Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
Message boards :
Number crunching :
GTX 1660 Ti performance
| Author |
Message |
|---|
 Azmodes  Volunteer tester
Send message
Joined: 30 Dec 16
Posts: 184
ID: 479275
Credit: 2,495,167,703
RAC: 395,664
                          
|
|
So because I was curious (and also impulsive and pretty addicted to BOINC at this point), I recently purchased a 1660 Ti. I was mainly interested in how it performs at Genefer manual sieving, considering how the RTX cards are crushing it at that particular project. The 1660 Ti is a Turing chip as well, but without the tensor cores of the RTX series, so I wasn't sure what that would imply for sieving and PG performance in general.
Well, without further ado, here are some results. I will add more from other subprojects as they come in.
Some specs from the rig:
CPU: i9-9900k (running 14 GCW sieve tasks)
RAM: 2x4 GB DDR4 at 3100 MHz
OS: Ubuntu 18.04
GPUs: 1x PNY GeForce GTX 1660 Ti XLR8 Gaming Overclocked Edition+ 1x GTX 980 (running GPUGrid, with one thread assigned via SWAN_SYNC); driver 410.78*
No CPU or GPU overclocks (if you don't count the i9 turbo-ing).
GFN21 manual sieving: (all with w1)
b7 160 P/day
b8 230 P/day
b9 330 P/day (this is already faster than a 1080 Ti on b13 by 60 P/day)
b10 360 P/day
b11 400 P/day
b12 420 P/day
b13 430 P/day (cf. 2070 ~600, 2080 with slight core OC ~840)
I can't attest to screen lag, since the monitor is connected to the 980. I haven't tried multiple instances at once, core load is pretty maxed out already and I've seen actual decreases in throughput on another Linux machine that way. But yeah, safe to say Turing rocks at msieving and I will definitely throw this baby at finishing the GFN21 range.
Genefer primality tests: (sample size only one task)
15: 28 secs
16: 87 secs
17Low: 304 secs
17Mega: 330 secs
19: 45-50 mins (non-client projection)
20: ~2.5-2.75 hours (non-client projection)
AP27: 1,194 secs
PPS sieve: 248 secs (one task) / 464 secs (2 tasks in tandem)
*While the 980 is recognized as such, the 1660 Ti is only referred to as "Graphics Device" by the OS. Everything seems to work fine, though (OpenCL and CUDA stuff both evidently), so not sure if that's just because it's brand new or I should update the driver. Or perhaps Maxwell and Turing make strange bedfellows?
____________
Long live the sievers.
+ Encyclopaedia Metallum: The Metal Archives + | |
|
|
|
|
Good to know.
I wasn't aware of a GTX 1660 Ti.
For Genefer and AP27 it looks twice as fast as my GTX 1060 3GB.
____________
"Accidit in puncto, quod non contingit in anno."
Something that does not occur in a year may, perchance, happen in a moment. | |
|
Dad
Send message
Joined: 28 Feb 18
Posts: 284
ID: 984171
Credit: 182,080,291
RAC: 0
             
|
|
WOW
That's quick
____________
Tonight's lucky numbers are
555*2^3563328+1 (PPS-MEGA)
and
58523466^131072+1 (GFN-17 MEGA) | |
|
Michael Goetz Volunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14695
ID: 53948
Credit: 1,042,938,327
RAC: 2,140
                                       
|
*While the 980 is recognized as such, the 1660 Ti is only referred to as "Graphics Device" by the OS. Everything seems to work fine, though (OpenCL and CUDA stuff both evidently), so not sure if that's just because it's brand new or I should update the driver. Or perhaps Maxwell and Turing make strange bedfellows?
You should update your drivers. Not only is that likely the cause of the name not showing up, but since the driver is actually the compiler that in real time compiles the GPU program into the machine code for the GPU, an up to date driver may produce faster code. A driver that doesn't know about the GPU model may not be using all of the available GPU features.
____________
My lucky number is 75898524288+1 | |
|
|
|
|
NVIDIA GeForce GTX 1080 Ti (4095MB) driver: 41735
AP27: 760sec (1900MHz)
CPU: i7-8700K SR5 -t 6 | |
|
Azmodes Volunteer tester
Send message
Joined: 30 Dec 16
Posts: 184
ID: 479275
Credit: 2,495,167,703
RAC: 395,664
|
|
GFN18: 1,213 secs
GFN19: 2,830 secs
GFN20: 10,218 secs
GFN21: 43,661 secs
GFN22: <44 hours (non-client projection)
*While the 980 is recognized as such, the 1660 Ti is only referred to as "Graphics Device" by the OS. Everything seems to work fine, though (OpenCL and CUDA stuff both evidently), so not sure if that's just because it's brand new or I should update the driver. Or perhaps Maxwell and Turing make strange bedfellows?
You should update your drivers. Not only is that likely the cause of the name not showing up, but since the driver is actually the compiler that in real time compiles the GPU program into the machine code for the GPU, an up to date driver may produce faster code. A driver that doesn't know about the GPU model may not be using all of the available GPU features.
I'm having some trouble with updating the driver, working on it... 415 is the highest I've managed, so to speak. The identification issue remains.
____________
Long live the sievers.
+ Encyclopaedia Metallum: The Metal Archives + | |
|
|
|
GFN18: 1,213 secs
GFN19: 2,830 secs
GFN20: 10,218 secs
GFN21: 43,661 secs
GFN22: <44 hours (non-client projection)
As fast as the GTX 1080 !?
| |
|
Scott Brown Volunteer moderator
Project administrator
Volunteer tester
Project scientist
Send message
Joined: 17 Oct 05
Posts: 2542
ID: 1178
Credit: 30,932,444,644
RAC: 6,381,061
                                                              
