CPU/GPU/Accelerators sub-WG phoneconf

Europe/Zurich
31/S-027 (CERN)

31/S-027

CERN

10
Show room on map
  • Local: Andrea Sciabà, Servesh Muralidharan
  • Remote: Michele Michelotto, Chris Hollowell, Shigeki Misawa, Charles Leggett, Mattieu Puel

Charles notes that RTX cards are low in double precision performance. He suggests to add information about the Intel CSA architecture. About inference, he suggests to explain what can be done, he thinks that for HEP there is little benefit to use something other than CPUs because even if a different type of processor can run faster, you still need to get data in and out of the accelerator.

Servesh asks if anybody knows somebody working on FPGA; in case please contact him.

Servesh will contact Aritz Brosa to see if he is still available to provide material on MIPS and RISC-V.

Shigeki wonders whether the document should separate architectures for embedded processors and for server processors. Andrea thinks that being our focus on server processors, at least for the immediate future we can avoid investing effort on embedded processors.

He also suggests to better explain what is the current machine learning software ecosystem: nowadays most people develop software using CUDA, which drives the hardware. However there are alternatives to Nvidia (AMD, FPGAs, ...) and we should say something about it. For example, Tensorflow has an implementation for CUDA and also for the Xilinx FPGA.

Having reiterated that it is appropriate to host interconnect technologies in our document when they are used by CPUs, Servesh asks what to do about Infinibad and Intel Omnipath. It is not obvious if they should be covered by us or by the networking subgroup. They are used for node-to-node communication e.g. in supercomputers. The fact that some Intel CPUs may have an Omnipath controller on-die is not a good discriminator (they can have also an Ethernet controller). We will make sure that either we or the networking subgroup are covering them.

We should also mention somewhere the (lack of) future of Xeon Phi.

The next meeting will be the 8th of March at 16:00.

There are minutes attached to this event. Show them.