Unreal physics. NVIDIA PHYSX testing on SLI Multi-Card configuration

  • Post author:
  • Post category:Blog
  • Post comments:0 Comments

As you know, NVIDIA graphics cards are able to take part of the duties of the central processor: to transcode the video, accelerate mathematical calculations and, of course, work with physical engines - with the same Physx. Such work takes the lion's share of the performance of the board, and the iron enthusiasts have been trying to increase the computer’s power and save money for several years, deceiving NVIDIA SLI and connecting two video cards of different series in one system. They learned to do this for a long time: computers with two different boards gathered back in mid-2008, but for this it was necessary to make a lot of effort. Get two monitors, or one, but with a pair of entrances - for example, DVI and VGA. After installing the first video card, turn off the computer, insert a second video card, connect another monitor to it, remove all drivers from the safe Windows mode with the Driver SWEEPER program and then install the NVIDIA Forceware driver again.

In October 2009, technology appeared NVIDIA Multi-Card , which allowed it to easily collect ligaments from two Nvidia video cards of different families, say, from GeForce GTX 295 And GeForce 9600 GT. In the last issue, we conducted the first tests of this technology, and today we are ready to tell in thoroughly which cards are better to put on Physx calculation.

With what is

Physics in the game is needed so that you believe in the reality of what is happening. In games she appeared very, very long ago. Take the same Pong Superb Casino : If the virtual ball did not bounce when hit, as the real ball does, the developers did not convince us that this is a ping-pong. Almost forty years have passed since then, but nothing has changed, only the scale has increased. Now almost any game is a huge world with thousands of objects, and each of them must behave reliably. Unfortunately, it is not so easy to arrange.

If you shoot from a shotgun in a box, then it will not only fly into the air, but also break into shreds. And when in the game she only grows holes from the bullets, without moving, it looks narly. However, this is what happens, because a three -dimensional model is just a set of landfills that form the desired shape. In order for the game to distinguish a cardboard box from the reinforced concrete block, you need to explain to the computer what gravity, blow, acceleration, friction, indicate the weight of the box and the material from which it is made.

Those who have gone beyond the school physics course or studied the sopromat and theoretical mechanics, know that even to calculate with what speed the wheel will roll from the mountain is incredibly difficult. And to reliably describe how the box will respond to the fraction that falls into it - that is still a task. For six months, students of technical universities expect how some spring reacts to compression-and imagine how many such wheels, boxes, springs in the game like Burnout Paradise. Yes, even if the entire composition Electronic Arts Plant all this to program, then their next game we will see in ten years! Fortunately, there are physical engines in the world.

Replacing reality

In fact, any engine, whether it is Havok or Physx, is a database with a description of materials and their properties. In the case of the same box, for example, the developer is enough to say the engine: “It is cardboard and weighs 200 grams,” and the engine will automatically select the necessary formulas and make it clear to the game how to respond to certain manipulations with a box. And it works fine. A striking example - Half-Life 2 with his engravipushka, which allows you to scatter the boxes or rearrange the barrels.

But boxes and barrels are, unfortunately, only a small part of the real world. If you look at the same Half-Life 2, you will notice that all objects in the game are absolutely solid bodies. They know how to jump, tumble in the air, bounce off the walls, but there even the easiest object when hit does not change its shape. The reason is insufficient performance of modern computers. Write formulas for working with water, fabric, rubber or fog is not difficult, there are a huge amount of them, but there is nothing to calculate such formulas yet.

Take the fabric for example. Just think how many points in a skirt of an ordinary girl. Thousands! And in order for the hem to flutter correctly in the wind, it is necessary to calculate the position of each point every second, take into account the behavior of all neighboring ones, the strength of the wind, the direction of movement of the character. This is not to mention the fact that the wind creates many small flows that also affect fabric behavior. And imagine that your Core i7, which puffs with Windows, settles problems with drivers, monitors AI of characters, suddenly gets such a skirt. Even the most modern processors are not ready for such loads.

