Hacker News new | past | comments | ask | show | jobs | submit login

Intel has some great software like MKL and related libraries, but, speaking from experience, due to its internal politics Intel is unable to innovate in a meaningful way or build a reasonable SW strategy for GPUs.

The oneAPI heavily contributed to their GPU failures IMO. Choosing SYCL and promoting non-existing performance portability and pouring resources in the trying to shove CPU and GPU (and at some point FPGA!) into the same programming model while disinvesting from OpenCL was a big mistake, but it was probably the only way to please both CPU and GPU management.

And the problem is not SYCL per se but the fact that they try to make it an open standard rather then trying to provide a way to extract maximum performance from their HW. They needed to build a CUDA alternative, where, after jumping through considerable hoops, you can get to actually programming tensor cores for those who need it. Not surprisingly, Intel is not using SYCL for their GPU kernels (look at oneDNN sources). With Raja gone there may be hope, but I'm not holding my breath.

The only thing that they have that is competitive is Habana's stuff which is outside of the oneAPI.

I would not be surprised if AMD has a similar issue of infighting due selling both CPUs and GPUs.

And NVIDIA does not have this problem.




What are your thoughts on the move from ICC to the LLVM-based ICX? Does that fit in with this product strategy somehow?




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: