by Alexey Stepin , Anton Shilov
06/23/2006 | 11:18 AM
Besides color, shape and smell each object in this world has a dozen of physics characteristics such as weight, density, elasticity and others that affect their behavior when these objects interact with other objects or people.
<%BANNER[article]%>The rubber ball jumps off the floor, the water fills the gaps, the bomb ruins the wall, the fragments of broken glass scatter all over – there are millions of examples like that in our everyday life. However until recently nothing like that could be seen in computer games that were developed to model physical world – real or fictitious. The player could unload his entire reserve into the wooden barrel or thin building wall, but the barrel will remain their as solid as before and the wall will in the best case have a few tracks left by the bullets emulated with textures.
The game developers have lately achieved quite impressive results with the virtual words visualization, since the graphics cards have become much more powerful. It has become possible to demonstrate extremely complex scenes without fearing that the game would turn into a slide show. This has certainly added more realism to the games of today compared with what we saw 5 years ago, for instance. However, the realistic feeling you get from the game builds up from multiple factors. One of them is the interactivity of the game world, and it is certainly not less important than photorealistic graphics. The interactive character of the world in the game has always been a very tough nut to crack.
The thing is that accurate emulation of the objects’ behavior using real-life physics requires a lot of computational power. If it possible to implement the behavior emulation for a relatively small number of objects, which has been done in Half-Life 2 in particular, then the emulation of real physics for all objects of the scene can easily stall even the most powerful CPU. Yes, today it is the CPU that bears the complete workload of the physics models calculations in the games. When dual-core processors appeared it looked like it was possible to assign one of the cores to this complex task, while the other core would continue working on the game A.I., pathfinding, etc.
This logical solution is though just a semi-measure: the specifics of modeling complex objects and particle systems for such phenomena as smoke, water, etc. requires a lot of parallel calculations, otherwise the performance may turn out unsatisfactory. In other words, we need a computational device with parallel architecture that can quickly process a number of complex calculations if we want to be able to implement a complex realistic physical model in the today’s games. I have to stress that contemporary GPUs or CPUs like Cell, for instance, have exactly the architecture like that. Moreover, ATI Technologies and Nvidia have already announced the ability of their Radeon X1000 and GeForce 7 graphics cards to work as physics coprocessors.
A small and relatively young company , AGEIA Technologies, founded in 2002 has been long working on a special PhysX co-processor designed to process calculations of the physical models in contemporary games. The first solutions based on this chip have recently hit the streets. We are going to take a closer look at this new solution and ASUS PhysX P1 card will help us here.
So what does this physics accelerator from AGEIA Technologies look like?
The Gaming Power Triangle concept suggested by AGEIA Technologies implies that the gaming system should consist of three major parts in order to ensure that we achieve maximum realism. Each of these components is responsible for processing a certain part of the game.
The CPU works with the gaming process and artificial intelligence calculations. The graphics processor renders and displays the gaming scene and the PPU (Physics Processing Unit) has to calculate the physics model of the gaming world. In other words, the PPU in this concept is responsible for the movements and interaction of all objects in the game starting with the models of the players and monsters and finishing with the liquids and scattered pieces of broken or destroyed objects. This task requires huge computational power to, but will the solution from AGEIA be powerful enough to ensure this potential?
Unfortunately, we do not have any detailed information about the PhysX chip architecture that is the heart of AGEIA physics accelerator, we can just share with you the major technical specifications of the new PPU (Physics Processing Unit).
AGEIA PhysX physics processor is manufactured by TSMC using 0.13micron technology and consists of 125 million transistors. It is a relatively small number of transistors compared with contemporary GPUs (ATI R580, for instance, features close to 400 million transistors), but is quite comparable with the number of transistors insingle-core CPUs. Although in the latter case the majority of transistors are used for L2 cache, while the majority of PhysX transistors build the computational cores, and unfortunately we don’t know how many of those there are inside a single PhysX processor. The developer claims there are dozens of them. Since these are not very complex FP32 math1ematical units, there can be quite a lot of them actually, up to 20 or even more.
The clock frequency of the AGEIA chip is unknown. Since there are a lot of computational cores and the processor is manufactured with far not the today’s finest production technology, the working frequency may not be too high. As for the performance of the PhysX processor, we know it: the chip can execute up to 20 billion instructions per second. If we take the data from AGEIA for granted, this should be enough to calculate:
Note that the number of object collisions will hardly ever get close to the second number in real games, so I don’t think we should be concerned about insufficient performance level of AGEIA PhysX processor.
The processor is equipped with the GDDR3 controller that communicates with the memory via the 128-bit bus. The memory frequency is 366 (733) MHz and the peak memory bus bandwidth equals 11.7GB/s. While top-of-the-line contemporary graphics cards boast over 50GB/s memory bus bandwidth, this number is relatively small, but AGEIA PhysX is very unlikely to suffer from insufficient memory bus bandwidth. Firstly, the PPU doesn’t need to transfer textures (which eats up a lot of memory). And secondly, AGEIA physics accelerator supports regular 32bit PCI interface with 133MB/s bandwidth, so it will turn into a real bottleneck much sooner than the local memory on the accelerator card.
The fact that they used PCI interface indicates that the physics accelerator doesn’t have to shift a lot of data back and forth to the other components of the Gaming Power Triangle. AGEIA, however, claims that the PCI bus bandwidth may become a bottleneck in some games. Here it is important to mention that the early AGEIA PhysX samples supported PCI Express x1, although the final modifications of this accelerator didn’t have it any more for some reason.
