Support for the nVidia A100 GPU
The new A100 GPU offers over twice the performance of the current V100. Bencmarks testing systems wityh 8-A100s have achieved over 100 Tflops. It is important to run the version of FMS corresponding to the driver on your system. A list of the appropriate drivers is containd on the Download page.Automatically configure CPUs to skip hyperthreads (Linux)
"Processors" on Linux systems may point to a physical core or be hyperthreads, which run in parallel on the same cores. All routines in FMS are designed to use all the reasources on each physical core. Scheduling additional processors on the same hardware degrades performance.On startup FMS analyzes the CPU configuration and identifies the sockets (physical chip) and cores within each chip for each processor. When a processor is found that uses the same socket and core as a previous one, it is identified as a hyperthread. The maximum number of physical cores found is used to set the FMS Parameter MAXCPU.
The CPUS page displayed by Dashboard displays detail information about each processor, including the socket number, core nmber and if it is a hyperthread.
Memory to GPU bandwidth test (Linux)
Achieving maximum performance with FMS requires a fast and uniform transfer of data between memory and each GPU. FMS now contains the functionality to perform these tests. The FMS Parameter CGTEST is used to control which tests are performed.- 0, no test is performed
- 1, GPU threads are left bound to their CPU cores as setup by FMS. The first test measures the transfer rate to each GPU from it's bound CPU core(s). The second test measures the transfer rate with all GPU threads running in parallel.
- 2, Measure the transfer rate from each CPU to each GPU.
- 3, Perform tests 1 and 2.
Dashboard now contains a CPU-GPU page, which lists the results of these tests.