The GTX750(Ti) is one of the most energy efficient cards and therefore one of the most interresting cards for crunching.
I thought I start a thread for just this type of card so ppl who own this can share their experience and discuss settings, command lines, etc.
I'm interrested to hear what cards exactly you own and how you use them. Do you crunch MB or AP? How many tasks do you run at the time? What command line do you use? Do you overclock and if yes, how and how much? And overall, are you satisfied with the performance of your card?

Let me start here: I own a MSI GTX750 OC Edition with 1GB DDR5 RAM.Clock speed is 1059/1137 Mhz. At least that's what GPU-Z is showing. When crunching I see a clock speed of 1202 Mhz.
I usually crunch AP on my card, 2 tasks at a time and I currently use the command line which Mike suggests for the 750Ti:

-use_sleep -unroll 10 -oclFFT_plan 256 16 512 -ffa_block 12288 -ffa_block_fetch 6144

Crunching 2 AP7 tasks usually takes about 5000 seconds, sometimes a bit more, sometimes a bit less.

Today I installed MSI Afterburner and made my first steps in overclocking my card. I raised clockspeed by 100 and memory clock by 200. Seems stable and temperature didn't go up (max 62 degrees with fan running at 38%) so I could maybe try a little more next time. I only crunched 4 tasks with this settings yet but it seemed to speed up crunching by about 10 minutes.

Overall I'm satisfied with this card. It's not the fastest out there but it gives you a lot of computing power per watt and that is very important for me.


Ok, your turn folks :-)

---

Asus GTX750TI-OC-2GD5, core 1188Mhz, memory 2048MB DDR5 @ 1350Mhz, factory defaults.

Running stock MB, one task at time. RAC about 11k.

Just switched to stock AP v7, one task@time, waiting MB cache to clear out.

Computer 6864212

---

"Please keep Your signature under four lines so Internet traffic doesn't go up too much"

- In 1992 when I had my first e-mail address -

I'm running 2 x EVGA GTX750Ti FTW w/ 2GB DDR5 @ 1189 MHz Base & w/ 1268 MHz Boost Clock (factory defaults). AP is 2 x ea. card as the primary tasks with 3 x MB to take up the slack on each. With the speed of my 4700K, I'm able to allocate .5 CPU per task with the AP tasks usually averaging 4.7k~4.8k seconds per task. I've run AP tasks since they started and switched to v7 when it became available. I use EGA Precision X 4.2.1 to control fan speed. With a lot of assistance from fellow beta testers, I came up with the following command line which seems to work for me (bold). The -tune 2 & -ocIFFT_plan, came from the readme.txt to try out and I left them there. I'm able to use such high amounts because the GPUs are standalone and I use the HD4600 iGPU for the monitor.

-use_sleep -unroll 12 -ffa_block 12288 -ffa_block_fetch 6144 -tune 1 64 8 1 -tune 2 128 8 1 -ocIFFT_plan 256 16 256 -hp

My host http://setiathome.berkeley.edu/results.php?hostid=5501972&offset=0&show_names=0&state=4&appid=20.

With AP tasks also working on the CPU, I use 80% CPU when I'm on the machine and 90% when I'm off it.

---

I don't buy computers, I build them!!

Cliff,


Mike posted this in the AP7 thread about the unroll

Make sure unroll needs to be double of your number of CU`s.

I have 2 Evga GTX 750Ti that share space with 2 Evga GTX 780 SC. I know, I shouldn't mix and match. I run 3 APs per card or 3 MB per card. Since I run with AMD chips I use slightly higher percentage of CPU 0.6 CPu for AP and 0.4 for MB. For the MB that really is overkill since they don't use that much but I'm lazy and since I don't run CPU work units it's fine.

-use_sleep -unroll 16 -oclfft_plan 256 16 256 -ffa_block 8192 -ffa_block_fetch 4096 -tune 1 64 4 1 -tune 2 64 4 1 -hp

With these setting I'm using 97% of the GPUs and 32% of the CPU. I could try to run some CPU task but I've found that it sometimes causes lock ups with the AMD CPU so I only run the GPUs.



Zalster

I have an EVGA GTX 750 (01G-P4-2761-KR). I am currently crunching Astropulse only and running two tasks at the same time.

I used -unroll 10 -ffa_block 4096 -ffa_block_fetch 2048 -hp -use_sleep

but after the previous post I changed it to

-unroll 8 -ffa_block 4096 -ffa_block_fetch 2048 -hp -use_sleep

I tried higher values for -ffa_block and -ffa_block_fetch earlier (over on Beta), but this seemed to slow things down.