Tests

"Anabiosis: Dream of reason"

It is difficult to call the game herself a masterpiece, but as a benchmark for Physx, it fits one hundred percent. There is everything that was developed by Physx SDK: pouring water, games with fabric, jumping boxes. At the same time, Anabiosis is extremely demanding on the schedule. Beautiful models, complex lighting - all this creates a serious load on the video card.

Batman: Arkham asylum

Actually, there are few interesting ideas with physics in Batman, but the objects in the game represented a variety of. The raincoat of the main character flutters on the go, enemies beautifully scatter hundreds of paper leaves. And we are also shown the work of Physx with particles: a fog is very beautifully made here, which is in a dense layer on the ground.

Physx Fluidmark

This benchmark was developed by Nvidia itself. There is nothing breathtaking in it: in the middle of the screen there is an ordinary ball, on which tens of thousands of particles form on top forming a viscous mass. Fluid does not load the video card with graphics and squeezes all the juices to the miscalculation of particles - using this test it is easy to evaluate the overall performance of each board.

Bright thought

The company was the first to understand this Ageia : In the mid-2000s, she began to make a physical engine Physx SDK , who taught to consider the behavior of solids, fabric, liquid - and process everything with hardware. Ageia has developed the Physx fee on Ageia, which was supposed to help the central processor. However, Ageia did not have enough money or effort to bring the revolutionary project to the end and get the necessary support from game developers. In February 2008, the company bought NVIDIA.

The basis of the PhySX board is a general -purpose nucleus, approximately the same as on all NVIDIA graphics cards. Given that in 2008 it was already advantageous Cuda , It was not difficult to integrate Physx support in drivers: all Nvidia video cards after GeForce 8000 know how to work with the Physx engine. But here there are shortcomings.

When Physx is turned on, we really unload the central processor, but at the same time we select the resources from the video card, and after all, it is also a graphics in the game to consider. It turns out a stick about the two -edges. Yes, video cards consider physics faster, they are imprisoned for this, but the overall performance of the system is still reduced. And not everyone wants to lose additional frames per second for some drawers. NVIDIA from the very beginning offered a way out: buy, they say, the second video card and give it to a torn physics. But the conditions were painful. In SLI it was possible to put only the identical first fee, and if you already had a GTX 285, then for the second such a card you had to give another 12,000 rubles, which was too expensive for boxes and skirts.

Time of change

NVIDIA Multi-Card technology is, conditionally, a variety of SLI. With its help, you can combine the GTX 285 with the old GeForce 8800 GT or GTX 275 C GeForce 9500 GT, without using additional bridges or expensive maternal boards. The procedure is simple: set a more powerful fee (GTX 285, GTX 275 and T.D.), reinstall the drivers, turn off the computer, insert the second video card. Windows itself will pick up another fee and pick up drivers under it. You only need to enter Nvidia Control Panel and choose which of the boards will be responsible for Physx.

For tests, we took five boards at once MSI : GeForce 9500 GT And 9600 GT , GeForce GT 240 , GT 220 And 210. All these cards cost less than 3,500 rubles today (which can be paid for reliable physics), and we were interested in how much they would pull Physx. In order not to rest against the graphics, in the bunch to them we installed the GeForce GTX 285. Naturally, we also checked her in work without hardware acceleration and with Physx.

The computer for tests was assembled powerful. Motherboard - MSI X58 Platinum , processor - Intel Core i7 920 , Memory - three strips Kingston Hyperx DDR3-1600 2 GB. They put it on top Windows Vista 64-bit and downloaded the last drivers Forceware 195.62. To evaluate performance, we used a couple of games - "" Anabiosis: Dream of reason " And Batman: Arkham asylum - and special benchmark Nvidia Physx Fluidmark. In games, we twisted to the maximum not only physics, but also the overall quality of the graphics.

One is good, but two are better

The central processor lost the calculations of Physx SDK - it became clear right away. Even our powerful Core i7 920 could not draw out any of the games for an acceptable level of performance - look at the tables, everywhere no more than fifteen frames. After the PhySX switch to the GeForce GTX 285, the situation has changed dramatically - the games went at decent speed. But compared with tests with Physx included without the support of the second graphics card, it was still notable: without Physx on the GTX 285, 144 frames ran, with Physx - already 58.

