More cuda cores or more GPU RAM?

Hello.
I need some help to choose a GPU.
I do architectural renders, but not that kind of super realistic visualizations, with a lot of furniture, plants,… What I do is more like working renders, to get an idea of how the project looks like. I use image textures, to figure out how different materials will combine. Of course, at the end I want to show the renders to other people, but it’s not the main purpose.
While designing, I’m doing changes all the time, and I have almost permanently a viewport in render mode. So it’s good for me if this render refresh is fast.
I also use several lights, specially when rendering interiors.

I’ve found 2 GPU with similar prices:
-1050ti with 4Gb… 199€
-1060 with 3Gb…215€

So, what would be better for me? The more RAM of the first one, or the more cuda cores of the other?
I understand that cuda cores are the “engine” that makes all the calculating, so more cores should give more rendering speed, and RAM is necessary to load textures, and I’ve read that if you get short of RAM then Blender/Cycles crashes. I don’t know what I need when using a lot of lights.

I’ve been doing a test with 2 architecture example models: Mies Pavillion and Classroom. While rendering, there is some kind of data at the top of the rendering window (time, remaining time,…), and seems like the peak RAM used is 297M.
So, can I consider that 3Gb is good enough to these type of renders?

Thank you.

1060 is faster than 1050 Ti. Is 1060 with 6GB vRAM out of your budget?

One clarification, for GPU you should monitor vRAM usage while rendering with an external application like GPU-z (nvidia-smi on linux). Blender does not correctly indicate the vRAM usage with CUDA. if you use CPU perhaps what blender informs is more precise, but, use an external System/RAM monitor to see what are the peaks reached.

Hi Yafu,
Thanks for your answer.

I’ve not an specific budget, but searching GPU’s it’s like “for 20€ more you have this…and for other 20€ this…and for a little more…”. So you start looking at 1050 GPU and in a few minutes you are looking for 1070 GPU…ha, ha…
Of course I know that with more cores, more RAM, more money, you have a better GPU. But I want to know how different characteristics work. I’m loosing too much time to take a decision.
I don’t want to spend 300€ in a GPU with a lot of RAM if I’m not going to use all these Gb’s. At least I think that if I have more cores, I will always take some profit of them, because a faster rendering is always welcome.

Now I’m rendering with CPU (Intel i7 6700 3.40GHz, 16Gb RAM, Windows 10 64 bits). So I’m almost sure that any GPU will be faster in rendering, but I don’t know what’s going to happen with the vRAM.

You should not be so sure about this:
https://blenderartists.org/forum/showthread.php?375718-Cycles-GPU-CUDA-slow-with-some-materials

If you really want to be sure that GPU will always be really faster than your CPU in complex scenes, I would go at least for a 1070.

More CUDA cores do not necessarily mean more speed. So try to see some benchmarks:

Speed on Cycles from lowest to highest: 1050 Ti - 1060 - 1070 -1080 -1080 Ti
More vRAM is always good.

So everything depends on what you think is enough for you :slight_smile:

Edit:
Clarification with BMW scene used in Phoronix: this is the scene downloaded from blender.org - demo files.This is not BMW27.blend from blenderartists.

Hello again.

Yafu, thank you for the links, but it seems to me that my use of Blender is different (I don’t use “hair”, for example). The benchmarks that are closer to my way of using Blender are “Mies Pavillion” and “Classroom”, and both show a faster rendering time when more cuda cores.

I don’t create complex scenes. I don’t use a lot of furniture, plants, people, cars,… I simply model a building, use a few good image textures (.jpg), and some lights (sometimes just 1 sun, sometimes 5 lights, sometimes 15,…). It looks more like a traditional model, not like those ultra realistic visualizations.

I have a viewport in “render” model while using Blender. And I do a lot of changes (cameras, lights, textures,…) all the time, so the render viewport is refreshing constantly to see these changes.

So I’m trying to understand what is better for the way I use Blender: more cuda cores or more vRAM. Of course I undertand that more of everything always works.

