I was curious hows the work going ?
Hi, was there anything in particular you were interested in? I hope that I can help you if you provide a bit more info ![]()
I’m interested in writing the low level kernels i’ve done work with blas,Trilinos, openmpi and similar low level programming on freebsd before. I know everything is wrapped around fp32 or migrating to dual path fp4 but I think better software can make fp64 perform better, especially since the vram is a better value here.
Are you guys looking for developers who can solve tough challenges?
I think the best way is contributing rocm-libraries to apply these skills. I’m trying to do it, we might try it together if you’re interested in it.
imao, it might be valid for deeptech or academy tasks
They don’t respond on github, so where do is development community active? Do you need to be invited in? I have issue on rocm that’s very old with no response.
Will AMD set some unrealistic performance targets to motivate developers to push the limits?
I’ve worked with Blas, Trilinos, OpenMPI, and other low-level programming on FreebSD, and I’m interested in developing the low-level kernels. Although I am aware that everything revolves around fp32 or switching to dual path fp4, I believe that fp64 can perform better with better software, especially since the vram is a better value in this case.
I have experience working with BLAS, Trilinos, OpenMPI, and other low-level programming tools on FreeBSD, and I’m particularly interested in developing low-level compute kernels. While I understand that most AI workloads currently focus on FP32 or are moving toward dual-path FP4, I believe FP64 performance still has significant potential with better software optimization. Given the available VRAM, I think FP64 could offer a more efficient balance for certain high-precision workloads.