GPU-Z (page 1) reports default clock 1020 MHz, Memory 1253 MHz and Boost Clock 1085 MHz.
GPU-Z (page 2) reports Core Clock 1162.8 MHz, Memory Clock 1252.8 MHz
I don't understand why the Core Clock is reporting a higher number than the Boost Clock.

I am very happy with this card but have not tried to overclock it yet. I have EVGA Precision X installed, so it should be easy enough to give it a try (and I will when I have a couple of days to monitor the system).

Tom

---

Cliff,


Mike posted this in the AP7 thread about the unroll

Make sure unroll needs to be double of your number of CU`s.

I used -unroll 10 -ffa_block 4096 -ffa_block_fetch 2048 -hp -use_sleep

but after the previous post I changed it to

-unroll 8 -ffa_block 4096 -ffa_block_fetch 2048 -hp -use_sleep

Yes, that's an interresting point. If it's true that unroll should be twice the number of the cards CUs then we must use 8 for our GTX750s. But IIRC Mike told me that 10 is ok.
Also, Mike said that using the "tune" switch may result in a slowdown on my card. Still, I see many ppl use it. Or is there maybe a difference between the regular 750 and the Ti-version?

Anyway, I will try unroll 8 now and see if I can detect any differences.

---

OK,

I'm showing unfamiliarness with the command lines. I thought CU was number of cpu cores. What does CU actually mean?

CU = Compute Units

The 750 has 4, the 750Ti has 5

---

CU = Compute Units

The 750 has 4, the 750Ti has 5

That's a new one for me.. Where do you find that in the specifications? I have several 780s and would like to know how to configure that unroll

Cliff,


Mike posted this in the AP7 thread about the unroll

Make sure unroll needs to be double of your number of CU`s.

I used -unroll 10 -ffa_block 4096 -ffa_block_fetch 2048 -hp -use_sleep

but after the previous post I changed it to

-unroll 8 -ffa_block 4096 -ffa_block_fetch 2048 -hp -use_sleep

Yes, that's an interresting point. If it's true that unroll should be twice the number of the cards CUs then we must use 8 for our GTX750s. But IIRC Mike told me that 10 is ok.
Also, Mike said that using the "tune" switch may result in a slowdown on my card. Still, I see many ppl use it. Or is there maybe a difference between the regular 750 and the Ti-version?

Anyway, I will try unroll 8 now and see if I can detect any differences.

Two times the CU`s is standard value.
Some cards can run unroll 3 times the CU`s while use_sleep is in place.

I have to make sure it works for everyone.
When i suggest uroll 10 you can safely use it.
You didn`t get any overflows did you ?

---

With each crime and every kindness we birth our future.

CU = Compute Units

The 750 has 4, the 750Ti has 5

That's a new one for me.. Where do you find that in the specifications? I have several 780s and would like to know how to configure that unroll

http://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units#GeForce_700_Series
SMX count column.
Or looks in a completed result for

OpenCL Platform Name: NVIDIA CUDA
Number of devices: 3
Max compute units: 12

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[

CU = Compute Units

The 750 has 4, the 750Ti has 5

That's a new one for me.. Where do you find that in the specifications? I have several 780s and would like to know how to configure that unroll

Just use the values i wrote in the read me.

Who do you think found this out ?

Anything above uroll 18 has no effect using oclfft_plan speeds up by 15%.

---

With each crime and every kindness we birth our future.

OpenCL Platform Name: NVIDIA CUDA
Number of devices: 3
Max compute units: 12

Ah, thanks Hal.. Always good to learn something new.

Just use the values i wrote in the read me.

Who do you think found this out ?

Anything above uroll 18 has no effect using oclfft_plan speeds up by 15%.

I figured it was you, but sometimes people like to know the how not just the why. I appreciate all the help you have given us over time Mike. I didn't mean to offend you.

OpenCL Platform Name: NVIDIA CUDA
Number of devices: 3
Max compute units: 12

Ah, thanks Hal.. Always good to learn something new.

Just use the values i wrote in the read me.

Who do you think found this out ?

Anything above uroll 18 has no effect using oclfft_plan speeds up by 15%.

I figured it was you, but sometimes people like to know the how not just the why. I appreciate all the help you have given us over time Mike. I didn't mean to offend you.

No offense taken Zalster.

I spend 3 weeks just testing on nvidia cards to sweet you guys good parmeters.
You know i`m usually an ATI person.