Let me ask something a little bit different. What is better for me?:
-2 GPU’s 1060 with 3Gb each one (which means 3Gb total, because they don’t add, but with 2 GPU’s calculating 50% of the tiles at the same time). Price: (2x215€) 430€.
-1 GPU 1070 with 8Gb. Price: 429€.
In Blenchmark I can see:
-2x1060 (display)…41s.
-1x1070…72s.
Of course, there’s a big difference in vRAM (3Gb vs. 8 Gb), but I don’t know if that’s important to me or not. That’s the question. I need to know how vRAM works in rendering. Specially when refreshing a viewport in render mode.
(More info: I don’t play games with the computer. And my PC is ready for 2 GPU’s: 2 PCI slots and a 700W source with two 6+2 pin connectors)

Yes, but I was talking in general, is that what you said: “so more cores should give more rendering speed”…It is not entirely true when comparing different families. And also by the title of the thread that is general, someone could be confused.

I’ve read people working on Archviz say they use large amount of large textures. If that’s not your case, then I guess 3GB vRAM might be enough for you.

Classroom scene takes 800 MiB vRAM with my GTX 960. So if you work with those kinds of scenes, those two GTX 1060 may be a good choice.

Get the 6GB GTX1060, period. It has the best price/performance of NVIDIA GPUs right now and it’s also faster than the 3GB version. If you run out of RAM, you’re dead in the water, so don’t skimp on it to save a few bucks.

Disregard the RAM usage displayed in Blender, it doesn’t include RAM used by tile data and the kernel itself. Also some RAM is required for display, which Blender can have no knowledge of. You can use a tool like GPU-Z to see total RAM usage.

Hello.

it’s also faster than the 3GB version

blenchmark.com (BMW test):
-GeForce GTX 1060 3Gb: 75s.
-GeForce GTX 1060 6Gb: 77s.
3Gb GPU is 2s. faster (but this difference is irrelevant to me)

What I’m trying to figure out is how to know how many RAM is enough, and if it’s better to put money in cuda cores or in vRAM memory. Any comparison among solutions with different prices is not “fair”.

As far as I know (which is not very far… he, he…), cuda cores are always “working”. So any more money you spend in more cuda cores has a reward as a faster rendering. But you can buy a lot of vRAM and perhaps you will never use all of it. That’s the point.
I know that running out of vRAM means crash, but perhaps it never happens. It sounds like “pay an expensive car/home/whatever insurance, because perhaps someday something happens”. 3Gb or 6Gb are numbers with no sense to me. But 41s. or 72s. have a lot of sense.

When I see these benchmark tests, it seems that using 2 cheaper GPU’s gives more speed than 1 more expensive. Another example:
-2 GTX 1050ti 4Gb (display): 49s. Price 159€x2= 318€
-1 GTX 1080 8Gb (display): 61s. Price: 595€.
-1 GTX 1070 8Gb (display): 62s. Price: 419€.
-1 GTX 1060 6Gb (display): 82s. Price 309€.
I know…you have less vRAM with the two 1050ti, but…what if you never need more than 4Gb?

The fastest renderings in Blenchmark are made with a combination of various GPU’s, which means that more cuda cores are used, unless the vRAM has not increased.
I can understand that somebody decides to buy one 1070 instead of two 1050ti, for example, because requires a lot of RAM, or needs an SLI-ready GPU, or wants to use VR. And these differences make the higher price reasonable. But what if you don’t need any of these things?

blenchmark.com (BMW test):
-GeForce GTX 1060 3Gb: 75s.
-GeForce GTX 1060 6Gb: 77s.
3Gb GPU is 2s. faster (but this difference is irrelevant to me

Blenchmark results are useless, the tile sizes chosen are too small. The 6GB has 1280 CUDA cores, the 3GB one has 1152, which is not necessarily something that shows up in a benchmark. Also, these GPUs have variable clocks, which adds additional variance to testing.

But you can buy a lot of vRAM and perhaps you will never use all of it.

You can use up VRAM easily, especially with a GPU used for display.

