[Gordon] The A100s have different amounts of memory, how much did CERN get?
A. 40GB AFAICR.
[Gordon] Will partitions be used or can you get a whole card?
A. It will be a mixed layout, depending on request. One limitation is that one cannot re-partiion “live” (have to stop workloads).
[Attila] How are the partitions visible to the user? Does this look like a single GPU?
A. A full GPU would be 1 PCI device; partitioned cards appear as additional PCI devices. It’s one PCI device per partition (even the fat ones). Limitation is that only certain splits are allowed (the horizontal lines in the diagram).
[Attila] The current drivers (470) are now a bit old. Is there a problem with having the latest versions all the time?
A. If you have a VM you can use your own drivers, at whichever version. The batch version is managed by the batch system, etc. There was not a big push to update the drivers - more for the CUDA version.
[Antonio] Instabilities with vGPUs - what are they?
A. Started to see kernel freezes on PCI passthrough for some workloads. Have contacted Nvida support, but not so predictive to date. Freeze kills all workloads on that physical machines. Hence reluctance to make this general.
[Antonio] ATS-IT indico is closed.
A. Ricardo will check this.
[Charles] Can you request systems with multiple instances? To test multi-GPU workloads?
A. Yes, subject to pre-partitioning, we can give multiple slots to a single “workload”.
Useful links:
[Attila] I took a look at the code - the scripts don’t specify any optimisation levels. It uses default. The Intel compiler is much more aggressive ‘out of the box’ than other compilers. For host code some optimisations are enabled.
A. CUDA is maybe overeager with loop unrolling, there is more branchy code. Yes, will look at compiler flags. However, experience was the default DPC++ flags were not that aggressive.
[Attila] managed to get very similar performance with SYCL, but never got code running faster in SYCL.
[Attila] DPCT is an Intel product. DId Codeplay help develop this tool?
A. No, we didn’t develop this. Worked on DPC++.
[Attila] WIth the Intel acquisition of Codeplay, is there a plan for ComputeCPP?
A. All very fresh, so can’t really say anything yet. ComputeCPP has a lot of good features.
[Attila] Would really like to be able to use DPC++ with plugins, so that we don’t have to recompile the compiler.
A. This is a tricky thing. Some gotchas are old CL headers in CUDA, can be a real pain to get rid of.
[Charles] How does SYCL handle atomic when the backend is a multi-core system (like 70 CPU cores)?
A. Not completely sure - there are software fallbacks.
[Attila] this would be in the openCL cpu driver?