Hi Asim, The GPUs on JLSE are definitely different from anything that is available publicly and therefore conform to NDA/TOU. If you look at this page, https://wiki.jlse.anl.gov/display/inteldga/Getting+Started And the presentation titled “NDA Presentation on Aurora Testbed from ECP AM'22 for quick reference” in slide 9 you will find the theoretical max and then slides 21,22,23 will show the measured peaks on very simple benchmark. You had mentioned dense linear algebra, so I was wondering if you were looking for specific GEMM or other BLAS3 performance, which we do not have documented as its highly dependent on the config/opt/implementation. As for the meaning of those env, some of them should have documentation if you do module help on the files. Many still needs this to be added so they are WIP. They are simply setting or rather showing the config of the platform, for ex, in the case of Arcticus, we have multi_gpu which means both the GPUs are visible. -- Servesh From: Asim YarKhan <[email protected]> Date: Thursday, August 18, 2022 at 5:53 PM To: Muralidharan, Servesh <[email protected]> Cc: [email protected] <[email protected]> Subject: Re: [Ecp-aurora-sdk-discuss] Arcticus and florentia theoretical peak performance Hi Servesh,
Is a table on the JLSE "Intel Discrete GPU" Confluence site that talks about the theoretical peaks of the arcticus and florentia GPUs? For 1, what specific operation peak are you looking for on these systems? I am looking for something that gives the peak theoretical (or practical) performance of a single GPU on arcticus running double precision operations. I found the following documentation online for the Arctic Sound 2T, https://www.techpowerup.com/gpu-specs/arctic-sound-2t.c3503 The double precision performance here is shown as 3.4 TFlops [cid:ii_l6zmlpg10] That does not match up with my expectations, so I was looking on the JLSE confluence site for more information. I was hoping that there was a page on the JLSE confluence that simply listed theoretical peak GFlops for each precision.
For the second question re the meaning of " intel_gpu_env/multi_gpu ", I will wait for more details... Thanks, Asim On Wed, Aug 17, 2022 at 12:41 PM Muralidharan, Servesh <[email protected]<mailto:[email protected]>> wrote: Hi Asim, For 1, what specific operation peak are you looking for on these systems? There is a variety of config options that can influence these peaks and given these are early testbeds that go through upgrades that could influence these numbers as well. For 2, i.e. intel_gpu_env/<modulefile> these are under testing at the moment and we would provide an update aligned with the next SDK release. These modules set various environment options that otherwise would need to be added to user’s job script. Currently, they have a config that defaults to the recommend settings for each of the platform that provides the best out of the box performance/WA for the early hardware. -- Servesh From: Ecp-aurora-sdk-discuss <[email protected]<mailto:[email protected]>> on behalf of Asim YarKhan via Ecp-aurora-sdk-discuss <[email protected]<mailto:[email protected]>> Date: Wednesday, August 17, 2022 at 11:12 AM To: [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>> Subject: [Ecp-aurora-sdk-discuss] Arcticus and florentia theoretical peak performance The ECP SLATE project is working on distributed dense linear algebra libraries. We are trying to figure out how well our implementations are performing on these platforms. To that end, I was wondering if there is a table on the JLSE "Intel Discrete GPU" Confluence site that talks about the theoretical peaks of the arcticus and florentia GPUs? Also, is there a page that describes best practices in launching jobs and gives explanations for the meaning of these modules intel_gpu_env/multi_gpu
