LL with Raspberry Pi 3B
 

I tested a number on Prime 95 in GIMPS and it said it would take about 8 weeks to test. So instead I started the LL test in Mathematica on a RPi 3B. But tomorrow will be 10 weeks and it is still running a candidate with a large exponent.
- Is the RPi 3B not as fast as the Intel 5i I used on GIMPS to find how long it takes. My RPi is new and the computer I used with the 5i is about 6 years old.

I'm looking at alternatives like Pandalatte and RPi 4, but are these really any faster? Are the newest Intel CPUs and GPUs much faster than single board computers.

Can I get some advice?

1. Prime95 is likely 20 to 50 times faster than Mathematica at LL testing.
2. All 4 cores of the raspberry Pi are likely as fast as one core of a core i5, even 6 years old.
So, you're looking at 100 to 200 times longer to test a candidate, using the software you've chosen.

[QUOTE=nickargenta]Can I get some advice?[/QUOTE]
Get something resembling a modern PC with an Intel core I5/I7 processor. Run prime95 on it.

[QUOTE=nickargenta;523695]I tested a number on Prime 95 in GIMPS and it said it would take about 8 weeks to test. So instead I started the LL test in Mathematica on a RPi 3B. But tomorrow will be 10 weeks and it is still running a candidate with a large exponent.
- Is the RPi 3B not as fast as the Intel 5i I used on GIMPS to find how long it takes. My RPi is new and the computer I used with the 5i is about 6 years old.

I'm looking at alternatives like Pandalatte and RPi 4, but are these really any faster? Are the newest Intel CPUs and GPUs much faster than single board computers.

Can I get some advice?[/QUOTE]
[LIST=1][*]The Raspberry Pi is too slow.[*]The Pi uses 32 bit operating system?[*]Mathematica is slow[*]Prime95/mprime or mlucas are fast.[/LIST]
I have a test running 64 bit mlucas on an Odroid N2, which is 3 times faster than the 3B+. At ~90 million bits it will take 3 months. I suggest you abandon the Mathematica run and run somthing that is fun on your Pi.

The N2 is 3 times quicker than the 3b+. AFAIK there is no 64 bit OS for the 4B+ yet. A single core of an i5/i7 with AVX2/FMA3 is 3-4 times faster than an N2. An AMD Radeon VII is 5-10 times faster than a i7; Which makes it over 500 times as fast as the 3B.

[QUOTE=nickargenta;523695]I tested a number on Prime 95 in GIMPS and it said it would take about 8 weeks to test. So instead I started the LL test in Mathematica on a RPi 3B. But tomorrow will be 10 weeks and it is still running a candidate with a large exponent.
- Is the RPi 3B not as fast as the Intel 5i I used on GIMPS to find how long it takes. My RPi is new and the computer I used with the 5i is about 6 years old.

I'm looking at alternatives like Pandalatte and RPi 4, but are these really any faster? Are the newest Intel CPUs and GPUs much faster than single board computers.

Can I get some advice?[/QUOTE]

Mathematica is probably highly generalized, and not optimized for large numbers and the Lucas-Lehmer test in particular. Sure, it will chug away, but it might take years...

Mlucas is the fastest Lucas-Lehmer test program that you can run on ARM and the Raspberry Pi, or at least I'm not aware of any others that are even close. As a reference, L-L double-check exponents (around M50700000 currently) take about 90 days on a Raspberry Pi 3A+ or 3B+, and about 45 days on a Raspberry Pi 4B. That is with some tweaks like running nothing else on the board and turning the graphical interface off. Other ARM SBCs are likely to be somewhat faster, depending on design choices. For example, the Jetson Nano has faster memory (same LPDDR4 but the memory bus is doubled in width to 64 bits) so the same test takes about 41 days, even with an older core and slightly slower clock rate than the Pi 4B.

The Raspberry Pi boards were designed to fit a price range, and unfortunately that means, that performance-related sacrifices had to be made. Other manufacturers may not be bound by the premise of a $35 computer, and with the latest Pi 4, even they had to let it slide, because only the smallest memory model is still nominally $35.

But Mlucas won't run on the default 32-bit Raspbian distribution, it needs a 64-bit kernel to access all the available power, and in particular, the NEON vector extensions in the ARM core. That's another compromise the Raspberry Pi Foundation has had to make. The software side, for now, needs to stay compatible across the range of boards, not just one target. So the same distribution needs to be able to run on a Raspberry Pi 1 and 4B, so for now, it can't be 64-bit.

And yes, x86 hardware is that much faster. Wide fast memory buses, large caches, fast cores with hyperthreading, highly advanced branch prediction and lots of vector floating point capability.

But I still run LL on a bunch of various ARM boards (total now 24 boards, across 21x RPI3A+, one 3B+, one 4B and one Jetson Nano), not because it makes sense in any way, but because fiddling is fun.

Thanks

So it seems you know a lot about RPi speeds. I understand from your comments that the more important factor for speed is the choice of software. You state some very high speeds for 64 bit (I assume from your comments) on single board. How much faster would this be on x86? You made no comment here.

Can I chat with you privately? Do you use Signal?
Nick

So I understand from your comments that the AMD Radeon is the fastest. Do you have one that can be tested
? I'm interested in platforms at the highest speed

Thanks

[QUOTE=nickargenta;523715]So I understand from your comments that the AMD Radeon is the fastest. Do you have one that can be tested
? I'm interested in platforms at the highest speed[/QUOTE]

The card, although expensive, can be put into any old computer [u]with[/u] a PCIe3 slot and a 750w PSU.

You may PM (private message) me through this message board if you wish to do so.

[QUOTE=nickargenta;523715]So I understand from your comments that the AMD Radeon is the fastest. Do you have one that can be tested
? I'm interested in platforms at the highest speed[/QUOTE]
I recommend looking at this benchmark page if you want to see relative GPU performance for LL:
[url]https://www.mersenne.ca/cudalucas.php[/url]

edit: Radeon VII really excels because it has an unprecedented (for consumer graphics) 1:2 double precision performance (this means the clock ticks needed to perform DP operation compared to SP,single precision) while many others are crippled to 1:32 or so. This coupled with high bandwidth memory makes it really excel at LL.