We began to select the perfect fee for physics, driven quite a lot of tests and found out that not every video card is able to work with Physx: too large calculation volumes. Moreover, the inappropriate fee can even become a narrow place of the entire system. Compare the results! The first video card, GeForce 210, lost my responsibilities and slowed down: compared to the test with one GTX 285, we lost 27 frames per second in Batman. But starting with GeForce 9500 GT, the indicators crawled up. Despite the fact that the absolute performance of PhySX compared to the single GTX 285 is three times lower (see. result Fluidmark), in games, such a ligament adds 10 frames per second. The same can be said about GeForce 9600 GT - an increase of almost 40%. But the cards of the new series, GT 220 and GT 240, behaved ambiguously. In the game “Anabiosis: Dream of Reason” the results were at the level of a single GTX 285, and in Batman: Arkham asylum, performance increased by 37%.

Table 1
Technical characteristics
MSI N285GTX SUPERPIPE OC MSI N9500GT-MD512/D2 MSI N9600GT Zilent 1G
The core GT200B G96 G94
The number of transistors 1.4 billion 314 million 505 million
Process process 55-nm 55-nm 65-nm
The number of stream processors 240 32 64
The frequency of the graphic nucleus 680 MHz 550 MHz 650 MHz
Frequency of stream processors 1476 MHz 1375 MHz 1625 MHz
Type, memory volume GDDR3, 2 GB GDDR2, 512 MB GDDR3, 1 GB
Memory frequency 2500 MHz 800 MHz 1998 MHz
Data tire 512 bits 128 bits 256 bits
The number of texture blocks 80 16 32
The number of rasterization blocks 32 8 16
Interface PCIE X16 2.0 PCIE X16 2.0 PCIE X16 2.0
Price for January 2009 13,500 rubles 1900 rubles 2800 rubles
Technical characteristics
MSI N210-MD512H/D3 MSI N220GT-MD512 MSI N240GT-MD1G
The core G218 G216 G215
The number of transistors 260 million 486 million 727 million
Process process 40-nm 40-nm 40-nm
The number of stream processors 16 48 96
The frequency of the graphic nucleus 589 MHz 625 MHz 550 MHz
Frequency of stream processors 1400 MHz 1360 MHz 1340 MHz
Type, memory volume GDDR3, 512 MB GDDR2, 512 MB GDDR3, 1 GB
Memory frequency 1580 MHz 810 MHz 1580 MHz
Data tire 64 bits 128 bits 128 bits
The number of texture blocks 8 16 32
The number of rasterization blocks 4 8 8
Interface PCIE X16 2.0 PCIE X16 2.0 PCIE X16 2.0
Price for January 2009 1500 rubles 2200 rubles 3500 rubles
Table 2
Comparative tests
Physx Fluidmark, 1680x1050 Physx Fluidmark, 1920x1080
GTX 285 + Core i7 920 23 23
GTX 285 214 249
GTX 285 without Physx - -
GTX 285 + GF 9500 GT 74 65
GTX 285 + GF 9600 GT 128 108
GTX 285 + GF 210 31 33
GTX 285 + GF GT 220 77 69
GTX 285 + GF GT 240 - -
Comparative tests
“Anabiosis: Dream of reason”, 1680x1050 “Anabiosis: Sleep of the mind”, 1920x1080 Batman: Arkham Asylum, 1680x1050 Batman: Arkham Asylum, 1920x1080
GTX 285 + Core i7 920 11.4 10.2 12 13
GTX 285 38.2 38.3 58 51
GTX 285 without Physx - - 144 128
GTX 285 + GF 9500 GT 48.9 42.9 55 53
GTX 285 + GF 9600 GT 54.8 45.6 77 73
GTX 285 + GF 210 35.5 32.2 31 31
GTX 285 + GF GT 220 46.7 38 69 69
GTX 285 + GF GT 240 47 42.9 80 76

Leave a Reply