Nvidia for professional 3D applications
Half a year ago, I was looking for a video card, in which I could do 3d modeling, and GPU rendering. In connection with the appearance on the market of a large number of renderers on CUDA, I was impatient to purchase a video card with CUDA support, namely Nvidia .
As some already know, Nvidia puts up for sale video cards of several models Geforce, Quadro, Tesla, ION, Tegra. In this short comparison, we miss ION and Tegra, because Designed for mobile devices and weak in performance.
We need power!
')
WHAT A MANUFACTURER SAYS
Geforce - video cards, focused on the consumer market and gamers, in particular.
If you are interested in games - Geforce is the best option for this.
Video cards show themselves best in games; they have high frequencies, are not expensive, and are most voracious under load.
Zhefors are rarely mentioned as common computational tasks (Cuda, OpenCL).
It has PhysX, called the coolest hardware solution to accelerate physics.
Leisure holder Geforce (Battlefield 3).
Quadro - video cards for users of professional 2D and 3D applications.
If you deal with 3D modeling packages, CAD, complex vector graphics, then Quadra will suit you.
Complex models on the screen are rendered faster, less "jerks".
Quadras comparable in performance to Gefors in games will be several times more expensive.
On the pictures of the site nvidia you can already see more Kuda than on the Zhefors.
I mean, video cards are professional, even general purpose computing will be!
The work of the owner of Quadro (Autodesk Alias ​​Studio).
Tesla - computing systems for general-purpose scientific and technical computing.
Here CUDA is advertised in its entirety as the coolest general-purpose computing tool. Everywhere posters with aerodynamic calculations, voxel scanning of the human body, graphic models of loads, and unrealistically fast rendering on iRay.
There are no video outputs on Tesla, as well as no hardware rasterization: neither OpenGL nor DirectX works.
The work of the owner of Quadro + Tesla (Quadro - 3d graphics, Tesla - molecular dynamics).
***
SMALL RESEARCH
When I began to understand their differences, I was surprised by the fact that GeForce, Quadro, Tesla video cards use the same graphics chips.
Consider a video card with the same, not the newest, GF100 chip (512 CUDA cores):
Single-chip:
GeForce: GTX465, GTX470, GTX480
Quadro: 4000, 5000, 6000
Tesla: C2050, C2070, M2050, M2090
Consider one representative from each family in more detail.
GeForce GTX480
Once the top gaming graphics card.
Cost: at the time of release about $ 500 (now used and for 300 saw), currently not produced ( GTX580 512 cores replaced, and GTX680 1536 cores)
The number of CUDA cores is 480.
Memory size 1.5 Gb.
Float performance:
Single precision: 1344.9 Gflops.
Double accuracy: 168.1 Gflops.
(There is a more trimmed version of the GTX470, now you can find it for less than $ 250, 448 CUDA cores, 1.25 Gb)
Quadro 5000
One of the best video cards for professional applications.
Cost: according to Amazon about $ 1700. Released.
The number of CUDA cores is 352.
Memory capacity 2.5 Gb.
Float performance:
Single precision: 718.08 Gflops.
Double accuracy: 359.04 Gflops.
(It is worth paying attention to the Quadro 6000, 448 cores, 515 Gflops double precision, $ 4000)
Tesla C2075
Cost: according to Amazon about $ 2,200. Also released .
The number of CUDA cores is 352.
Memory capacity 6 Gb.
Float performance:
Single accuracy: 1030 Gflops.
Double accuracy: 515 Gflops.
What do we see?
Note that float performance GeForce GTX480 wins. The reason for this is the largest number of working cores and the highest frequencies among analogs. It is necessary to convert the coordinates of objects in games, calculate shadows, calculate pixel and vertex shaders. In the end - to the game "flew."
But for Tesla and Quadra to buy for scientific research, modeling the dynamics of liquids and gases, in double precision the performance is severely curtailed and inferior to counterparts.
Performance Ratio:
GeForce : double / float - 1/8
Quadro and Tesla : double / float - 1/2
In addition, the smallest memory has the same GTX480. It is enough for games, but if you want to calculate aerodynamics - buy something more serious.
***
WHAT DO I NEED?
(People involved in 3d graphics)
1. Smaller brakes when editing a 3d model.
2. Some are interested in the possibility of fast rendering on the GPU.
3D performance GeForce vs Quadro
From the information above, it may seem that professional applications on the GeForce are not used due to the fact that it has a small amount of memory, but it is not.
The video will show you why the “bad Quadra” is better than the “good Zhefors” in professional applications.
Quadro 600: 1Gb, 96 CUDA cores, 150u.e.
