gpuOwL: an OpenCL program for Mersenne primality testing - Page 229

Viliam Furik · 2020-10-21, 11:17

Quote:

Originally Posted by preda

In general I would recommend a smaller -maxAlloc, e.g. "-maxAlloc 18G" (just to be sure you don't overfill the GPU RAM which would produce a slowdown), but in this case you ran without P-1 so that doesn't matter.

RTX 3090 has 24 GB of VRAM. I think -maxAlloc 18G is ok... (Unless I don't understand, how maxAlloc works.)

aheeffer · 2020-10-21, 13:18

Thanks a lot indeed! Now we finally know. With 619µsec for this exponent, the Radeon VII remains the king. Even underclocked and running at 150 W, the Radeon VII is still twice as fast as the RTX 3090. I will cancel my order for the RTX 3080 now.and hope for the best with the new RX 6900 XT.

ATH · 2020-10-21, 13:52

It is no surprise really: https://en.wikipedia.org/wiki/GeForce_30_series

Gpuowl and CUDALucas uses "double precision", and 3090 has only 1/32 the double precision performance compared to single precision performance (918 GFLOPS vs 29389 GFLOPS). It has been like that for many years now with Nvidia consumer cards.

With 29389 GFLOPS single precision it should do really well at trial factoring with mfaktc, probably one of the fastest cards for trial factoring including the expensive Tesla cards.

moebius · 2020-10-21, 14:12

Quote:

Originally Posted by aheeffer

With 619µsec for this exponent, the Radeon VII remains the king.

The AMD RX Vega 64 remains the second winner in the consumer card category.

Benchmarks of 77936867 for the following graphics cards would be interesting.
AMD Radeon RX 6900/6800/6700 XT when available
Nvidia GeForce RTX 3080/3070, 3060 TI, 2080 Ti
Nvidia Titan RTX, Titan V
Nvidia A100 SXM4

Viliam Furik · 2020-10-21, 16:08

Quote:

Originally Posted by aheeffer

Thanks a lot indeed! Now we finally know. With 619µsec for this exponent, the Radeon VII remains the king. Even underclocked and running at 150 W, the Radeon VII is still twice as fast as the RTX 3090. I will cancel my order for the RTX 3080 now.and hope for the best with the new RX 6900 XT.

DON'T CANCEL! There is still a chance.

Firstly, it can still be darn fast in TF, if the issue of it not being as-darn-fast-as-expected resolves. (It should be more than twice as fast in TF as RTX 2080Ti, but it is only a bit faster. I guess that there's something wrong with the way the mixed GPU cores (half of all cores) are supposed to work. They are meant to be able to do either INT32 or FP32 (BTW FP32 is what we want for TF), but they don't seem to fill the "or FP32" part of the deal.

Secondly, Preda is researching the possibility of computing the DP (FP64) numbers by expressing them as SP (FP32). It may or may not make it even faster than Radeon VII if working as intended.

Viliam Furik · 2020-10-21, 16:09

Quote:

Originally Posted by moebius

The AMD RX Vega 64 remains the second winner in the consumer card category.

Benchmarks of 77936867 for the following graphics cards would be interesting.
AMD Radeon RX 6900/6800/6700 XT when available
Nvidia GeForce RTX 3080/3070, 3060 TI, 2080 Ti
Nvidia Titan RTX, Titan V
Nvidia A100 SXM4

I am not sure, where did you pull the name 3060Ti from, but I can give you 2080Ti results.

Viliam Furik · 2020-10-21, 16:18

Quote:

Originally Posted by Viliam Furik

RTX 3090 has 24 GB of VRAM. I think -maxAlloc 18G is ok... (Unless I don't understand, how maxAlloc works.)