---

With each crime and every kindness we birth our future.

Love the thread! I need some good TI info:)

I run Quad 750TI's and I'm getting ready to start testing with -unroll and drivers and other stuff to squeeze more credits.. So far I haven't had time to mess with AP 7 much so I've just run a generic two per card with 1 core reserved just like with AP 6..

4x EVGA 750 TI Superclocked.. Default clocks are 1176 core, 1350 memory, and it shows 1255 boost but with Seti they run higher..

GPU Boost speeds
device 0: 1320 mhz
device 1: 1333 mhz
device 2: 1294 mhz
device 3: 1306 mhz

I run 100% CPU with the GPU's at 96-99%. 2 MB and/or 2 AP at once with a very conservative -unroll 4 -ffa_block 2048 -ffa_block_fetch 1024.. AP 7's are taking from 2000-4600 each so I guess an average of 3300 looks about right.

I've yet to see the RAC with a healthy week of AP's, but it broke 54,000 at the end of the last AP run then took a nose dive.. It's climbing back up now, I was expecting her to break 56,000. I've since upgrade from a 965 BE to FX-8370 clocked at 4.5 with some nice fast 2133 memory, so she should run a tiny bit faster now.. I'm looking for 58,000 with my current settings, maybe more after I spend tomorrow testing -unrolls and locks and drivers and feed the hamsters some Red Bull:)

Cliff,


Mike posted this in the AP7 thread about the unroll

Make sure unroll needs to be double of your number of CU`s.

I used -unroll 10 -ffa_block 4096 -ffa_block_fetch 2048 -hp -use_sleep

but after the previous post I changed it to

-unroll 8 -ffa_block 4096 -ffa_block_fetch 2048 -hp -use_sleep

Yes, that's an interresting point. If it's true that unroll should be twice the number of the cards CUs then we must use 8 for our GTX750s. But IIRC Mike told me that 10 is ok.
Also, Mike said that using the "tune" switch may result in a slowdown on my card. Still, I see many ppl use it. Or is there maybe a difference between the regular 750 and the Ti-version?

Anyway, I will try unroll 8 now and see if I can detect any differences.

Two times the CU`s is standard value.
Some cards can run unroll 3 times the CU`s while use_sleep is in place.

I have to make sure it works for everyone.
When i suggest uroll 10 you can safely use it.
You didn`t get any overflows did you ?

No, I didn't get any overflows. I also tried unroll 12 yesterday and it also works. But to be honest, I can't see any difference in speed with 8, 10 or 12.

Yesterday I also tried to overclock the GPU a bit more. I applied +150 clock and +300 on memory but after some minutes the card seemed to stop working. Nothing crashed but there was no progress on tasks and the temperature went back to 41 degrees. So maybe +100 clock/+200 memory is max for my card because it is already slightly overclocked by default.

---

Anything above uroll 18 has no effect using oclfft_plan speeds up by 15%.

using oclfft_plan??? Can anyone explain more please. I´m far away from my hosts so i can´t do any serious test.

An example of a command line for the 780 or 670 could help.

Thanks.

---

-use_sleep -unroll 16 -oclfft_plan 256 16 256 -ffa_block 16384 -ffa_block_fetch 8192 -tune 1 64 4 1 -tune 2 64 4 1 -hp



Juan,

This is what I am using.... Aren't you supposed to be on a beach somewhere enjoying a cool cold one??? I'm going to report you for disobeying orders ;)

Edit.. I had to change a few things there. Juan note the second tune for the second kernal



Zalster

Aren't you supposed to be on a beach somewhere enjoying a cool cold one??? I'm going to report you for disobeying orders ;)

Ups, My mistake... i will going for another relaxing cold beer as doctor´s orders.

Thanks for remind me. :)

---

-use_sleep -unroll 16 -oclfft_plan 256 16 256 -ffa_block 16384 -ffa_block_fetch 8192 -tune 1 64 4 1 -tune 2 64 4 1 -hp

Juan,
This is what I am using.... Aren't you supposed to be on a beach somewhere enjoying a cool cold one??? I'm going to report you for disobeying orders ;)

Edit.. I had to change a few things there. Juan note the second tune for the second kernal

Zalster

I suppose the -oclfft_plan is for the newer drivers/cards. It was a total disaster on my old NV 8800 GT with the old CPU friendly driver, Computer 6813106
It appears all four of those will be Invalid :-(
Worked fine until then.

On the other hand, it seems to work Great on my ATI cards. Just look at the difference it made on this host, Computer 7258715