Virtual Reactor can be used on either Linux or Windows. The following table lists minimum and recommended system requirements.
Minimum | Recommended | |
---|---|---|
Operating System | 64-bit RHEL 7 / CentOS 7, Ubuntu 16.04, or other recent 64-bit Linux 64-bit Windows 10 |
64-bit RHEL 7 / CentOS 7 or Ubuntu 20.04 LTS Windows 10 Professional 64-bit |
CPU | Any 64-bit Intel compatible from the last 5 years | Intel Core i7-9800X (4.4 GHz, 8 cores, 16.5 MB cache) or better. Higher clock speed and newer Intel architecture are better. |
Memory (RAM) | 8 GB | At least twice as much as your GPU Memory. Faster is better. |
Hard drive space | 500 GB | Boot Drive: 1 TB SSD or M.2 NVMe Data Drives: 2 x 8 TB 7200 RPM HDDs |
GPU * | NVIDIA GPU required
Pascal series or later is required for multi-GPU |
NVIDIA A100 (40 GB) NVIDIA Quadro GV100 (32 GB) NVIDIA Quadro RTX 8000 (48 GB) NVIDIA Titan RTX (24 GB) |
CUDA Compute Compatibility * | 3.5 (Kepler architecture) | 6.0 or higher (Pascal architecture or newer) |
NVIDIA Driver * | 450.36.06 | Latest available |
GPU RAM * | 4 GB | 24 GB or more |
* – GPU requirements only apply if Virtual Reactor will be running in GPU parallel mode. The required GPU RAM is dependent on the size of the simulation being run. A larger simulation will require more GPU RAM. Multi-GPU capability is only available for Linux operating systems.
CPFD has partnered with Mark III Systems who have pre-configured multi-GPU systems available for purchase.
GPU-specific Considerations
Guidelines for selecting a GPU for Barracuda Virtual Reactor acceleration
- Only NVIDIA brand GPU cards are supported by Barracuda Virtual Reactor.
- CUDA Compute Capability versions 3.5, 3.7, and 5.0 are marked for deprecation by NVIDIA and may not be supported in the future. This includes the remaining non-deprecated Kepler microarchitecture GPUs and some of the Maxwell architecture GPUs.
- GPU cards must be CUDA Compute Capability version 6.0 or newer to support a multi-GPU parallelization. This corresponds to Pascal microarchitecture or newer cards; all cards released since April 2016 fit this criteria.
- NVIDIA Data Center GPUs/Tesla cards will yield the best performance for Virtual Reactor simulations on both Linux and Windows.
- NVIDIA Quadro cards have great performance for Virtual Reactor simulations on both Linux and Windows (when set to TCC mode, see below).
- NVIDIA TITAN series cards have good performance on both Linux and Windows (when set to TCC mode, see below).
- NVIDIA GeForce series cards will work with Virtual Reactor, but are not recommended on Windows systems because they do not support TCC mode (see below).
Important notes for GPU acceleration of Virtual Reactor calculations on Windows
- Calculation speed is best when Tesla Compute Cluster (TCC) mode is set. Only Quadro and TITAN series cards support TCC mode. GeForce series cards do not support TCC mode, and can only be run in Windows Display Driver Model (WDDM) mode. It is strongly recommended that Virtual Reactor calculations be run on cards that have TCC mode enabled (see Setting TCC Mode). On some timing tests, CPFD has found that TCC mode can be up to twice as fast as WDDM mode.
- Because TCC mode cannot be set for a graphics card that is connected to a monitor, it is recommended that a lower-power / lower-cost GPU card be installed in the first motherboard slot. This lower-power GPU card should be used to run all connected monitors (running in WDDM mode), and the TCC mode GPU cards should be dedicated to running Virtual Reactor calculations.
- Multi-GPU simulations and GPU memory oversubscription are not supported when running Barracuda on Windows.
Multi-GPU System Requirements
- Linux operating system support only. Multi-GPU is not supported on Windows.
