Running Intel MKL-DNN On 2 x Xeon Platinum 8280 CPUs With GCC 9 "Cascadelake" Tuning
For those wondering about Intel's MKL-DNN "Math Kernel Library for Deep Neural Networks" performance on Cascade Lake, here are some reference benchmarks using the dual Xeon Platinum 8280 setup including when using the GCC 9 compiler for building this Intel open-source library and employing the "cascadelake" compiler tuning.
MKL-DNN is one of Intel's open-source deep learning libraries and in turn is used by Caffe, Nervana Graph, OpenVINO, Tensorflow, PyTorch, and other popular software projects.
As part of the Cascade Lake benchmarking fun, I finally got around to adding a "mkl-dnn" test profile to the Phoronix Test Suite / OpenBenchmarking.org so now via the Phoronix Test Suite you can run performance tests of it in different configurations simply by phoronix-test-suite benchmark mkl-dnn and is currently using the latest Git code of this Intel-optimized library.
On the dual Xeon Platinum 8280 server built on a Gigabyte Xeon Scalable barebones setup while running Ubuntu 18.04 LTS, I did some quick tests of this initial MKL-DNN profile while using the current GCC 9.0.1 compiler. The GCC9 compiler will debut as stable in the next few weeks in the form of "GCC 9.1" as the first stable release and with this annual GNU compiler update is the initial "cascadelake" target that includes enabling AVX-512 VNNI support over the existing "skylake-avx512" target that is used for 1st Gen Xeon Scalable CPUs. I ran MKL-DNN benchmarks both when built by GCC9's skylake-avx512 target and then again with cascadelake while "-O3" was also part of the CFLAGS/CXXFLAGS.
The cascadelake tuning did help a number of the tests.
For those wanting to compare your own system's MKL-DNN performance to this dual Xeon Platinum 8280 server, there is this OpenBenchmarking.org result file. With the Phoronix Test Suite installed on your Linux system simply run phoronix-test-suite benchmark 1904187-HV-CASCADELA78.
More extensive MKL-DNN tests will be coming on Phoronix soon now that the test profile is finally in place.
MKL-DNN is one of Intel's open-source deep learning libraries and in turn is used by Caffe, Nervana Graph, OpenVINO, Tensorflow, PyTorch, and other popular software projects.
As part of the Cascade Lake benchmarking fun, I finally got around to adding a "mkl-dnn" test profile to the Phoronix Test Suite / OpenBenchmarking.org so now via the Phoronix Test Suite you can run performance tests of it in different configurations simply by phoronix-test-suite benchmark mkl-dnn and is currently using the latest Git code of this Intel-optimized library.
On the dual Xeon Platinum 8280 server built on a Gigabyte Xeon Scalable barebones setup while running Ubuntu 18.04 LTS, I did some quick tests of this initial MKL-DNN profile while using the current GCC 9.0.1 compiler. The GCC9 compiler will debut as stable in the next few weeks in the form of "GCC 9.1" as the first stable release and with this annual GNU compiler update is the initial "cascadelake" target that includes enabling AVX-512 VNNI support over the existing "skylake-avx512" target that is used for 1st Gen Xeon Scalable CPUs. I ran MKL-DNN benchmarks both when built by GCC9's skylake-avx512 target and then again with cascadelake while "-O3" was also part of the CFLAGS/CXXFLAGS.
The cascadelake tuning did help a number of the tests.
For those wanting to compare your own system's MKL-DNN performance to this dual Xeon Platinum 8280 server, there is this OpenBenchmarking.org result file. With the Phoronix Test Suite installed on your Linux system simply run phoronix-test-suite benchmark 1904187-HV-CASCADELA78.
More extensive MKL-DNN tests will be coming on Phoronix soon now that the test profile is finally in place.
Add A Comment