I have realised that you recommended 18G, against original 20G. Still, I think that's safe. Unless he is not running another VRAM hungry program.

moebius · 2020-10-21, 16:21

Quote:

Originally Posted by Viliam Furik

I am not sure, where did you pull the name 3060Ti from, but I can give you 2080Ti results.

https://www.techradar.com/news/nvidi...an-rtx-3060-ti
Yes, please do so, I'm particularly interested in the comparison of the 2080 TI and the fictitious RTX 3060 TI and the RTX 3070. The forum user xx005fs apparently even has a Titan V, maybe he is reading this thread.

Viliam Furik · 2020-10-21, 16:42

Quote:

Originally Posted by moebius

https://www.techradar.com/news/nvidi...an-rtx-3060-ti
Yes, please do so, I'm particularly interested in the comparison of the 2080 TI and the fictitious RTX 3060 TI and the RTX 3070. The forum user xx005fs apparently even has a Titan V, maybe he is reading this thread.

Code:

2020-10-21 18:22:29 gpuowl v6.11-364-g36f4e2a
2020-10-21 18:22:29 config: -proof 8
2020-10-21 18:22:29 config: -nospin
2020-10-21 18:22:29 config: -iters 300000 -log 10000
2020-10-21 18:22:29 device 0, unique id ''
2020-10-21 18:22:29 GeForce RTX 2080 Ti-0 77936867 FFT: 4M 1K:8:256 (18.58 bpw)
2020-10-21 18:22:29 GeForce RTX 2080 Ti-0 Expected maximum carry32: 583B0000
2020-10-21 18:22:30 GeForce RTX 2080 Ti-0 OpenCL args "-DEXP=77936867u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=8u -DPM1=0 -DMM_CHAIN=1u -DMM2_CHAIN=2u -DMAX_ACCURACY=1 -DWEIGHT_STEP_MINUS_1=0xa.c42d0d7cec038p-5 -DIWEIGHT_STEP_MINUS_1=-0x8.0e50c8817ddf8p-5  -cl-unsafe-math-optimizations -cl-std=CL2.0 -cl-finite-math-only "
2020-10-21 18:22:30 GeForce RTX 2080 Ti-0

