Discussion DC Tasks number/bit formats

[pandemonium]

Hey guys,

This may be a trivial question so I apologize in advance.

What are most/all/some distributed compute tasks aligned with in terms of performance metrics / number/bit formats? (INT4, INT8, FP32, etc...)

Reading through some of the SETI forums I see a lot of INT8 designations in the formats, but if I were to guess I'd say there's a good mix of a lot of the formats being used depending on the task and projects at large.

Anyways, the reason I ask is if there's a list of tasks (or projects) where we know what format is being utilized more. That way when people are curious about what projects to jump onto maybe they can more effectively contribute to the right tasks. It'd also be nice to know when buying our fancy calculators for a specific project or task.

I have little idea of what I'm talking about and I tried doing some searching, but didn't come up with answers that satiated my question. So here I am. Of course, I may have missed it in the exabytes. /shrug

 

[StefanR5R]

First of all, there is this big differentiation between projects which run only on CPUs, projects which run best on CPUs but have a GPU application too, and ones which have applications for both but run best on GPUs. (I'm not sure if there is any active project which only has a GPU application.) Though your question sounds like it is specifically about GPU applications.

Milkyway@home: FP64 (or on GPUs which don't have good FP64 support: some workarounds with performance hit)

Molecular simulations such as Folding@home and GPUGrid: FP32

PrimeGrid's PPS-Sieve: INT32
PrimeGrid's GFN: ?

This may be no longer true. See link in #6.

 

[pandemonium]

Not particularly about GPU applications, but that does bring up a good question.

I'm obviously dumbing it down and not diving into the wide range of workloads that go on with compute tasks, but was curious if there were any relevance of the format, the throughput of the hardware being utilized, and the ending performance.

 

[Orange Kid]

The simple answer is, yes.
The unfortunate reality is that each project has its own best combination of not only hardware, but also software.
There is no magic combination or one size fits all.
We all have our own personal favorite combos and strive to maximize the performance.
It comes down to the metric you wish to use to rate your performance with. Points, work or contribution to the greater good.

 

[Ken g6]

PrimeGrid's GFN: ?

I'm given to understand that GFN is FP32 as well these days.

 

[StefanR5R]

The GPU port of PrimeGrid's new "Wieferich and Wall-Sun-Sun" (short: "ww"), OCL_cuda_WW and opencl_ati_WW, use INT64. (source)

 

[Markfw]

I can only speak to a few points. F@H works MUCH better on any GPU compared to CPU's. Everything, CPU, GPU, regardless of project seems to work much better under linux.