Have you actually confirmed that you are memory bandwidth limited? What happens if you halve the memory speed?
Have you actually confirmed that you are memory bandwidth limited? What happens if you halve the memory speed?Yes.
My hardware has 10 cores capable of two threads each. The CPU is overclocked, typically achieving 5.12 GHz under heavy loads. The board is liquid cooled.
My Windows O/S is set to prioritize applications.
I instructed Mathematica to run 12 parallel kernels, which results in 12 computational subprocesses plus 2 coordinating subprocesses. Each computational subprocess accounted for about 6% of available CPU load until its task was finished. Memory usage rarely exceeded 20% and is no longer a bottleneck. In conjunction with this, the cores ran about 4°C cooler. However, I believe the number of memory channels (2) are still limiting runtime. Keep in mind I've been historically spoiled by supercomputers with hundreds of cores all on the same memory backplane with 1:1 channel to core ratios. Additionally, the number of cores in my system and number of parallel processes permitted by my Mathematica license inhibit faster times to completion -- although the memory channel issue restricts these gains in a less than linear fashion.
I have clocked the start-to-finish time of parallel processes from within each process, and compared them to the start-to-finish times of singleton parallel processes with the same per-process load. The singleton execution times were about 75% of the parallel execution times.Have you actually confirmed that you are memory bandwidth limited?
I haven't checked.What happens if you halve the memory speed?
I agree. Memory use is available by library call to the Mathematica runtime system. It is independent of Windows.The memory usage information can be tricky. ...
That tells you that having more cores/threads isn't scaling linearly, but it doesn't give any hints to why. It could be software or hardware, or both.I have clocked the start-to-finish time of parallel processes from within each process, and compared them to the start-to-finish times of singleton parallel processes with the same per-process load. The singleton execution times were about 75% of the parallel execution times.
This is why I ran many tests with varying parallelism and problem sizes. The latter is three dimensional. There are inflections in the runtime curves. One of them is from singleton to multiple parallel processes.That tells you that having more cores/threads isn't scaling linearly, but it doesn't give any hints to why. It could be software or hardware, or both.
Seems like it would be worthwhile to prove your theory like LightningZ71 suggested. Choose a scenario where you think it is bandwidth limited and lower your memory speed.This is why I ran many tests with varying parallelism and problem sizes. The latter is three dimensional. There are inflections in the runtime curves. One of them is from singleton to multiple parallel processes.
Have you successfully scaled shared-memory parallel computer programs on hardware platforms with 1:1 to 1:n memory channel ratios?Seems like it would be worthwhile to prove your theory like LightningZ71 suggested. Choose a scenario where you think it is bandwidth limited and lower your memory speed.
I have not, I may be missing some context for pinpointing a bottleneck in that regard.Have you successfully scaled shared-memory parallel computer programs on hardware platforms with 1:1 to 1:n memory channel ratios?
In a toy benchmark, that may or may not correlate with the problem at hand But more importantly it would be worth to check how Mathematica runs on WoA. According to the support page there is no official WoA version. So it doesn't really matter how much faster X Elite is in geekbench if it will need to go through emulation layer in WindowsSorry if this makes you feel bad but this is a frickin' $900 devkit embarrassing your huge desktop PC: https://browser.geekbench.com/v6/cpu/compare/8010211?baseline=8143535
It's basically a large heatsink so just put it in the path of cool airflow from an airconditioner(Also, cooling constrained form factor.)
Up to 23.8x faster basecalling for DNA sequencing in Oxford Nanopore MinKNOW when compared to the 16-inch MacBook Pro with Core i9, and up to 1.8x faster when compared to the 16-inch MacBook Pro with M1 Pro.
Nice trolling attempt. ;-) Definitely doesn't sound like CPU processing, but as if one of the accelerators is used. The GPU maybe?Apple said:Up to 23.8x faster
Why can't it be multithreaded processing with SME?Nice trolling attempt. ;-) Definitely doesn't sound like CPU processing, but as if one of the accelerators is used. The GPU maybe?
Against which baseline on i9? Singlethreaded x87?Why can't it be multithreaded processing with SME?
Have you tried running anything on even the base M1? It is lightning fast in certain workloads.Against which baseline on i9? Singlethreaded x87?
None of the computing intensive software which I use at work, and almost none of which I use at home runs natively on it. For different reasons, there just happens to be no intent to port the software. — But the software which the OP uses has got a macOS-on-Apple-Silicon port. We lack performance profiling insight though to make good guesses how it would perform there, and if it would be a viable platform for the OP's longer term goal of scaling the dataset size up.Have you tried running anything on even the base M1?
Out of curiosity, is any of the SW you use available on Intel macOS (I mean recent versions)?None of the computing intensive software which I use at work, and almost none of which I use at home runs natively on it. For different reasons, there just happens to be no intent to port the software.
At my work: Semi-OT, because that's shared memory parallel computing too but mostly floating point, solving of systems of linear equations, and miscellaneous postprocessing.Out of curiosity, is any of the SW you use available on Intel macOS (I mean recent versions)?