When I see these benchmark tests, it seems that using 2 cheaper GPU’s gives more speed than 1 more expensive. Another example:
-2 GTX 1050ti 4Gb (display): 49s. Price 159€x2= 318€
-1 GTX 1080 8Gb (display): 61s. Price: 595€.
-1 GTX 1070 8Gb (display): 62s. Price: 419€.
-1 GTX 1060 6Gb (display): 82s. Price 309€.
I know…you have less vRAM with the two 1050ti, but…what if you never need more than 4Gb?

Where did you get these numbers? Again, forget about blenchmark, it penalizes GPUs with more cores even more.

The prices that I see have the 1060 starting at 250€, the 1050Ti starting at 150€. The 1060/6GB has 60% more cores and 75% more memory bandwidth, which would make it a better deal by all metrics. If you want to get two GPUs, get two 1060s, or get one now and add one later.

Yeah the 3GB gtx1060 is a weird thing, don’t buy it. Years and years ago 2.5GB GPU was production standard and back then that still wasn’t enough either. 6GB gtx1060 you have access to about 5.4GB of that so those 2 and 3GB cards are becoming outdated by now even if they have modern speeds. The more memory dump in cycles gives you that carefree feeling on expanding your scenes in ways you otherwise would never even attempt.

I’ve been doing some tests, with the Resource Monitor of Windows 10. All the time using CPU, no GPU. All data is referred to Blender, not the complete PC.

Model “Classroom”:
-with a wireframe viewport, Blender uses 675Mb of RAM, and 0% of CPU.
-with a render viewport, Blender uses 1040Mb of RAM, and 97-99% of CPU while rendering tiles. Once viewport rendering is finished, still uses 1040Mb of RAM, but 0% of CPU. I guess Blender maintains the memory used because the rendered viewport can need a refresh at any moment.
-when creating a render image (render button), uses also 1040Mb of RAM and 99% of CPU. Once finished the rendering (or interrupted), RAM use goes to 700Mb and CPU use to 0%. While rendering, viewport shows “Mem,297.00M”, which is close to the difference between 1040 and 675.

Model “Barcelona Pavillion”
-with a wireframe viewport, 315Mb of RAM. 0% CPU.
-with a rendered viewport, 530Mb of RAM. 75-85% CPU (constantly changing). Once viewport is rendered RAM maintains the 485Mb, but CPU use goes to 0%. Viewport shows “Mem.158.82M”.
-when rendering an image (render button), RAM goes to 550Mb, and CPU at 99%. Viewport shows “Mem.161.02M”.

Model “BMW”
-Uses 440Mb of RAM while doing the rendering test (Blenchmark). 290Mb when the test is finished. Viewport shows “Mem.119.07M”, wich is pretty close to the difference between 440 and 290.

While rendering some of my working models, RAM use goes to 200-300Mb (I don’t create complex scenes), and CPU use is 99%.

So…
It seems that, in my case, the use of RAM in Blender is far from 6Gb, or 4Gb, or even 3Gb. I don’t know how percentage of RAM will pass to vRAM when using a GPU, but I guess that it never will be more than the total CPU RAM used now.
But with any kind of model, CPU use goes to the maximum.
Classroom model uses more RAM, probably because uses more image textures (and perhaps also because uses more light points, I don’t know).
My conclusion is that more calculating power (more CPU or a lot of cuda cores) will give more speed, but more RAM is not necessary in this examples.
Somebody agrees?

It seems that you are looking for someone to approve a risky decision for you :slight_smile:

People recommend you something safe, otherwise you are the one who knows best what kind of scenes you work with.

Some points to consider:
Render with CPU may not consume the same memory as with GPU. According to developers, there are certain things CUDA uses in vRAM that are out of control of what Blender developers can handle (as I think it was the explanation they had told to me sometime about why Blender can not display the correct use of vRAM)

Luckily vRAM usage in Cycles has been very well optimized for a few versions, but not so long ago this was a headache for many:
https://developer.blender.org/T43310

And that had taken a long time to get fixed.

