Hi,
I was wondering if there are plans to support some of the Huawei Ascend product line? Here in Europe, the call to government-paid institutions to diversify to non-US hardware has intensified over the last few months. China-based GPU’s appear to be a short-term solution to align with these policies.
It would be nice if I could also use Mojo, should my local institution decide to buy some of that hardware.
Kind regards,
Hylke
The standard Huawei hardware on NPUs and GPUs are written on Ascend C.
Yesterday on the Zoom meeting you should have seen how Tatiana handled the case for HIP and CUDA in Nvidia H100 and AMD MI300A.You are right to point out this as we normally phrase it that “Mojo can be one language for all GPU programming”. We’ll climb into that definitely i suggest before version 1.0 - on our current phase.
Taking the feedback on who should handle this is the person who demonstrated that MAX can outperform HIP and CUDA.
Tatiana’s work focuses on making Mojo the “one language to rule them all” by utilizing MLIR. The suggestion that MAX could outperform Ascend C is based on the same pattern we saw with HIP (AMD) and CUDA (NVIDIA):
Removing the “Vendor Tax”: Vendor-specific languages (like CUDA or Ascend C) are often tied to specific hardware quirks.
Global Optimization: Because the MAX Graph Compiler sees the entire neural network at once, it can perform “kernel fusion” and memory planning that a manual C-style coder writing individual operators in Ascend C might miss.
The Goal: Just as Mojo has shown it can reach 99% of CUDA’s performance (and sometimes beat it in memory-bound tasks), the goal for the Huawei stack would be to let you write in Mojo and have MAX generate code that is actually more efficient than hand-written Ascend C code.
Let’s make this effort Hylke this are the kind of things that prove the power of Mojo in Domain Agnosticism as it is a “systems language “ . I’m very sure the GPU programming leaders will rock on this. If anyone got some skepticism DM or comment. How do you see this Hylke.