llrCUDA - Page 3

Ralf Recker · 2011-01-09, 20:17

Quote:

Originally Posted by em99010pepe

For

GPU: GTX470 TDP of 215 W
CPU: Q9550 TDP of 121 W ( 4 cores at 3.6 GHz)
CPU timing for 5*2^1282755+1 of 1056 secs

Then

the GPU client should be 7.1 times faster than the CPU version to achieve the same relation Watts/candidates of CPU LLR client in 24 hours. This is the optimal for this pair, Q9550 and GTX470.

At PrimeGrid's CPU/GPU sieves (PPS Sieve / CW Sieve) the 460 is faster than 12 Q9550s (48 cores).

em99010pepe · 2011-01-09, 20:31

Quote:

Originally Posted by Ralf Recker

At PrimeGrid's GPU sieves (PPS Sieve / CW Sieve) the 460 is faster than 12 Q9550s (48 cores)...

One thing is to be faster other is to be better energy efficient, ok? If your statement is true then the GTX 460 is more efficient on sieving than 12 Q9550's.
For the calculations above I didn't consider if the GPU client uses some CPU power, I don't know if it is the case. If so and considering the investment of a machine with GPU, probably the GPU client would need to be more than 10x faster than the CPU client.

Anyway, next test should be at n=3M, like 7*2^3015762+1, also prime.

Karl M Johnson · 2011-01-09, 20:43

It's also good to compare GPU's vs CPU's computational performance on perf/$.
Oh, and there are motherboards with 8 PCI-E slots.
I didnt encounter a MB with 8 cpu sockets

em99010pepe · 2011-01-09, 20:51

Quote:

Originally Posted by Karl M Johnson

It's also good to compare GPU's vs CPU's computational performance on perf/$.
Oh, and there are motherboards with 8 PCI-E slots.
I didnt encounter a MB with 8 cpu sockets

You can have a motherboard with 8 PCI-E slots but you can also have quad socket motherboards, each socket can take 4 to 6 cores CPU (Tyan motherboars with Xeon processors). You need to compare specific energy, perf/$ is dependent of the latter.

I think msft needs to concentrate on having a GPU client at least 10x faster than a CPU client. Then we need to know if the ratio stands as we increase the size of the tests (higher n for k*2^n+1).

Ken_g6 · 2011-01-09, 23:42

Very nice work, Shoichiro! Thank you!

I've been going over the code, and so far I've made a few optimizations with double2 that improve the speed of Ralf's test on my GTX 460/768@750 from about .84 to about .82 ms/bit. But now I'm looking at the normalization kernels.

First, I'm wondering about the error checking. Does maxerr really have to be a double? Or can it be just a float?

msft · 2011-01-09, 23:54

Quote:

Originally Posted by Ken_g6

First, I'm wondering about the error checking. Does maxerr really have to be a double? Or can it be just a float?

I think float have Sufficient accuracy.

msft · 2011-01-10, 02:18

Support RE64. Only 2 line change. Wait Ken_g6's work.

Ken_g6 · 2011-01-10, 04:07

5*2^1282755+1 is prime! Time : 998.783 sec. .765 s/bit, down from .84 or so.

Changes from v0.12 also included.

That seems to be about as far as blind tweaking will take me. I've been unable to get cudaprof nee computeprof to run the app, which is why I'm tweaking blind. If someone could give me a ranking of the most time-costly kernels, I might be able to do more. Edit: Looks like I hit pretty much all the bases, so probably not.

Also, cuda_normalize2_kernel just bugs me, because it looks like it's an entire kernel using only one thread. But since I don't understand the algorithm well enough - I don't even see how the split between #2 and #3 works - I don't see a way to fix it.

msft · 2011-01-10, 04:29

Quote:

Originally Posted by Ken_g6

Also, cuda_normalize2_kernel just bugs me, because it looks like it's an entire kernel using only one thread. But since I don't understand the algorithm well enough - I don't even see how the split between #2 and #3 works - I don't see a way to fix it.

I concern index conflict with x[].
I can not understand "wrapindex".

Ralf Recker · 2011-01-10, 05:16

A first glance at v0.13:

ralf@quadriga ~/llrcuda.0.13 $ time ./llrCUDA -q5*2^1282755+1 -d
Starting Proth prime test of 5*2^1282755+1, FFTLEN = 131072 ; a = 3
5*2^1282755+1, bit: 80000 / 1282757 [6.23%]. Time per bit: 0.722 ms

Around 11% faster than v0.11...

Ralf Recker · 2011-01-10, 05:30

ralf@quadriga ~/llrcuda.0.13 $ time ./llrCUDA -q5*2^1282755+1 -d
Starting Proth prime test of 5*2^1282755+1, FFTLEN = 131072 ; a = 3
5*2^1282755+1 is prime! Time : 935.885 sec.

real 15m36.000s
user 4m27.065s
sys 6m40.837s

2011-01-10, 05:16	#32
Ralf Recker Oct 2010 191 Posts	A first glance at v0.13: ralf@quadriga ~/llrcuda.0.13 $ time ./llrCUDA -q52^1282755+1 -d Starting Proth prime test of 52^1282755+1, FFTLEN = 131072 ; a = 3 52^1282755+1, bit: 80000 / 1282757 [6.23%]. Time per bit: 0.722 ms Around 11% faster than v0.11... Last fiddled with by Ralf Recker on 2011-01-10 at 05:20*

2011-01-10, 05:30	#33
Ralf Recker Oct 2010 191 Posts	Nearly 2 minutes faster ralf@quadriga ~/llrcuda.0.13 $ time ./llrCUDA -q52^1282755+1 -d Starting Proth prime test of 52^1282755+1, FFTLEN = 131072 ; a = 3 52^1282755+1 is prime! Time : 935.885 sec. real 15m36.000s user 4m27.065s sys 6m40.837s Last fiddled with by Ralf Recker on 2011-01-10 at 05:31*

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
LLRcuda	shanecruise	Riesel Prime Search	8	2014-09-16 02:09
LLRCUDA - getting it to work	diep	GPU Computing	1	2013-10-02 12:12

2011-01-09, 20:43	#25
Karl M Johnson Mar 2010 110011011₂ Posts	It's also good to compare GPU's vs CPU's computational performance on perf/$. Oh, and there are motherboards with 8 PCI-E slots. I didnt encounter a MB with 8 cpu sockets

2011-01-09, 23:42	#27
Ken_g6 Jan 2005 Caught in a sieve 5·79 Posts	Very nice work, Shoichiro! Thank you! I've been going over the code, and so far I've made a few optimizations with double2 that improve the speed of Ralf's test on my GTX 460/768@750 from about .84 to about .82 ms/bit. But now I'm looking at the normalization kernels. First, I'm wondering about the error checking. Does maxerr really have to be a double? Or can it be just a float?