|
GFN18: 1,213 secs
GFN19: 2,830 secs
GFN20: 10,218 secs
GFN21: 43,661 secs
GFN22: <44 hours (non-client projection)
Those times (excluding the GFN22 result where I don't have a comparison) are consistently faster than a GTX 1080 by a good margin, except for GFN21 where the times are about the same between the 1660 and 1080.
A closer look at the GFN20 vs. GFN21 results is interesting. The GTX 1660Ti used the OCL4 transform for GFN20, but used the OCL5 transform for GFN21. The RTX cards tend to use OCL5 instead of OCL4 on these higher GFN tasks, but I haven't seen this behavior on my 2070. Perhaps it is a driver issue and we can anticipate better performance on GFN21 soon with a new release, but maybe not? | |
|
Azmodes Volunteer tester
Send message
Joined: 30 Dec 16
Posts: 184
ID: 479275
Credit: 2,495,167,703
RAC: 395,664
|
|
I've fixed the driver issue. Now running 418.43 and the card is properly recognized. I haven't tested any Genefers with the new driver yet, but msieving performance is the same.
____________
Long live the sievers.
+ Encyclopaedia Metallum: The Metal Archives + | |
|
mikey
Send message
Joined: 17 Mar 09
Posts: 2495
ID: 37043
Credit: 1,150,189,458
RAC: 234,623
              
|
|
If anyone wants one here they are on sale in the US..
http://www.nowinstock.net/computers/videocards/nvidia/gtx1660ti/ | |
|
compositeVolunteer tester
Send message
Joined: 16 Feb 10
Posts: 1282
ID: 55391
Credit: 2,332,117,278
RAC: 186,145
                 
|
GFN18: 1,213 secs
GFN19: 2,830 secs
GFN20: 10,218 secs
GFN21: 43,661 secs
GFN22: <44 hours (non-client projection)
As fast as the GTX 1080 !?
NVidia says... "performance that rivals the GeForce GTX 1070" (nvidia.com) | |
|
mackerel Volunteer tester
Send message
Joined: 2 Oct 08
Posts: 2976
ID: 29980
Credit: 787,861,900
RAC: 6,873
                 
|
|
nvidia general statements are probably focused around gaming performance. Whatever they did in Turing, it certainly helps out more in compute. One of the changes for example is the int32 and fp32 pipes can be used at the same time, whereas before you use one or the other. If there is a mix of instructions that can be run in parallel like this, this might explain part of that speedup. | |
|
Azmodes Volunteer tester
Send message
Joined: 30 Dec 16
Posts: 184
ID: 479275
Credit: 2,495,167,703
RAC: 395,664
|
|
Times from after the driver update, averages from 4-5 tasks:
15: 27 secs
16: 82 secs
17Low: 300 secs
17Mega: 317 secs (most around 325, but one outlier came in at 282 for some reason)
By the way, temps are mostly <70C.
____________
Long live the sievers.
+ Encyclopaedia Metallum: The Metal Archives + | |
|
Azmodes Volunteer tester
Send message
Joined: 30 Dec 16
Posts: 184
ID: 479275
Credit: 2,495,167,703
RAC: 395,664
|
|
I finally figured out how to manipulate GPU clocks in Linux. Without touching voltages, I pushed the 1660 Ti to 450 P/day with +150 on the core and -2000 on memory (memory clocks don't matter for sieving and reducing them can help with heat/power consumption).
That's 1,935,000 credit per day.
____________
Long live the sievers.
+ Encyclopaedia Metallum: The Metal Archives + | |
|
robish Volunteer moderator
Volunteer tester
Send message
Joined: 7 Jan 12
Posts: 2264
ID: 126266
Credit: 10,804,259,671
RAC: 1,988,834
              
