Asteroids@home has long been a favorite for GPU crunching, with solid support for NVIDIA and AMD via OpenCL and CUDA.

However, there is a massive untapped resource being ignored: the Apple Silicon iGPU.

For a long time, the excuse for not supporting the M-series GPU was a lack of native FP64 or the deprecation of OpenCL on macOS.

But in 2026, those excuses are no longer valid.

Projects like Einstein@Home and PrimeGrid have already proven that the Apple Silicon iGPU is more than capable of handling high-precision scientific math with incredible efficiency.

On my base M4 Mac Mini, I am currently running 10 concurrent tasks on the iGPU for other projects with perfect stability.

Apple Silicon’s unified memory architecture offers a unique advantage for memory-intensive projects like Asteroids, allowing the GPU to access huge amounts of data without the traditional PCIe bottleneck.

By refusing to release an arm64-apple-darwin GPU application, Asteroids@home is leaving thousands of highly efficient M1, M2, M3, and M4 nodes on the table.

We have the hardware, and other projects have already provided the roadmap.

It is time to bring Asteroids@home into the modern era of Green Crunching on Apple Silicon.

When can the community expect a test version for the Mac iGPU?

> Projects like Einstein@Home and PrimeGrid have already proven that the Apple Silicon iGPU is more than capable of handling high-precision scientific math with incredible efficiency.

This is a misleading claim...

EaH has the O4MDG and BRP7 applications, which require FP64 support to operate. Apple Silicon iGPU isn't supported.

On the other hand, EaH also includes the BRP4 and BRP4A applications, which do not require FP64 and therefore allow the use of GPUs that lack FP64 support.

PrimeGrid applications use INT64 math, which IS supported by Apple Silicon iGPUs.


The only way to support Apple Silicon iGPU on Asteroids would be to emulate FP64 with FP32. It is theoretically possible, but incredibly inefficient because each FP64 operation requires ~4-12 FP32 operations (it's not 2x as you might expect). That's a lot of work for making an app with disappointing performance for a very few devices. The fact that PrimeGrid apps run well does not guarantee similar performance for other applications, as each app performs different types of work.

For these reasons, it’s highly unlikely that an Asteroids app will ever be released for Apple Silicon iGPUs, unless Apple adds FP64 support or a completely new app is developed that doesn’t rely on this feature.

I appreciate the technical breakdown regarding native FP64 support. You are correct that the M-series lacks dedicated hardware double-precision units, but framing this as an impossible barrier for 2026 is a bit of a legacy viewpoint.

While PrimeGrid does leverage INT64, its success proves the stability and efficiency of the M-series iGPU for high-precision math. On my M4 Mac Mini, the iGPU handles these intense workloads with a fraction of the heat and noise of the CPU, finishing tasks 9x faster than the CPU cores for the same power draw.

Regarding the "inefficiency" of FP64 emulation: even with a 4-12x penalty for emulating double precision via FP32, the parallel nature of the Apple GPU still offers a massive net gain over CPU crunching. Modern libraries like libMetalFloat64 have already demonstrated that "hardware double precision is not needed" to achieve competitive performance on Apple Silicon.

By dismissing an Asteroids app based on 2018 hardware assumptions, we are ignoring the reality of 2026: Apple Silicon is the most efficient global compute resource we have. If projects like Einstein@Home and PrimeGrid have found the roadmap, it's time for Asteroids to stop looking for excuses and start looking at emulation or Metal-optimized alternatives.

Science shouldn't wait for "perfect" hardware when we have "great" hardware sitting idle right now.

I understand that you’re excited about how the Mac iGPU performs on other projects, but the Asteroids GPU app isn’t particularly efficient on GPUs in general. If you add a significant emulation overhead, the app will likely be slower and less efficient than a CPU version. The expectation that it'll be much faster and super efficient is simply false, and you'll be disappointed.
Of course, someone would have to rewrite the app this way first, which would be a significant effort. It’s not just a matter of "enabling it" for Apple GPUs, which is all PrimeGrid had to do, since their app doesn't depend on FP64 math.

The app is open source https://github.com/AsteroidsAtHome/PeriodSearch/tree/dev/period_search_opencl_amd so if anyone wants to make it 9x faster, rewrite it for Metal and FP32, feel free to do so. Don't make excuses, send a pull request :)

The Asteroids gpu app is much less productive compared to the cpu app when deployed across a multi-core cpu. Why most of us don't bother with the gpu app and just run the cpu app.

---

A proud member of the OFA (Old Farts Association)