2020-10-21 18:22:30 GeForce RTX 2080 Ti-0 OpenCL compilation in 0.01 s
2020-10-21 18:22:31 GeForce RTX 2080 Ti-0 77936867 OK        0 loaded: blockSize 400, 0000000000000003
2020-10-21 18:22:31 GeForce RTX 2080 Ti-0 validating proof residues for power 8
2020-10-21 18:22:31 GeForce RTX 2080 Ti-0 Proof using power 8
2020-10-21 18:22:33 GeForce RTX 2080 Ti-0 77936867 OK      800   0.00%; 1900 us/it; ETA 1d 17:08; 1579c241dc63eca6 (check 0.80s)
2020-10-21 18:22:51 GeForce RTX 2080 Ti-0 77936867 OK    10000   0.01%; 1880 us/it; ETA 1d 16:42; fc4f135f7cf4ad29 (check 0.80s)
2020-10-21 18:23:11 GeForce RTX 2080 Ti-0 77936867 OK    20000   0.03%; 1881 us/it; ETA 1d 16:42; 3cd1bd9d5e09cbc5 (check 0.80s)
2020-10-21 18:23:30 GeForce RTX 2080 Ti-0 77936867 OK    30000   0.04%; 1880 us/it; ETA 1d 16:41; c4e0ff35e3290d98 (check 0.80s)
2020-10-21 18:23:50 GeForce RTX 2080 Ti-0 77936867 OK    40000   0.05%; 1880 us/it; ETA 1d 16:41; dffe1b1b0d748128 (check 0.80s)
2020-10-21 18:24:10 GeForce RTX 2080 Ti-0 77936867 OK    50000   0.06%; 1880 us/it; ETA 1d 16:41; 52e286945371ed29 (check 0.80s)
2020-10-21 18:24:29 GeForce RTX 2080 Ti-0 77936867 OK    60000   0.08%; 1893 us/it; ETA 1d 16:58; 0945da4dc08bdd95 (check 0.80s)
2020-10-21 18:24:49 GeForce RTX 2080 Ti-0 77936867 OK    70000   0.09%; 1880 us/it; ETA 1d 16:39; 7131fa4eb77f4bb2 (check 0.80s)
2020-10-21 18:25:09 GeForce RTX 2080 Ti-0 77936867 OK    80000   0.10%; 1881 us/it; ETA 1d 16:40; 8d76071d27ee4221 (check 0.80s)
2020-10-21 18:25:28 GeForce RTX 2080 Ti-0 77936867 OK    90000   0.12%; 1880 us/it; ETA 1d 16:39; 0bacff453b2f470e (check 0.80s)
2020-10-21 18:25:48 GeForce RTX 2080 Ti-0 77936867 OK   100000   0.13%; 1880 us/it; ETA 1d 16:39; 6d7296b9e2830f50 (check 0.80s)
2020-10-21 18:26:07 GeForce RTX 2080 Ti-0 77936867 OK   110000   0.14%; 1880 us/it; ETA 1d 16:39; 8cbfd4435622bda7 (check 0.80s)
2020-10-21 18:26:27 GeForce RTX 2080 Ti-0 77936867 OK   120000   0.15%; 1880 us/it; ETA 1d 16:39; 79ae5dad855057ad (check 0.80s)
2020-10-21 18:26:47 GeForce RTX 2080 Ti-0 77936867 OK   130000   0.17%; 1880 us/it; ETA 1d 16:38; 50c97bcbf876231f (check 0.80s)
2020-10-21 18:27:06 GeForce RTX 2080 Ti-0 77936867 OK   140000   0.18%; 1881 us/it; ETA 1d 16:39; e1db15f897271496 (check 0.80s)
2020-10-21 18:27:26 GeForce RTX 2080 Ti-0 77936867 OK   150000   0.19%; 1893 us/it; ETA 1d 16:54; 127631386c6a9b17 (check 0.80s)
2020-10-21 18:27:46 GeForce RTX 2080 Ti-0 77936867 OK   160000   0.21%; 1880 us/it; ETA 1d 16:37; 25b7b6206fc6f085 (check 0.80s)
2020-10-21 18:28:05 GeForce RTX 2080 Ti-0 77936867 OK   170000   0.22%; 1880 us/it; ETA 1d 16:37; 416816b0d9f4bba8 (check 0.80s)
2020-10-21 18:28:25 GeForce RTX 2080 Ti-0 77936867 OK   180000   0.23%; 1880 us/it; ETA 1d 16:36; 6bee5d054f770861 (check 0.80s)
2020-10-21 18:28:44 GeForce RTX 2080 Ti-0 77936867 OK   190000   0.24%; 1880 us/it; ETA 1d 16:36; f37f068f014b18a0 (check 0.80s)
2020-10-21 18:29:04 GeForce RTX 2080 Ti-0 77936867 OK   200000   0.26%; 1881 us/it; ETA 1d 16:37; f0b04b45b0855bd2 (check 0.80s)
2020-10-21 18:29:24 GeForce RTX 2080 Ti-0 77936867 OK   210000   0.27%; 1880 us/it; ETA 1d 16:35; 43eb2fc2424d8aac (check 0.80s)
2020-10-21 18:29:43 GeForce RTX 2080 Ti-0 77936867 OK   220000   0.28%; 1881 us/it; ETA 1d 16:36; a1081c6dc6a7689f (check 0.80s)
2020-10-21 18:30:03 GeForce RTX 2080 Ti-0 77936867 OK   230000   0.30%; 1880 us/it; ETA 1d 16:35; 2387818d3d3d0d01 (check 0.80s)
2020-10-21 18:30:23 GeForce RTX 2080 Ti-0 77936867 OK   240000   0.31%; 1886 us/it; ETA 1d 16:43; a9deae45055e5216 (check 0.84s)
2020-10-21 18:30:42 GeForce RTX 2080 Ti-0 77936867 OK   250000   0.32%; 1887 us/it; ETA 1d 16:44; 89fcab15218f7cac (check 0.80s)
2020-10-21 18:31:02 GeForce RTX 2080 Ti-0 77936867 OK   260000   0.33%; 1881 us/it; ETA 1d 16:35; 55da428da4cf928a (check 0.80s)
2020-10-21 18:31:21 GeForce RTX 2080 Ti-0 77936867 OK   270000   0.35%; 1880 us/it; ETA 1d 16:34; dc349756c5f05abf (check 0.80s)
2020-10-21 18:31:41 GeForce RTX 2080 Ti-0 77936867 OK   280000   0.36%; 1880 us/it; ETA 1d 16:34; 3564af24488443f4 (check 0.80s)
2020-10-21 18:32:01 GeForce RTX 2080 Ti-0 77936867 OK   290000   0.37%; 1880 us/it; ETA 1d 16:33; 63fb281a06f78198 (check 0.80s)
2020-10-21 18:32:19 GeForce RTX 2080 Ti-0 Stopping, please wait..
2020-10-21 18:32:20 GeForce RTX 2080 Ti-0 77936867 OK   300000   0.38%; 1880 us/it; ETA 1d 16:33; 990aa099aad5bf9c (check 0.80s)
2020-10-21 18:32:20 GeForce RTX 2080 Ti-0 Exiting because "stop requested"
2020-10-21 18:32:20 GeForce RTX 2080 Ti-0 Bye