|
I finally figured out how to manipulate GPU clocks in Linux. Without touching voltages, I pushed the 1660 Ti to 450 P/day with +150 on the core and -2000 on memory (memory clocks don't matter for sieving and reducing them can help with heat/power consumption).
That's 1,935,000 credit per day.
That is just....cool. Result!!!!! :) May need to purchase soon. ;)
____________
My lucky number 10590941048576+1 | |
|
|
|
|
Ubuntu's repository must have more current drivers than Manjaro's (I did a system update about 5 minutes ago...) Then again, every time I tried to get the proprietary drivers to work with Ubuntu, I got that damned log-in loop, and Nouveau is about as useful as a light bulb with a burnt out filament.
____________
 | |
|
|
|
I finally figured out how to manipulate GPU clocks in Linux. Without touching voltages, I pushed the 1660 Ti to 450 P/day with +150 on the core and -2000 on memory (memory clocks don't matter for sieving and reducing them can help with heat/power consumption).
That's 1,935,000 credit per day.
Der8auer has a video up about overclocking cards, specifically the 1080 Ti and he says that +150 on the clock is something that every card should be able to do. I haven't done anything with the clocks because GPUs arent exactly cheap, I'm running a Founder's Edition, which doesn't have the best cooling in the world and I'm way too lazy to do a custom loop right now, and I'd prefer, if I buy another, that it be for added crunching, not as a replacement. What are you boosting to?
____________
| |
|
|
|
|
Looking at this thread, i have desided to get me a 1660ti and see what i can do with it. | |
|
mikey
Send message
Joined: 17 Mar 09
Posts: 2495
ID: 37043
Credit: 1,150,189,458
RAC: 234,623
|
Looking at this thread, i have desided to get me a 1660ti and see what i can do with it.
Good things of course!! | |
|
|
|
Looking at this thread, i have desided to get me a 1660ti and see what i can do with it.
Good things of course!!
Got my Zotac AMP 1660 ti today, and uninstalled mu old driver for my good old asus gtx 1060 3gb and replaced the card and installed the latest driver from nvidias website and all was fine, games and all. But opencl for the card is not there! Ok, my amd cards in the same pc were working and opencl was present, hmm? OK then...
Went through some forums and found that nvidia disables opencl for different reasons for different people. So then i unistalled the nvidia driver and what do you know my amd cards lost they'r opencl capability, niice nvidia... ok then. Downloaded the oldest nvidia driver available for the card and after installing it all gpus are now blessed with opencl capability. Reminds me of the old days with nvidia physixs, when amd card was present, nvidia physixs is disabled on ANY nvidia card.
Same old, same old. Ok, i'm going to bed now, tired and angry. | |
|
|
|
|
So, a month has passed, and i have had no problems with my 1660ti or driver stability issues. Recently nvidia released a gaming driver 436.15 and i decided to try that and it enabled all of the compute capabilitys, so its even better now, thanks nvidia.
Using it on my good old PII hexa on asus sabertooth with 16gb ddr3 and 256gb SSD runnin on win7 pro 64. Runtimes compared to my old gtx 1060 3gb:
Primegrid AP 27:
gtx 1060 3gb 1 wu at a time: runtime per wu ~2180 sec
gtx 1660ti 6gb 1 wu at a time: runtime per wu ~1190 sec
gtx 1660ti 6gb 2 wu-s at a time: runtime per wu ~2400 sec
Primegrid pps sieve:
gtx 1060 3gb 2 wu-s at a time: ~540000 points per day
gtx 1660ti 6gb 2 wu-s at a time: ~1,40000 points per day (410sec per wu)
gtx 1660ti 6gb 3 wu-s at a time: no difference in points per day. (cpu load is higher)
gtx 1660ti 6gb 1 wu at a time: ~ 350sec per wu or 800 000+ points per day
collatz:
gtx 1060 3gb 2wu-s at a time: 1700+ sec runtime per wu. (around 3 200 000 points per day).
gtx 1660ti 2 wu-s at a time: 700+ sec runtime per wu. (around 7 450 000 points per day).
Gtx 1060 and 1660ti are both installed on a chinese pci extender cable of ~35cm in lenght.
Asus OC 1060 3gb boost holds at 1910mhz:
https://www.upload.ee/image/10497788/gtx_1060_3gb.png
Zotac AMP OC 1660ti 6gb boost holds at 1965 mhz:
https://www.upload.ee/image/10497789/gtx1660ti_.png
Asus OC 1060 3gb rated tdp is 120w, Zotac AMP OC 1660ti 6gb rated tdp is 130w (120w tdp is for non oc models, oc models are rated 130w tdp).
Measured the power draw difference for my pc with a kill a watt. My psu is corsair AX860 platinum, cause i got 3 gpus.
The cards were not installed in the pc simultaneously, and difference was around 15watts from the socket. The 1660ti pulls around 15watts more from the socket, as it should.
I used collatz for testing, cause it loads the gpu heavily.
While the 1660ti is not exactly cheap, since it is very power efficient and very fast, its a must have, and i am not dissappointed at all with its performance. Infact, it performs better than i expected.
pps sieve config that i am using for the gtx 1060 and 1660ti:
<app_config>
<app>
<name>pps_sr2sieve</name>
<max_concurrent>16</max_concurrent>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.875</cpu_usage>
</gpu_versions>
</app>
</app_config>
collatz config that i am using for the gtx 1060 and 1660ti:
<app_config>
<app>
<name>collatz_sieve</name>
<max_concurrent>16</max_concurrent>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.875</cpu_usage>
</gpu_versions>
</app>
</app_config>
and:
<app_config>
verbose=1
kernels_per_reduction=48
threads=8
lut_size=17
sleep=1
reduce_cpu=0
sieve_size=28
</app_config> | |
|
Message boards :
Number crunching :
GTX 1660 Ti performance |
