Emulating lower-end graphics cards

[InertiaSelling]

Hi. I'll be spending next year in Vancouver, and unless I shrink my toolcase computer into an ITX build, I'll have to buy an all-purpose laptop. Since I'm not a heavy gamer, I'm considering several ultrabooks, with GPUs ranging from the mx150 (mobile 1030) to the 1050ti. Is there a way for me to downclock my 1060 and/or disable cores and memory chips, in order to get an idea of how these cards would perform for my use case? The few topics I've seen on other forums seem to indicate that this isn't possible, but I know there are some, like, really smart, very stable geniuses here.

[TheRandomness]

No there is not unless you write your own BIOS to the card, which has a much higher chance of bricking the thing. Just look at benchmarks online. 

[Jurrunio]

At most you can emulate, are mobile 1060 and Max-Q 1060.

[Bananasplit_00]

you can BIOS mod it, which can also break your card if you screw it up

[faziten]

You can "emulate" the results but you need to know this first:

 

If the software you will be using scales linearly with #cuda cores and Frequency and memory you can predict fairly realistic results.

Pascal is a scalar architecture, there is a strong correlation between performance and cuda core count and frequency assuming the software is capable of using the higher cound and higher speed cudas efficiently.

 

Lets do quick maths: (Worst Case is considered to be a perfect scaling software -not games-)

So 1060 and 1050ti relative performance would be:

1060=    1280cudas*2000freq= 2560000z

1050Ti= 768cudas  *1700freq= 1305600z giving a worst case scenario of 51% relative performance.

 

Lets take a peek on real world gaming benchmark:

I'm arbitrary using this video, take any other and do the same math.

https://www.youtube.com/watch?v=PM2rGZkqHbk

 

Battlefield 1: (stock cards).

1060 6gb= 1280*1860= 2380800

1050Ti= 768*1772=1360896

Relative expected performance (worst case): 57%

Now lets take some readings:

I pause randomly at some points in the video:

0:34 shows 1050Ti= 40FPS; 56FPS=1060

Relative performance observed: 71%.

0:44 shows 1050Ti=41fps; 1060=62FPS.

Rel. Perf Obs= 66%.

0:54 shows 1050Ti= 51FPS; 1060=65FPS

Rel. Perf. Obs= 78%.

 

Not as tight as expected but within parameters. (this is not to extend for every case, it's just ONE observation).

 

Another observation would be like in this video (2:20) mining virtual coin.

1050Ti= 114.724 MHash/s

1060 6GB= 193.464Mhash/s

Rel Perf Obs=59% (thighter compared to the theoretical 57%, miners scale really nice).

 

Disclaimer:

I know this user did not disclose working speeds, im asuming they are stock (and i do understand this assumption is quite heavy but i don't intend to write an academic paper, just show you how you could take an educated guess about expected performance margins).

Apply this method and you'll have a better guess about what you can expect.

This is not far from anecdotal data, dont be fooled by the presentation, i used only two examples to show the procedure.

Remember: This only applies to PASCAL based GPUs, and remember i ruled out the VRAM speed and bus width (because it only impact a handful of software, its rare to say the least).

 

Last words:

Limit your search to the critical software you want to work with and try to find data about performance and scaling. If you are lucky enough, you should end up with a realistic aproximation.

 

 

tl;dr: 1050Ti should give between 1/2 and 4/5 of 1060 6GB performance depending on the software used.