That's it, 1880 us/it, tested at stock settings (or rather stock overclocking settings), but with AIO slapped on it, so it doesn't suffocate from the heat under the chunky stock cooler.

BTW:
ASUS ROG Strix 2080Ti O11G
Stock cooler on stock settings -> about 70 to 80 °C under max load
AIO on stock settings -> never above 45 °C under max load

Watercooling is the way to go!

kriesel · 2020-10-21, 16:46

Quote:

Originally Posted by Viliam Furik

(BTW FP32 is what we want for TF)

I don't think so. Consider that the source code for mfaktc contains lots of uint32, occasional uint or uint64, and no float declarations in gpusieve.c, my_types.h; a mix of int and float in the kernel code. We want good performance in int 24 and 32, and float.

moebius · 2020-10-21, 16:49

Thank you, the values will be more like an RTX 3080, of course without further SP/DP improvements from Preda.

2020-10-21, 16:49	#2519
moebius Jul 2009 Germany 1000100011₂ Posts	Thank you, the values will be more like an RTX 3080, of course without further SP/DP improvements from Preda. Last fiddled with by moebius on 2020-10-21 at 16:56

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfakto: an OpenCL program for Mersenne prefactoring	Bdot	GPU Computing	1668	2020-12-22 15:38
GPUOWL AMD Windows OpenCL issues	xx005fs	GpuOwl	0	2019-07-26 21:37
Testing an expression for primality	1260	Software	17	2015-08-28 01:35
Testing Mersenne cofactors for primality?	CRGreathouse	Computer Science & Computational Number Theory	18	2013-06-08 19:12
Primality-testing program with multiple types of moduli (PFGW-related)	Unregistered	Information & Answers	4	2006-10-04 22:38

2020-10-21, 13:18	#2510
aheeffer Aug 2020 37 Posts	Thanks a lot indeed! Now we finally know. With 619µsec for this exponent, the Radeon VII remains the king. Even underclocked and running at 150 W, the Radeon VII is still twice as fast as the RTX 3090. I will cancel my order for the RTX 3080 now.and hope for the best with the new RX 6900 XT.

2020-10-21, 13:52	#2511
ATH Einyen Dec 2003 Denmark 3,037 Posts	It is no surprise really: https://en.wikipedia.org/wiki/GeForce_30_series Gpuowl and CUDALucas uses "double precision", and 3090 has only 1/32 the double precision performance compared to single precision performance (918 GFLOPS vs 29389 GFLOPS). It has been like that for many years now with Nvidia consumer cards. With 29389 GFLOPS single precision it should do really well at trial factoring with mfaktc, probably one of the fastest cards for trial factoring including the expensive Tesla cards.