So today, fine. Tomorrow, I do not know.

As it seems the scenes in which you work do not use much vRAM, and 3Gb or 4Gb can walk safely for you. Again, you have the final decision.

You literally haven’t paid attention, at all. You cannot measure these things with just a CPU.

You need to measure your VRAM usage with a tool like GPU-Z, for the entire system. Your desktop, your windows, your applications all take physical video memory which may not be freeable when you need it. Windows 10 reserves even more memory (almost 20%) that applications cannot use.

You also need to measure with the appropriate tile sizes. Larger tiles are required for GPUs to perform fast (256x256+), but they also take more memory. More passes also take more memory. The GPU kernel takes memory.

I believe you don’t want advice, you want your preconception fulfilled. Go ahead, get the smaller VRAM, but then don’t post about your “CUDA out of memory” errors…

Hello again.

I had no preconception of this question, because my knowledge in this matter is so small. But I started this thread to get an idea of what could be better for my way of using Blender, not to find the best GPU for a general use. And thinking about this is how I started to put the focus on vRAM, because I see that more cuda cores are always good (more calculating power…more speed…), but perhaps a lot of vRAM is not necessary for me. Different GPU’s also have other functions, like SLI, VR, more outputs,…and I’m sure that lot of people think about this functions when choosing a new GPU. So, I don’t understand why I can not try to figure out how many vRAM would be good enough for me.

For the same price, I can buy one 1060 with 6Gb or two 1050ti with 4Gb. One solution gives more RAM, and other gives more speed. That’s the kind of question I wanted to solve. And some of the information I’m receiving in this thread helps me, of course. But the typical (and obvious) “more RAM is better” is not the kind of thing I was expecting for.

I have test RAM usage with CPU because I don’t have GPU. But I understand now that data could be different measuring vRAM. On the other hand, perhaps not all the memory that Blender is needing from my CPU now will be used in the future GPU. I guess that some of the actual RAM usage will not pass to vRAM. Anyway, I understand that there’s not any formula to solve this.

Well, thank you for your comments.

You were told everything there is to know about this, you just ignored that information.

Nobody here cares if you run out of VRAM, you’re given reasonable advice, which is to favor more VRAM at the risk of not needing it.

You’re also looking for arguments to buy the 1050Ti, which is inferior even when taking into account price/performance. Again, nobody here cares if you spend your money unwisely.

For the same price, I can buy one 1060 with 6Gb or two 1050ti with 4Gb. One solution gives more RAM, and other gives more speed.

You don’t even know that. You brought out some numbers, but they’re probably done with the wrong (i.e. default) settings. Again, everything from blenchmark is useless.

Let me do some math for us all:
One 1060/6GB gives you 1280 (1.0x) cores at 1506Mhz (1.0x)
Two 1050Ti give you 1536 cores (1.2x) at 1290Mhz (0.85x)

Your “ideal” speedup from just cores would be (1.2 * 0.85) = 1.02 and that’s ignoring the much lower memory bandwidth of only 112GB/s vs 196GB/s.

So, a back-of-the-envelope calculation for “have more cores” doesn’t even pan out.

Hi again.

BeerBaron, I’m not looking for arguments to buy a 1050 Ti. I’m looking for arguments to decide what is better for me.
Actually, if you read some of the options I have exposed, I’m opened to other GPU’s, but always under the condition of having the best relation cost/speed.

I’ve found 2 GPU with similar prices:
-1050ti with 4Gb… 199€
-1060 with 3Gb…215€

In that example I was suggesting that perhaps 1060 3Gb was a better option for me than 1050Ti 4Gb.

-2 GPU’s 1060 with 3Gb each one (which means 3Gb total, because they don’t add, but with 2 GPU’s calculating 50% of the tiles at the same time). Price: (2x215€) 430€.
-1 GPU 1070 with 8Gb. Price: 429€.
In Blenchmark I can see:
-2x1060 (display)…41s.
-1x1070…72s.

