|
1)
Message boards :
Number crunching :
Possible support for Intel iGPU Raptorlake?
(Message 2075)
Posted 23 May 2025 by ahorek's team Post: You can enable it through an anonymous platform if you want to try it. The iGPU should technically work, but performance with this app will be extremely poor, potentially around a day per work unit. Worse yet, it might throttle your CPU tasks due to thermal constraints, especially on a Laptop. It's better to keep the iGPU idle or run Einstein@Home instead, their Arecibo app performs better on non-NVIDIA GPUs. Even then, using the iGPU will very likely reduce your overall system performance, so it’s often not worth using it at all. |
|
2)
Message boards :
Number crunching :
Android support?
(Message 2071)
Posted 15 May 2025 by ahorek's team Post: ok, it took about 3 days :D https://sech.me/boinc/Amicable/result.php?resultid=100264226 It technically works, but since even low-end GPUs can complete a work unit in about 10 minutes, it's not really practical. The only project that runs reasonably well on mobile GPUs is Einstein@Home. PrimeGrid and Amicable use 32-bit integers, which perform terribly, at least on Adreno GPUs. Mali might handle them better, but even then, we're still looking at work units that take days to complete. |
|
3)
Message boards :
Number crunching :
CPU VS GPU WU’s - massive completion time difference, why?
(Message 2070)
Posted 12 May 2025 by ahorek's team Post: The CPU application is outdated and doesn't fully leverage SIMD capabilities, whereas the algorithm is more GPU-friendly. For math-intensive and parallelizable workloads, GPUs deliver substantially better performance, as they are specifically designed for this type of computation. I recommend allocating your CPUs to projects that either lack a GPU version or where the GPU version isn't as efficient as the CPU counterpart... |
|
4)
Message boards :
Number crunching :
Android support?
(Message 2069)
Posted 12 May 2025 by ahorek's team Post: I had to rewrite the code a little because Android does not support named semaphores:
NOINLINE Semaphore::Semaphore()
{
if (sem_init(&mySemaphoreInternal, 0, 0) != 0)
{
std::cerr << "Failed to initialize unnamed semaphore, errno " << errno << std::endl;
boinc_finish(-1);
}
}
class Semaphore
{
public:
Semaphore();
FORCEINLINE ~Semaphore() { sem_destroy(&mySemaphoreInternal); }
FORCEINLINE bool Signal() { return sem_post(&mySemaphoreInternal) == 0; }
FORCEINLINE bool Wait()
{
while (sem_wait(&mySemaphoreInternal))
{
if (errno == EINTR)
{
errno = 0;
}
else
{
return false;
}
}
return true;
}
private:
sem_t mySemaphoreInternal;
};
|
|
5)
Message boards :
Number crunching :
Android support?
(Message 2068)
Posted 12 May 2025 by ahorek's team Post: I'm currently testing the Android version on Ardeno 750 (which is the 2nd fastest mobile GPU). Even if it functions correctly, the performance appears to be VERY bad > The GPU side, my phone is 450 GFLOPS but only OpenCL 2.0 and my 4060 Ti is 22,100 GFLOPS or 49 times faster. I hope this is the correct way to think about things but if it is, then my phone's GPU would do a WU in about 4 hours or 6 WU per day. That's doable but again the heat issue is what I would be afraid of. You can't do that, because there are other factors: 1/ Mobile GPUs use shared memory, which significantly limits memory bandwidth compared to a desktop GPU like the 4060 that has dedicated VRAM. 288.0 GB/s for 4060 Ti vs ~50 GB/s DDR5 (also shared with the CPU) 2/ NVIDIA GPUs have great INT32 performance, whereas mobile GPUs often emulate INT32 operations to conserve power, since these operations are generally relevant for gaming workloads. Even if FP32 performance numbers look good, INT32 performance is much worse. 3/ The cache sizes on mobile GPUs are very small (if any), which forces the app to rely on system memory. The system memory already suffers from limited bandwidth and higher latencies. 4/ In general, code isn't optimized for mobile GPUs. Achieving good performance often requires various tricks, like avoiding certain features or adapting to architectural quirks. It's challenging because these GPUs have very limited resources, and no developers go to that level of optimization. Especially when it takes around 100 top-end phones to match the performance of a single 4060 Ti. It simply doesn't make sense. |
©2025 Sergei Chernykh