GTX560Ti: 1Gb, 384 CUDA cores, 250u.e. (Pricing from Amazon)
It turns out that Nvidia carefully ensures that 3d productivity in professional Geforce applications yields to Quadro at comparable prices.
How can viewport brakes be implemented?
The fact is that the number of polygons in games is significantly less than that of professionals in professional applications. In games, it rarely reaches one million polygons, and in professional games - tens of millions.
Here you can do this: cut performance when converting vertex coordinates. If there are more vertices than a certain number, then put a delay before the next vertices are drawn.
Or set a delay when drawing triangles. If more than a certain amount - then put a delay before drawing each subsequent triangle.
A little lyrical digression, or Nitrous in 3ds Max .
I was misled by the Nitrous engine in 3ds Max, which stands next to OpenGL and DirectX. What is it like? In Autodesk, there is something that causes Nitros, whose hardware support, it turns out, is on every self-respecting video card, but only 3D Max knows about it?
Well, you can make a small logical chain. Autodesk is a rich corporation, and in good partnerships with manufacturers of ATI and Nvidia. Raise the need to sell your brainchild! How to interest consumers? Performance is the same!
So, GeForce GTX580 (yes, I bought it), 7.3 million triangles, 2560 Torus Knot, without shadows and without Adaptive degradation.
Nitrous - 42 fps; Direct3d - 13 fps; OpenGL - 2 fps.
OpenGL - slow. DirectX is much better. And Nitrous - the coolest of all, it turns out! What then nitros then?
Two options:
1. This is OpenGL / DX in which additional brakes have been removed in the viewport, created intentionally in OpenGL / DX modes.
2. This is OpenGL / DX, which is able to access the hardware functions of gaming video cards, and manifest quadra abilities in them!
And I am inclined to the 2 variant, since in Blender and in Rhino3D, this very wildly slows down (2fps).
It turns out that users of 3ds Max and other Autodesk products are not at all so fundamentally switching to Quadra? Unfortunately, I do not have Quadra to check the performance of Nitros compared to OpenGL.
If you have a GeForce or Radeon, there is no desire to fork out for Quadra, you do not use products from Autodesk, and you have very complex models, then:
1. Complex objects can be hidden. Objects can be shown in a viewport with a lower grid density.
2. Instead of objects, you can show "containers" containing them.
That is, follow the number of polygons in the viewport, if you have really "heavy" models.
But in the game you play normally.
GPU rendering
Since commercial producers do not talk about what data types (float or double) they use - you have only to guess.
iRay is shown everywhere with Quadro and Tesla, it may give the impression that iRay does not work with GeForce at all.
Picture from the office. nvidia site.
But no, it works, and how. It would seem, what could be better for non-graphic calculations than a Tesla video card, specially sharpened for non-graphic calculations?
(Taken from the post: "V-Ray and Iray. Comparison and Review" )
GeForce GTX580 is the fastest single-chip graphics card in iRay GPU rendering. And much cheaper than "serious" analogs of the same performance. And if you miss 1.5GB, there are GTX580s with 3GB of memory.
When using V-RayRT, Octane, Cycles, Arion, the GTX570 and 580 video cards are also the best to show themselves. Does this mean that all these renderers do not use double precision calculation for rendering?
In any case, if you want to render on the GPU, you can save well on the GeForce.
GTX680
But the corporation noticed that for calculations, the GTX580 began to take more and more often, the double performance of the GTX680 is inferior to the float not by 8 times, but by 24, which could not but affect some tests .
It is known that in Octane Render performance increased by 64%.
ATI Radeon vs FirePro
Similar to Nvidia, AMD has also divided the video card models. Radeon (analog GeForce), FirePro (analog Quadro), FireStream (analog Tesla). The performance of double-precision floating-point calculations is inferior to single by 4 times, in all ATI models. Interestingly, the performance of ATI's top-end gaming graphics cards ( Radeon HD 7970 , float - 3.79 Tflops, double - 947 Gflops) exceeds even the single-chip Tesla in double precision. It should be noted that the performance in flops is not always an indicator of the performance of iron in specific cases.
The reason why ATI is much inferior to Nvidia in the GPGPU market is not yet clear to me. Maybe the gaming segment is enough.
Choice?
I chose the GTX580 3Gb. The video card gives you the opportunity to enjoy new games and GPU render performance. And the brakes in the viewport of 3d modeling packages are not very critical for me.
The author of the article with respect to this manufacturer, and he is the happy owner of the card Nvidia.
Such marketing moves are an integral part of a market economy, they are resorted to by all manufacturers without exception.
But nevertheless, we will not be conducted on marketing tricks of corporations, and thoughtfully buy what is really useful for us!