I was comparing 1060 and 1070. Nothing about 1050Ti.
I have no budget, but if I spend 200€, or 400€, or 500€, I want to be sure that I’m choosing the best option for me.
So, definitely I’m not trying to get your approval for any particular GPU. I’m trying to get information like this:

Let me do some math for us all:
One 1060/6GB gives you 1280 (1.0x) cores at 1506Mhz (1.0x)
Two 1050Ti give you 1536 cores (1.2x) at 1290Mhz (0.85x)

Your “ideal” speedup from just cores would be (1.2 * 0.85) = 1.02 and that’s ignoring the much lower memory bandwidth of only 112GB/s vs 196GB/s.

Although I have to say that I don’t understand why 2x1050Ti are faster in Blenchmark test than any other GTX 1000 series GPU.
I know you don’t like this test, but it’s the best I’ve found about GPU’s and Blender. Other benchmark about GPU’s have a more general point of view, but I don’t use the computer for video games, Photoshop, video editing,… The only “hard” work I’m going to do with the computer is 3D modelling with Blender.

I don’t know what else to say. I don’t want you to think that “I’m not listening”. I was trying to get objective information: here, Blenchmark, other Cycles tests, other webs,…, but it seems that this matter is not easy to solve scientifically.

Thank you very much for your help.

Hi, we do like Blenchmark but it is not usable for GPU.
The results are so wrong if you have a closer look.
For me VRAM is more important than speed today.
I have a 2 and a 4 GB card, for small scenes I use both and for bigger scenes the 4 GB card.
But I cant render more complex scenes with particles, hair or high resolution meshes and textures.
I cant render any of the production files of the latest Blender movies.
Take a look at your scenes, what do you need on CPU.
Do you need more next year?
I got my GTX 670 2 GB very cheap but I would never buy a card smaller 6 GB today.

Cheers, mib

Huh? I quoted you multiple times where you put two 1050Tis against one 1060, so that’s what my answer is based on.

I have no budget, but if I spend 200€, or 400€, or 500€, I want to be sure that I’m choosing the best option for me.
So, definitely I’m not trying to get your approval for any particular GPU.

I don’t care which GPU you get. Quite frankly though, as a result of my reply you should’ve immediately fled the building and bought a 1060, that’s how solid my advice is.

How much more time are you going to waste on making a decision? Time is money!

Although I have to say that I don’t understand why 2x1050Ti are faster in Blenchmark test than any other GTX 1000 series GPU.

Because the tile sizes are wrong. That’s why blenchmark is useless. The 1050Ti has way less cores than the high-end GPUs, that’s why it doesn’t suffer from small tile sizes as much.

I don’t want you to think that “I’m not listening”.

That happens when I have to write things three times over.

BlenchMark is not good for GPU because it uses an old scene that is not for production and in addition it uses an old configuration of tiles sizes. The people in this forum have verified that the results of BlenchMark are not reliable for GPU.
You could look for more reliable benchmark results, with more robust scenes. You could see Phoronix results I showed you above for example. A quick calculation I can do by seeing some results is that two 1050 Ti could be about 20% or 30% faster than a single 1060 card.
About vRAM, I’ve already said it. You are the only one who knows what kind of scenes you will be working on (currently and in the future), and in which OS you will use Blender (because using of vRAM that the OS does). The only thing I can show you is the use of vRAM that does some scene in my GTX 960, so that you have an idea (I’ve already told you about classroom scene). To that you must add the use of vRAM that does the OS (if you are using the same card for display, remember you have an intel iGPU).

I can pretty much guarantee whichever one you choose you’ll be disappointed. If you take a cheaper one, you’ll either exceed the available vRAM because you’ll want to do more complex things later on (because we all do that sooner or later), or wonder if the more expensive one would indeed be faster. If you take a more expensive one, you’ll be wondering if you really could have gotten away with the cheaper one.

Jettison the math. Bite the bullet, get a decent mid-range one that exceeds your requirements. The extra capacity will be there if necessary and if not, it will give you some room to play with. If its a bit slow (which it probably won’t be) or you do want faster render times you can supplement it with another one later.