- RedHat/CentOS 7 or newer, or Ubuntu 16 or newer are supported. Other versions of Linux with new enough libraries will also likely work, but CPFD only officially tests on and supports RedHat/CentOS 7 and Ubuntu 16 or newer. Note that CentOS 6 cannot be used to run Virtual Reactor 21.0 because its libraries are too old.
- NVIDIA GPU cards must be Pascal architecture or newer, which corresponds to CUDA Compute Capability 6.0 and greater. NVIDIA released Pascal architecture GPU cards beginning in 2016. Older GPU cards can still be used for single-GPU acceleration, but cannot be used for multi-GPU simulations.
- NVIDIA driver version 450.36.06 or newer. It is recommended that you upgrade to the latest NVIDIA driver for your GPU cards.
- Disable IOMMU / VT-d in BIOS. For systems that are not being used as virtual machine hosts, it is necessary to disable IOMMU / VT-d in the system BIOS. This feature has been found to cause instability when running multi-GPU Virtual Reactor simulations.
GPU cards listed by architecture
The following table lists GPU cards known to work with Barracuda Virtual Reactor, grouped by NVIDIA architecture. Additional information about GPU cards in the different NVIDIA architecture groups can be found on Wikipedia’s CUDA article.
Architecture | Model | Memory (GB) | CUDA Cores | Release Date |
---|---|---|---|---|
Ampere (CUDA 8.0) | A100 | 40 | 6912 | 2020-05-14 |
Turing (CUDA 7.5) | Quadro RTX 8000 | 48 | 4608 | 2018-08-13 |
Quadro RTX 6000 | 24 | 4608 | 2018-08-13 | |
TITAN RTX | 24 | 4608 | 2018-12-18 | |
Quadro RTX 5000 | 16 | 3072 | 2018-08-13 | |
Volta (CUDA 7.0) | Quadro GV100 | 32 | 5120 | 2018-03-27 |
Pascal (CUDA 6.1) | Quadro P6000 | 24 | 3840 | 2016-10-01 |
Quadro P5000 | 16 | 2560 | 2016-10-01 | |
TITAN Xp | 12 | 3840 | 2017-04-06 | |
Titan X (Pascal) | 12 | 3584 | 2016-08-02 | |
GeForce GTX 1080 Ti | 11 | 3584 | 2017-03-05 | |
Maxwell (CUDA 5.2) | GeForce GTX Titan X | 12 | 3072 | 2015-03-17 |
Kepler (CUDA 3.5) | Tesla K40c | 12 | 2880 | 2013-10-08 |
Quadro K6000 | 12 | 2880 | 2013-07-23 | |
Quadro K5200 | 8 | 2304 | 2014-07-22 | |
GeForce GTX Titan Black | 6 | 2880 | 2014-02-18 | |
GeForce GTX Titan | 6 | 2688 | 2013-02-21 |
Additional Considerations
Though Virtual Reactor simulations can be run on laptops, or lower-performance desktop machines, doing so is generally not recommended. Investing in an up-to-date calculation machine, with the fastest hardware currently available, will provide much faster calculation speeds. Additionally, since computer hardware advances in capacity and speed at such a fast pace, it is recommended to purchase updated hardware every 2 to 3 years to obtain the fastest performance.
Virtual Reactor can be installed on compute nodes of a cluster. However, it will not take advantage of the multi-node parallel computing capabilities of the cluster. Each Virtual Reactor simulation utilizes a single CPU core, and parallelization is available solely through the use of an NVIDIA GPU. Since each individual node of a cluster is not usually optimized for the fastest possible single-core CPU performance, it is often the case that running Virtual Reactor on a cluster node will not give the best possible calculation speed. Instead, it is generally better to purchase a very fast single-CPU standalone calculation machine on which to run Virtual Reactor. This standalone machine will outperform a cluster node in the majority of cases, for the purpose of running Virtual Reactor.