What SIEVE_PRIMES are you using?

[Chuck]

Originally Posted by Prime95

I'm working on GPU sieving and I'd like to know what sieve_primes setting mfaktc most users are using.

5000. Two instances of mfaktc on GTX-580 with core-i7 970. No affinities.

[Jaxon]

~20,000 running 3 instances of mfaktc on a GTX 580. My CPU is a i5-3570k running at 4100MHz. GPU use is near 100%.

[ATH]

Geforce GTX 460 running on a Q9450 (Yorksfield):

1 instance of mfaktc 80M exponent 71-72 bits:
SievePrimes auto adjusts to 5000. avg rate 125M/s. GPU load 73-75%. CPU load 25% (1 full core)

2 instances of mfaktc 80M exponent 71-72 bits:
SievePrimes auto adjusts to ~ 17000-19000. avg rate 2 x 85M/s. GPU load 95-99%. CPU load 50% (2 cores)

[kjaget]

As with everyone else, depends on the number of instances of mfaktc.

I have a GTX560ti448 (basically a 570) with an OCd i5-750

1 or 2 instances don't max out the GPU even at sieve primes = 5000. Going from 1 to 2 instances doubles my effective throughput.

With 3 instances I can max the GPU so I only get a ~27% speedup, sieve primes are in the 10-12,000 range.

Going to 4 instances gives me a ~10% speedup, sieve primes go up to ~3x,xxx (from memory, might not be 100% accurate).

For your entertainment, here's how number of candidates is influenced by sieve primes on a really slow GPU. Candidates is in millions, time is run time per class and rate is M candidates / sec. This is from a 49M exponent from 70 to 71. I only ran 1 class per sieve primes value so there's some noise in the timings, but the candidates should be correct. Looks like ~5-6% reduction in candidates every time you double sieve primes, at least in the 5-200K range that mfaktc uses.

Sieve Primes	candidates	time	rate		
5,000	643.83	79.133	8.14
5,625	636.49	78.161	8.14
6,328	628.62	77.283	8.13
7,119	621.81	76.271	8.15
7,500	618.14	75.288	8.21
8,008	614.47	75.369	8.15
8,437	611.32	74.453	8.21
9,009	607.65	74.525	8.15
9,491	604.5	73.627	8.21
10,135	600.83	73.66	8.16
10,677	597.69	72.794	8.21
11,401	594.02	72.872	8.15
12,011	590.87	71.963	8.21
12,826	587.2	72.067	8.15
13,512	584.58	71.202	8.21
14,429	580.91	71.366	8.14
15,201	578.29	70.436	8.21
16,232	574.62	70.002	8.21
17,101	572	69.666	8.21
18,261	568.85	69.897	8.14
19,238	566.23	70.685	8.01
20,000	564.13	68.713	8.21
20,543	562.56	68.4	8.22
21,642	559.94	68.201	8.21
22,500	558.37	68.01	8.21
23,110	556.79	67.644	8.23
24,347	554.17	67.5	8.21
25,998	551.03	67.644	8.15
27,339	548.41	66.802	8.21
29,247	545.78	66.935	8.15
30,813	543.16	66.158	8.21
32,902	540.02	66.301	8.14
34,664	537.4	65.458	8.21
37,014	534.77	65.492	8.17
38,997	532.15	64.818	8.21
41,640	529.53	64.805	8.17
43,871	526.91	64.181	8.21
46,845	524.29	66.124	7.93
49,354	522.19	63.611	8.21
50,000	521.14	63.479	8.21
52,700	519.05	63.399	8.19
55,523	516.95	64.691	7.99
56,250	516.42	62.906	8.21
59,287	514.33	63.179	8.14
62,463	512.23	62.4	8.21
63,281	511.18	62.267	8.21
66,697	509.08	62.493	8.15
70,270	506.99	61.759	8.21
71,191	506.46	61.692	8.21
75,000	504.37	61.443	8.21
75,034	504.37	61.935	8.14
79,053	502.27	61.185	8.21
80,089	501.74	61.122	8.21
84,413	499.65	61.352	8.14
88,934	498.07	60.674	8.21
90,100	497.55	60.612	8.21
94,964	495.45	60.851	8.14
100,000	493.36	60.711	8.13
100,050	493.36	60.102	8.21
101,362	492.83	60.043	8.21
106,834	490.73	60.273	8.14
112,500	489.16	59.756	8.19
112,556	488.64	59.528	8.21
114,032	488.64	59.53	8.21
120,188	486.54	59.694	8.15
126,562	484.44	59.048	8.20
126,625	484.44	59.021	8.21
128,286	483.92	58.958	8.21
135,211	482.34	59.352	8.13
142,382	480.25	58.512	8.21
142,453	480.25	58.509	8.21
144,321	479.72	60.202	7.97
152,112	478.15	60.392	7.92
160,179	476.05	57.998	8.21
160,259	476.05	57.998	8.21
162,361	475.53	57.939	8.21
171,126	473.96	58.494	8.10
180,201	472.38	57.555	8.21
180,291	472.38	59.309	7.96
182,656	471.86	57.493	8.21
192,516	469.76	57.5	8.17
200,000	468.71	57.606	8.14

