Chris Elrod

But it varies by computer! I'll try again on the Haswell laptop before I issue a release, but performance has always been bad there.

Sure, I'll try. Note that even with 2000 GFLOPS, it'd take almost six and a half minutes: ```julia julia> 2e-9 * 73_000^3 / 2000 389.017 ``` My computer started overheating,...


This was on my 7980XE, which is clocked lower than my 10980XE was (which was featured in earlier benchmarks), in case anyone notices the GFLOPS are generally a bit lower...

https://github.com/chriselrod/BLASBenchmarksCPU.jl The docs also have a section on disabling turbo on Linux. I don't know other OSes, but it's definitely recommend for the sake of consistent multithreaded benchmarks. Another option...

Multithreaded x86 is failing. Note it is failing on master as well: https://github.com/JuliaLinearAlgebra/Octavian.jl/runs/8240502733?check_suite_focus=true

Would anyone complain if we just abandon 32 bit support? @DilumAluthge I'm not inclined to spend any of my free time maintaining it. Either Octavian will not be compatible with...

It's worth exploring if some sort of repacking could improve performance, in particular for large batches of small matrices.

There are a lot of branches based on the size of the arrays. Ideally, we'd be able to enable coverage for our tests over different array sizes. For the 20x20...

Two things that should be simple and help with this file: 1. Add a test dependency on something like StaticArrays.jl (or StrideArrays after it is registered). Testing with mutable statically...