Hello Asteroids@home Team,

I saw your fantastic progress with the new AVX-512 and ARMv8 applications this year. Given your clear commitment to supporting efficient new hardware, I am writing to request a native OpenCL 1.2+ application for the Qualcomm Adreno iGPU.

The latest Windows on Snapdragon drivers have fixed many of the previous OpenCL shared memory bottlenecks. As more users adopt these efficient devices, having a native Asteroids app would be a major win for the community. I would be happy to help test any 'alpha' builds on my Snapdragon X hardware to help create more competition in the distributed computing space.

A native AArch64 application could potentially offer better efficiency on Windows-on-ARM CPUs than an emulated x64 application. If you have access to real hardware, it would be great to compare the difference.

As for the GPU version, Adreno does not support FP64, which effectively rules out GPU usage for the current application. Apple GPUs and Intel Arc (Alchemist) aren't supported for the same reason.

But both einstein@Home and primegrid support the apple silicon iGPU's, so clearly It's possible.

PrimeGrid applications mainly depend on integer performance
GFN-16 took 20 minutes (with an older driver, it was much slower then using a CPU)

sieving didn't work (wrong results)

Amicable works with a few modifications, but the performance is terrible.

Einstein - BRP4A and BRP4 (it's the same app) don't require FP64 support. It's the only app that works and it's reasonably efficient - 10 minutes / WU

Asteroids, O4MDG, and BRP7 applications require FP64 support. Their calculations are fundamentally different and will never be supported on GPUs that lack this capability.


The Adreno X1-85 might be slightly faster than my Adreno 750, but don’t expect a significant performance improvement.
https://chipsandcheese.com/p/the-snapdragon-x-elites-adreno-igpu

Thanks for clarifying the technical divide between the integer-based workloads in PrimeGrid/Einstein and the specific FP64 requirements for Asteroids@home. I appreciate you taking the time to break down why the architecture currently creates such a hard wall for this specific project.

It is helpful to understand that it’s not just an OpenCL compatibility issue, but a fundamental hardware requirement for the science the project is doing. I will focus my testing on the AArch64 CPU-side optimizations, as that clearly seems to be the most viable path for efficiency on the Snapdragon platform for now.

Thanks again for the insights.

Thanks for clarifying the technical divide between the integer-based workloads in PrimeGrid/Einstein and the specific FP64 requirements for Asteroids@home. I appreciate you taking the time to break down why the architecture currently creates such a hard wall for this specific project.

It is helpful to understand that it’s not just an OpenCL compatibility issue, but a fundamental hardware requirement for the science the project is doing. I will focus my testing on the AArch64 CPU-side optimizations, as that clearly seems to be the most viable path for efficiency on the Snapdragon platform for now.

Thanks again for the insights.

you have literally been told this information about the FP64 requirement many times, even recently. why are you making the rounds on projects bumping months old threads with more AI slop responses?

---