As some already know, Nvidia puts up for sale video cards of several models Geforce, Quadro, Tesla, ION, Tegra. In this short comparison, we miss ION and Tegra, because Designed for mobile devices and weak in performance.
We need power!
Nvidia power ...
')
WHAT A MANUFACTURER SAYS
Geforce - video cards, focused on the consumer market and gamers, in particular.
If you are interested in games - Geforce is the best option for this.
Video cards show themselves best in games; they have high frequencies, are not expensive, and are most voracious under load.
Zhefors are rarely mentioned as common computational tasks (Cuda, OpenCL).
It has PhysX, called the coolest hardware solution to accelerate physics.
Leisure holder Geforce (Battlefield 3).
Quadro - video cards for users of professional 2D and 3D applications.
If you deal with 3D modeling packages, CAD, complex vector graphics, then Quadra will suit you.
Complex models on the screen are rendered faster, less "jerks".
Quadras comparable in performance to Gefors in games will be several times more expensive.
On the pictures of the site nvidia you can already see more Kuda than on the Zhefors.
I mean, video cards are professional, even general purpose computing will be!
The work of the owner of Quadro (Autodesk Alias ​​Studio).
Tesla - computing systems for general-purpose scientific and technical computing.
Here CUDA is advertised in its entirety as the coolest general-purpose computing tool. Everywhere posters with aerodynamic calculations, voxel scanning of the human body, graphic models of loads, and unrealistically fast rendering on iRay.
There are no video outputs on Tesla, as well as no hardware rasterization: neither OpenGL nor DirectX works.
The work of the owner of Quadro + Tesla (Quadro - 3d graphics, Tesla - molecular dynamics).
***
SMALL RESEARCH
When I began to understand their differences, I was surprised by the fact that GeForce, Quadro, Tesla video cards use the same graphics chips.
Consider a video card with the same, not the newest, GF100 chip (512 CUDA cores):
Single-chip:
GeForce: GTX465, GTX470, GTX480
Quadro: 4000, 5000, 6000
Tesla: C2050, C2070, M2050, M2090
Consider one representative from each family in more detail.
GeForce GTX480
Once the top gaming graphics card.
Cost: at the time of release about $ 500 (now used and for 300 saw), currently not produced ( GTX580 512 cores replaced, and GTX680 1536 cores)
The number of CUDA cores is 480.
Memory size 1.5 Gb.
Float performance:
Single precision: 1344.9 Gflops.
Double accuracy: 168.1 Gflops.
(There is a more trimmed version of the GTX470, now you can find it for less than $ 250, 448 CUDA cores, 1.25 Gb)
Quadro 5000
One of the best video cards for professional applications.
Cost: according to Amazon about $ 1700. Released.
The number of CUDA cores is 352.
Memory capacity 2.5 Gb.
Float performance:
Single precision: 718.08 Gflops.
Double accuracy: 359.04 Gflops.
(It is worth paying attention to the Quadro 6000, 448 cores, 515 Gflops double precision, $ 4000)
Tesla C2075
Cost: according to Amazon about $ 2,200. Also released .
The number of CUDA cores is 352.
Memory capacity 6 Gb.
Float performance:
Single accuracy: 1030 Gflops.
Double accuracy: 515 Gflops.
What do we see?
Note that float performance GeForce GTX480 wins. The reason for this is the largest number of working cores and the highest frequencies among analogs. It is necessary to convert the coordinates of objects in games, calculate shadows, calculate pixel and vertex shaders. In the end - to the game "flew."
But for Tesla and Quadra to buy for scientific research, modeling the dynamics of liquids and gases, in double precision the performance is severely curtailed and inferior to counterparts.
Performance Ratio:
GeForce : double / float - 1/8
Quadro and Tesla : double / float - 1/2
In addition, the smallest memory has the same GTX480. It is enough for games, but if you want to calculate aerodynamics - buy something more serious.
***
WHAT DO I NEED?
(People involved in 3d graphics)
1. Smaller brakes when editing a 3d model.
2. Some are interested in the possibility of fast rendering on the GPU.
3D performance GeForce vs Quadro
From the information above, it may seem that professional applications on the GeForce are not used due to the fact that it has a small amount of memory, but it is not.
The video will show you why the “bad Quadra” is better than the “good Zhefors” in professional applications.
Quadro 600: 1Gb, 96 CUDA cores, 150u.e.
GTX560Ti: 1Gb, 384 CUDA cores, 250u.e. (Pricing from Amazon)
It turns out that Nvidia carefully ensures that 3d productivity in professional Geforce applications yields to Quadro at comparable prices.
