gpuOwL: an OpenCL program for Mersenne primality testing

[LaurV]

Originally Posted by tServo

I would think that a Radeon Pro VII is pretty good AFA "bang for the buck":
It has 6.4 F64 Tflops

Won't scale. The Radeon Vii (classic) is memory band limited. James has already scores for the Pro, which are just a little bit better than the classic.


Edit: Whoa, I completely missed one page, but that doesn't change my answer. The Pro won't be two times faster than a classic, by no means. The limitation is easy to see: I run 2 Vii's classic, modified to water, and if I will OC them 20 to 30%, I'll see no much decrease in iteration time, but if I put the mem to 20% higher (from 1G to 1.2G), I see (about) 20% speed increase (but losing the "safety").

BTW, @MattL, what are your GPU/Mem clock scores, for those 640us? (I get 650 with a classic, air cooled, factory OC at ~1G7, mem at 1G0, for currently running M84321739; temperatures in the 90's, I try to keep them under 100°C)

[Zenzoma]

Originally Posted by Xyzzy

What are the temperatures like? Do you have a GPU-Z screen shot to share?

Here's the stats running PRP on M107700431.

[Zenzoma]

Originally Posted by LaurV

BTW, @MattL, what are your GPU/Mem clock scores, for those 640us? (I get 650 with a classic, air cooled, factory OC at ~1G7, mem at 1G0, for currently running M84321739; temperatures in the 90's, I try to keep them under 100°C)

I'll get those GPU-Z screenshots on the 640us exponent later this afternoon once the currently running PRP is done.

[M344587487]

I'm talking to someone who has a 6900XT and is willing to do some quick tests with gpuowl if they manage to get OpenCL and gpuowl working. What quick tests would be ideal? So far I've requested

gpuowl -prp 77936867 -iters 200000
 gpuowl -prp 57885161 -iters 200000

to get data for moebius and James. Any other switches that should be tested? "-carry short"? I don't want to bother them too much so nothing too involved please.

[moebius]

Originally Posted by M344587487

Any other switches that should be tested? "-carry short"? I don't want to bother them too much so nothing too involved please.

With this command line I can get about 15% out of some 16GB Nvidia cards at PRP under Linux.

gpuowl -carry short -use CARRY32,ORIG_SLOWTRIG,IN_WG=128,IN_SIZEX=16,IN_SPACING=4,OUT_WG=128,OUT_SIZEX=16,OUT_SPACING=4 -nospin -block 100 -maxAlloc 10000 -B1 750000 -rB2 20  -prp 77936867

[Zenzoma]

Originally Posted by LaurV

BTW, @MattL, what are your GPU/Mem clock scores, for those 640us? (I get 650 with a classic, air cooled, factory OC at ~1G7, mem at 1G0, for currently running M84321739; temperatures in the 90's, I try to keep them under 100°C)

This is a short log and the temperature data screenshot for it. It's air cooled.

I looked for a Radeon Pro VII water block but couldn't find one. Bykski reviewed the pictures I took of the pcb for them. They said it has mount holes in different locations than the standard Radeon VII. So their block won't work. They're not interested in making a block for a low volume Pro type card.