[Bdot]

One HD5770 with 3 instances at SP 55,000 fix each consuming a little over 2 cores of an Phenom X4 955 (3.2GHz).

Another HD5770 with 3 instances at auto-adjust, SP ~100,000 with 3 full cores of a Xeon hex-core. As my GPUs deliver only about 150-160M/s I can trade the 3rd core's 2.6 CPU-GHz-days against 10-12 GPU-TF-GHz-days. With just 2 instances SP would settle at ~30,000.

An NV Quadro2000, 2x ~95,000 auto-adjust, 2 cores of a Xeon-hex.

BTW, I'm on rcv's code to port it to OpenCL.

[bcp19]

Originally Posted by Prime95

I'm working on GPU sieving and I'd like to know what sieve_primes setting mfaktc most users are using.

I think mfaktc can auto-adjust this value. If so, please report the sieve_primes value mfaktc thinks is optimal.

My goal is to create a GPU sieve that is as fast as the CPU sieve for most users. To do this, the GPU sieve must be fast enough so that it can sieve deeper than the CPU -- where the time saved by reduced TF candidates equals the time spent doing GPU sieving.

This may not be possible, but it is good to have goals!

I have 6 machines TFing using 7 cards and each is set up differently. 2500K/480, 2 instances, SP=5000, will up it to 3 after my 100M LL finishes and probably have SP=20,000-30,000. 2400/480, 3 instances, SP=11000(when it had the 560Ti SP = 33000). Q8200/560 Ti, 3 instances, SP=14000. Q6600/560, 3 instances, SP=9500. Q8200/550 Ti, 3 instances, SP=40000. Phenom II 1055T/460, 3 instances, SP=35000, 1055T/6770, 2 instances, SP=5000.

[Prime95]

Thanks for all the data -- I was not expecting quite that much variety!

Just an update as to where I stand. I've built a test program that emulates the work required by a mfaktc GPU sieve. 95% of this code would be usable in a real mfaktc implementation.

On my 460, the GPU sieve reaches the point of diminishing returns at roughly 71265 Primes (First Prime Is 13, Last Prime Is 899851). The speed of the GPU sieve is such that it achieves as much throughput as a CPU sieve sieving 14177 primes from (first prime is 13, last prime is 153877).

I've got a few tweaks in mind that might improve things a tad.

Of course, this is all theory. A real implementation might run into roadblocks that I had not considered.

Originally Posted by Bdot

BTW, I'm on rcv's code to port it to OpenCL.

PM me if you'd like a copy of my test code. It is a blending of erato.cu from the "second dumbest cuda" thread and rcv's ideas as well as several ideas of my own.

[kracker]

I don't run mfaktc since I don't have a Nvidia card... but I have mfakto
Damn, wish the attachment limit file size was just a bit bigger...

[Xyzzy]

Damn, wish the attachment limit file size was just a bit bigger...

Just use a lower compression setting. The attached file at 123kb is very close to your original.

[kracker]

Originally Posted by Xyzzy

Just use a lower compression setting. The attached file at 123kb is very close to your original.

Ah I see. Thanks!

[TObject]

Also try GIF; it beats JPEG when there are a lot of continuous tones in the picture.