Seems that works on Zluda has restarted.


(as long as Nvidia allows it)

and CUDA was always faster on Nvidia.

now that AMD can run CUDA, even in this alpha state, CUDA ends up faster again.

Well....
https://www.phoronix.com/news/NVIDIA-Vulkan-AI-ML-Success

cool, but this has nothing to do with GPGPU compute or the types of compute that BOINC projects use.

AI/ML loads are just a bunch of matrix computations over and over. real world scientific compute is a lot more than that.

CUDA will reign supreme over OpenCL. projects aren't going to migrate to Vulkan for this stuff.

---

> just a bunch of matrix computations over and over. real world scientific compute is a lot more than that.
Matrix computations are frequently used in numerous algorithms, not just AI, but AI algorithms prioritize efficiency over high precision, so they use lower-precision number formats that store only a limited set of values. For example:1.0, 1.25, 1.5, 1,75, 2.0 etc. However, using these types for scientific computations will significantly reduce the accuracy of results. Those extensions are designed for a specific use case, accelerating AI algorithms even further, nothing else.

Vulkan is capable of performing the same tasks as CUDA or OpenCL with high or low precision, but it has never gained widespread popularity, at least for BOINC projects.

that's what i meant to say. low precision tensors.

---