4*4090 RTX is not twice as fast as 2*4090 RTX

Tutorials, Tips and Tricks
, ,
1

I’ve recently rented a Render Farm, I’ve used 2 out of many options: 2x4090 and 4x4090.

4x4090 costs twice as much as 2x4090 (100$ / 3.3 hours vs 50$ / 3.3 hours) but is not as twice as fast as 2x4090.

4x4090 rendering a scene takes 9.5 seconds:

cut
cut1288×680 428 KB

2x4090 rendering the same scene takes 11.5 seconds:

cut2
cut21302×722 450 KB

So the difference in speed is 2 seconds, which is 19%.

Is 4x4090 worth it?
By the way, what’s happening? Is this the result of the Diminishing Return phenomena?

Please share your thoughts. Thank you.

2

The CPU power isn’t unlimited, and has to supply the video card with (and receive from it) information, as well as do basic tasks. Blender has to keep up with the requests it’s making.

So it’s natural that there’s a point at which it’s doing all it can, regardless of how many GPUs.

3 Likes
3

From a business perspective, I would not hire 4 people to do the work of 2.2 people. The observed fact is that 2 gpus = 11.5 sec and 4 gpus = 9.5 sec. Not worth the money. Use 2 gpu’s.

As Thorn noted increasing overhead can reach a point of no further gain possible. That is true economically too but you would stop adding expense long before zero return. (Aww, there is no Ferengi emoji.)

However, render farms are designed to be very scalable and able to apply large numbers of gpu/cpu’s to renders. If this one is reaching no useful gain from 2 to 4 then either something is wrong or this is not a good render farm. Probably something is wrong. I recommend asking customer service about it or searching their forum/help/etc.

Am I on here again? From a business perspective, if I prepared and sent bills to my customers they might pay me. Discipline, focus …

3 Likes
4

Great answer! I noticed that you were typing your comment for so long lol

They don’t even know what Deminishing Return is and had to Google when I asked about it, yeah, 2*4090 is what Imma use from now on then.

5

I lacked discipline and focus while typing lol.

Don’t say stuff like diminishing returns. Just tell them about the render times and ask why it isn’t almost 2X. Or …

If an employee isn’t doing his job fire him and hire one who does. Maybe a different render farm is the solution.

Hmm, I’m not doing my job but if I fire me then I’ll end up living in box in the alley behind that suspicious massage parlor.

2 Likes
6

Noted, good point, I’ll ask a different Customer Support Agent then, there are many of them actually.

7

Everyone, I’ve contacted an employee there and they said that it took too short (i.e. seconds, in this case) to notice the difference. They said the difference would have been more comparable if it took multiple minutes to render a frame.

But, I rendered a frame on my 1050ti and it took 9 seconds whereas it took 1.5 seconds on their 2x4090. Even though it took seconds to render on both machines, the difference still was apparent, which is 83%.

8

Employee makes sense.

The original post said 2x4090 took 11.5 sec. Latest post says 1.5 sec. Typo? Which is right?

9 sec without render farm vs. 9.5 or 11.5 sec with it means little or no advantage in using the farm for this project.

Frame. This is an animation. More frames means more value in using a farm even if just to free your machine for other use.

1050ti. Upgrade to 1 or more bigger but old used cards could be worth it.

1 Like
9

There are two main factors that impact rendering speed, the initial CPU work required to prepare all the data for rendering and then the actual rendering itself.

Much of the initial CPU work can be single threaded (so a fast single core speed can be better then having many slower cores) and depending on how complex the scene is, the initial prep work can take a bit of time.

Even if the initial CPU time isn’t very much, if the total render time is fairly short (like under 60s or worse, under 15s) then a larger portion of the total render time is actually taken up by the CPU and hence throwing more GPU’s at the render only part has way less of an impact.

Your render is pretty much a worse case. The render farm is using a Threadripper as the CPU, so single core performance slower then many standard desktop PC’s, hence initial prep work takes longer.

Then, you are using 4090’s GPU’s, which are “DARN” fast and you are giving them very little to work on, the render is just to easy and hence done too fast. You likely would have been better off just using a single 4090, which I half bet would be still under 15s render time. It’s just too short and easy a scene to give the 4090’s a real chance to warm up and go to work.

Keep in mind, that using more GPU’s adds more work/CPU time, as Blender then has to manage the rendering across multiple devices and then bring it all back together for a final image.

If all you are doing is rendering single images that take less then 60s, then scale that back to just a single 4090 and save a bunch of money. The only reason I would use 2 or 4 GPU’s on such short render times, is if it was for an animation, in which case, you make each GPU render 1 frame, then you would see near 4 times speed increase.

Outside of that, if all you have is a 1050ti, consider a GPU upgrade (unless its a laptop), pretty much any cheap RTX based card will blow that 1050ti away and not using a render service would pay for it pretty fast.

3 Likes
10

Hi!

The 9.5 sec vs 11.5 sec (4x4090 vs 2x4090)
and
1.5 sec vs 9 sec (2x4090 vs 1x1050ti)
are 2 different cases, both are rendering a frame from an animation.

I personally it is worth it as the difference between 9 sec (1050ti) and 1.5 sec (2x4090) is 83%.

Thank you!

11

Seems unlikely that the render farm speed would vary that much so I suspected a possible typo of leaving out a “1” turning 11.5 into 1.5).

So 2x4090 took 11.5 sec in one case and 1.5 sec in another? Both on the same render farm? With that much variance you need more data to determine what the farm can provide. Several tests each covering several frames are needed. Average the per frame times to get the typical performance you can expect from the farm. Hopefully the 11.5 sec is an outlier.

1 Like
12

Yes sir, because I rendered different scenes for different cases.
Heavier settings took 11.5 sec, lighter settings took 1.5 sec.

Man, of course it must have had variance for completely different settings and completely different scenes. I don’t understand why you keep comparing 2x4090 in 2 different cases, I’m focusing on “4x4090 vs 2x4090” and “1x1050ti vs 2x4090”, not “2x4090 vs 2x4090”.

13

There was nothing in your posts to tell that a different scene and settings was used. I can’t read your mind.

It is good that the farm will be useful for some scenes. Happy rendering!

1 Like
14

Got it, I apologize for bothering you. It was my bad. I thought saying “rendering a frame” instead of “rendering the same frame” was obvious enough of a sign.