build 2020-01-AA) auto detect and utilise peer-to-peer access over NVLink or PCI bus on the motherboard. Note that not all GPUs support peer-to-peer access.
- When using multiple GPUs it is best to use cards of similar memory and performance. While it is possible (as of build 2020-01-AA) to re-balance a model over multiple GPUs, we do not recommend using cards with vastly disparate performance.
- Sufficient cooling and power supply should be considered if multiple cards are used. When installed in adjacent PCI slots, the preference is to use rear vented cards rather than side vented to avoid blowing hot air onto the neighbouring cards (which could lead to overheating).
In general terms a more recent architecture, higher clock speed CPU with a large cache will perform better than a slower clock speed chip. This section discusses CPU RAM, RAM speed, Processor frequency, Multi-core processing and hyper-threading.
The amount of CPU RAM will determine the size of the model that can be run or a number of models that can be run at one time. Faster RAM will result in quicker runtimes, however this is usually a secondary consideration to chip speed, cache size and architecture.
- TUFLOW HPC - Run on GPU Hardware: The parallel processing is being done on the GPU card. However, TUFLOW HPC-GPU still uses the CPU for model initialisation and for 1D calculations. If multiple GPU cards are used, TUFLOW will use the equivalent number of CPU threads for controlling the GPUs and migrating data. So for a machine dedicated to HPC-GPU modelling, the number of CPU cores should be higher than the number of installed GPUs.
- TUFLOW HPC - Run on CPU Hardware: HPC model can also be run on multiple CPU cores. For the comparison of simulation speed, please refer to HPC on CPU vs GPU.
- TUFLOW Classic: TUFLOW Classic simulation can only use one CPU core due to the implicit nature of the numerical solution, for more information see this TUFLOW forum post. More CPU cores will enable running more simulations at the same time most efficiently.
Processor Frequency and RAM Frequency
The frequency directly affects the run times. In general, the higher the frequency, the faster the model runs.
Solid state hard drives are preferred for temporary storage. Large data files can then be transferred to a more permanent location.