How can viewport brakes be implemented?
The fact is that the number of polygons in games is significantly less than that of professionals in professional applications. In games, it rarely reaches one million polygons, and in professional games - tens of millions.
Here you can do this: cut performance when converting vertex coordinates. If there are more vertices than a certain number, then put a delay before the next vertices are drawn.
Or set a delay when drawing triangles. If more than a certain amount - then put a delay before drawing each subsequent triangle.
A little lyrical digression, or Nitrous in 3ds Max .
I was misled by the Nitrous engine in 3ds Max, which stands next to OpenGL and DirectX. What is it like? In Autodesk, there is something that causes Nitros, whose hardware support, it turns out, is on every self-respecting video card, but only 3D Max knows about it?
Well, you can make a small logical chain. Autodesk is a rich corporation, and in good partnerships with manufacturers of ATI and Nvidia. Raise the need to sell your brainchild! How to interest consumers? Performance is the same!
So, GeForce GTX580 (yes, I bought it), 7.3 million triangles, 2560 Torus Knot, without shadows and without Adaptive degradation.
Nitrous - 42 fps; Direct3d - 13 fps; OpenGL - 2 fps.
OpenGL - slow. DirectX is much better. And Nitrous - the coolest of all, it turns out! What then nitros then?
Two options:
1. This is OpenGL / DX in which additional brakes have been removed in the viewport, created intentionally in OpenGL / DX modes.
2. This is OpenGL / DX, which is able to access the hardware functions of gaming video cards, and manifest quadra abilities in them!
And I am inclined to the 2 variant, since in Blender and in Rhino3D, this very wildly slows down (2fps).
It turns out that users of 3ds Max and other Autodesk products are not at all so fundamentally switching to Quadra? Unfortunately, I do not have Quadra to check the performance of Nitros compared to OpenGL.
If you have a GeForce or Radeon, there is no desire to fork out for Quadra, you do not use products from Autodesk, and you have very complex models, then:
1. Complex objects can be hidden. Objects can be shown in a viewport with a lower grid density.
2. Instead of objects, you can show "containers" containing them.
That is, follow the number of polygons in the viewport, if you have really "heavy" models.
But in the game you play normally.
GPU rendering
Since commercial producers do not talk about what data types (float or double) they use - you have only to guess.
iRay is shown everywhere with Quadro and Tesla, it may give the impression that iRay does not work with GeForce at all.
Picture from the office. nvidia site.
But no, it works, and how. It would seem, what could be better for non-graphic calculations than a Tesla video card, specially sharpened for non-graphic calculations?
(Taken from the post: "V-Ray and Iray. Comparison and Review" )
GeForce GTX580 is the fastest single-chip graphics card in iRay GPU rendering. And much cheaper than "serious" analogs of the same performance. And if you miss 1.5GB, there are GTX580s with 3GB of memory.
When using V-RayRT, Octane, Cycles, Arion, the GTX570 and 580 video cards are also the best to show themselves. Does this mean that all these renderers do not use double precision calculation for rendering?
In any case, if you want to render on the GPU, you can save well on the GeForce.
GTX680
But the corporation noticed that for calculations, the GTX580 began to take more and more often, the double performance of the GTX680 is inferior to the float not by 8 times, but by 24, which could not but affect some tests .
It is known that in Octane Render performance increased by 64%.
ATI Radeon vs FirePro
Similar to Nvidia, AMD has also divided the video card models. Radeon (analog GeForce), FirePro (analog Quadro), FireStream (analog Tesla). The performance of double-precision floating-point calculations is inferior to single by 4 times, in all ATI models. Interestingly, the performance of ATI's top-end gaming graphics cards ( Radeon HD 7970 , float - 3.79 Tflops, double - 947 Gflops) exceeds even the single-chip Tesla in double precision. It should be noted that the performance in flops is not always an indicator of the performance of iron in specific cases.
The reason why ATI is much inferior to Nvidia in the GPGPU market is not yet clear to me. Maybe the gaming segment is enough.
Choice?
I chose the GTX580 3Gb. The video card gives you the opportunity to enjoy new games and GPU render performance. And the brakes in the viewport of 3d modeling packages are not very critical for me.
The author of the article with respect to this manufacturer, and he is the happy owner of the card Nvidia.
Such marketing moves are an integral part of a market economy, they are resorted to by all manufacturers without exception.
But nevertheless, we will not be conducted on marketing tricks of corporations, and thoughtfully buy what is really useful for us!
UPD: thanks for ForhaxeD , Funcraft , podwhitehawk for found spelling and grammar mistakes.
Source: https://habr.com/ru/post/142369/
All Articles