2020-12-23 14:02:10 GpuOwl VERSION v7.2-21-g28dbf88
2020-12-23 14:02:10 Note: not found 'config.txt'
2020-12-23 14:02:10 config: -maxAlloc 15G -prp 77936867 
2020-12-23 14:02:10 device 0, unique id ''
2020-12-23 14:02:11 gfx906-0 77936867 FFT: 4M 1K:8:256 (18.58 bpw)
2020-12-23 14:02:11 gfx906-0 77936867 OpenCL args "-DEXP=77936867u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=8u -DAMDGPU=1 -DMM_CHAIN=1u -DMM2_CHAIN=2u -DMAX_ACCURACY=1 -DWEIGHT_STEP_MINUS_1=0.33644726404543274 -DIWEIGHT_STEP_MINUS_1=-0.25174750481886216 -DIWEIGHTS={0,-0.25174750481886216,-0.44011820345520131,-0.16213409745771243,-0.37306474779553728,-0.061788266441989627,-0.29798072935699788,-0.47471232907613115,-0.21390437908665341,-0.41180199020062258,-0.11975874301407295,-0.3413572830988989,-0.014337887291734644,-0.26247586476052853,-0.44814572555075455,-0.17414732433395128,}  -cl-std=CL2.0 -cl-finite-math-only "
2020-12-23 14:02:11 gfx906-0 77936867 ASM compilation failed, retrying compilation using NO_ASM
2020-12-23 14:02:13 gfx906-0 77936867 OpenCL compilation in 2.28 s
2020-12-23 14:02:13 gfx906-0 77936867 maxAlloc: 15.0 GB
2020-12-23 14:02:13 gfx906-0 77936867 P1(0) 0 bits
2020-12-23 14:02:13 gfx906-0 77936867 OK    400000 on-load: blockSize 400, c03f94396a5aa29e
2020-12-23 14:02:13 gfx906-0 77936867 validating proof residues for power 8
2020-12-23 14:02:13 gfx906-0 77936867 Proof using power 8
2020-12-23 14:02:14 gfx906-0 77936867 OK    400800   0.51% 1410fa502bb3a765  618 us/it + check 0.28s + save 0.09s; ETA 13:19
2020-12-23 14:02:20 gfx906-0 77936867       410000   0.53% bf44242560060429  617 us/it
2020-12-23 14:02:26 gfx906-0 77936867       420000   0.54% 7f656173fb521927  617 us/it
2020-12-23 14:02:32 gfx906-0 77936867       430000   0.55% b6c7618e60b71bb7  618 us/it
2020-12-23 14:02:38 gfx906-0 77936867       440000   0.56% 200cbc3b887fcfb2  619 us/it
2020-12-23 14:02:45 gfx906-0 77936867       450000   0.58% 4811fce58ccc9cab  619 us/it
2020-12-23 14:02:51 gfx906-0 77936867       460000   0.59% 15cb337858fe7eb1  619 us/it
2020-12-23 14:02:57 gfx906-0 77936867       470000   0.60% ad96b1f48c8bf011  619 us/it
2020-12-23 14:03:03 gfx906-0 77936867       480000   0.62% 93e184ebad6d3cd4  620 us/it
2020-12-23 14:03:09 gfx906-0 77936867       490000   0.63% 70160baec6378071  620 us/it
2020-12-23 14:03:16 gfx906-0 77936867       500000   0.64% 591eecd8448042ad  620 us/it
2020-12-23 14:03:22 gfx906-0 77936867       510000   0.65% 8afd187213816739  620 us/it
2020-12-23 14:03:28 gfx906-0 77936867       520000   0.67% 74e993308f33ac5b  620 us/it
2020-12-23 14:03:34 gfx906-0 77936867       530000   0.68% 57c0f9c504186096  620 us/it
2020-12-23 14:03:40 gfx906-0 77936867       540000   0.69% bcb42100a7c391ad  620 us/it
2020-12-23 14:03:47 gfx906-0 77936867       550000   0.71% ff6f9c39e0347941  620 us/it
2020-12-23 14:03:53 gfx906-0 77936867       560000   0.72% 9a740c005539ec84  620 us/it
2020-12-23 14:03:59 gfx906-0 77936867       570000   0.73% a82132fa0e95b673  622 us/it
2020-12-23 14:04:05 gfx906-0 77936867       580000   0.74% 200fc0c1347e2854  623 us/it
2020-12-23 14:04:11 gfx906-0 77936867       590000   0.76% 48edfb50a88114d1  620 us/it
2020-12-23 14:04:18 gfx906-0 77936867 OK    600000   0.77% b9decd65ca71b629  620 us/it + check 0.28s + save 0.09s; ETA 13:19

[ewmayer]

MattL, how many watts-at-wall (above system baseline) is the system pulling with the R7 Pro running?

[moebius]

Originally Posted by MattL

This is a short log and the temperature data screenshot for it. It's air cooled.

Your fans only run at 51% under 99% load , with my Vega 64 I can let the fans rotate up to 85% at an acceptable volume.The ambient temperature also has an influence. The card then also runs faster.
GPU-Z Values

[Zenzoma]

Originally Posted by ewmayer

MattL, how many watts-at-wall (above system baseline) is the system pulling with the R7 Pro running?

I have an AC power monitor coming Sunday. I'll let you know.

[Zenzoma]

Originally Posted by moebius

Your fans only run at 51%, with my Vega 64 I can let the fans rotate up to 85% at an acceptable volume. The card then also runs faster.
GPU-Z Values

Good idea. I moved that system to the basement recently so the noise doesn't matter. I had it in my office for the initial build and Radeon Pro VII testing and it was loud under PRP load! It's a blower style design.

[BigDukeSix]

Hi.
Are there any recommendations for a bored nVidia RTX3080? TF or PRP?