Before we find out how AGEIA PhysX actually works, let’s make sure that we understand what type of tasks this solution is designed for. There are two types of physics in today’s computer games:
The physics accelerator deals only with the visual effects. Until the game developers decide to increase the physical interactivity of the game on their end or add up some innovative features such as gravity gun in Half-Life 2, for instance, any physics accelerator will only be able to slightly improve the visual effects.
PhysX physics processor can handle objects of three types:
Unlike the existing software physics implementation in the today’s games, all objects calculated by PhysX PPU can interact with one another and the surrounding environment. Moreover, there are different sets of parameters for different types of objects, including weight, density, friction, etc. According to AGEIA Technologies, all this should help game developers to achieve unbelievable gaming realism in their products.
Of course, no technology, especially the recently announced one, is absolutely flawless. Therefore, I suggest that we devote some time to discuss the possible weak spots of the current implementation of the Gaming Power Triangle concept. Looks like there could be at least three potential bottlenecks that may limit the performance of the system equipped with a new PPU:
Despite all its great features the physics accelerator is absolutely useless for the end-user if the games do not support it. What does the situation look like so far for AGEIA?
At this time there are 6 games that support the new AGEIA physics PPU: Tom Clancy’s Ghost Recon Advanced Warfighter, Rise of Nations: Rise of Legends, Bet on Soldier: Blood Sport, Cell Factor, Gunship Apocalypse and City of Villains. In the nearest future this list should grow up to 20 titles including a very impatiently awaited Unreal Tournament 2007. As a result, the longer list of games will include:
This is a relatively modest list compared with the overall number of contemporary games out there. But if AGEIA manages to win the developers’ hearts with these few titles, then this list will inevitably grow bigger very soon. The company does everything they can to achieve this goal: they provide the corresponding API/SDR – the physical engine that used to be known as NovodeX. It features native PhysX support and is intended for developing games with advanced physical model for PC platforms as well as for PlayStation 3 and Xbox 360 gaming consoles.
In any case, the idea of an individual processor allocated to calculate gaming physics will live, evolve and spread independent of the hardware implementation in the form of a special chip or an extra add-on card performing the PPU functions in the system. The reasons for our optimism are very clear: it is impossible to get to the next stage of gaming realism without improving the behavior of the game objects. Moreover, special devices cam usually offer the developers more power than the general-purpose computational devices such as system CPUs.
According to the unofficial data we have at our disposal, AGEIA PhysX processes physics 200 times faster than a contemporary system CPU. However, the computational monster, ATI Radeon X1900 XTX, boasts even higher computational potential, although just in theory.
But let’s take a closer look at the actual living being from AGEIA. Please welcome ASUS PhysX P1.
AGEIA PhysX physics accelerator is a relatively small card equipped with a small cooler. At first glance it looks very similar to a graphics card, because you do not notice that there are no D-Sub/DVD connectors right away.
Looks like the PCB design follows closely one of AGEIA’s engineering samples, because there are a few technological connectors close to the fastening bracket that haven’t been laid out. The voltage regulator circuitry of AGEIA PhysX is quite complex and includes a lot of capacitors, even though the card hardly consumes more than 20-25W of power. The heart of this circuitry is two Intersil ISL6522CBZ PWM-controllers that control the work of AC-DC converters. Since PCI bus cannot provide enough power, the card is equipped with a standard four-pin Molex connector.
Since PhysX processor uses PCI 3.0 interface with 3.3V signals, it is tied up to this connector via three Texas Instruments TW222A chips. These elements consist of 23 NMOS transistors each and serve to limit the input-output signal level. This way, they protect the PhysX chip against damage if the card is installed into a 5V PCI slot.
Having removed the cooler we got access to the PPU chip and could read its marking.
The chip is marked as AG10011-P revision A1 and is manufactured on the 40th week of last year. In other words in the end of September – beginning of October 2005 AGEIA already had second revision of working PhysX chips at their disposal (the first revision was A0). Looks like it took them so long to get the cards into the market because they were working on the software all this time.
The die of the PhysX chip manufactured with 0.13micron technology is quite big and makes around 190sq.mm. As we have already said before, we do not know what frequency it works at.
The card is equipped with four GDDR3 Samsung K4J52324QC-BC20 chips in 136-pin FPGA packages. Despite the fact that these are 512Mbit chips, the total amount of onboard memory available for the PPU is 128MB. Maybe half of the memory banks are simply not used at all. The nominal frequency for the BD20 chips is 500 (1000) MHz, however the memory of PhysX P1 solution works at 366 (733) MHz. keeping in mind the low heat dissipation of the new 136-pin packages and lower working frequency they do not need any extra cooling.
PPU core is equipped with a cooler, which you should have already read about in our articles called NVIDIA Multi-GPU SLI Technology: New Approach to Old Ideas and ASUS N6600 GT/TD Graphics Card Review. The only difference between the PhysX P1 cooler from the other two is the color: in our case it is copper-colored. The cooler location is not the optimal one, because the PPU die appears right under the fan motor, which is a practically dead zone. For the most optimal cooling effect it should have been placed beneath the ribbed heatsink cooler by the air stream from the fan. However, since PhysX works at not very high clock speed, it shouldn’t be a big problem and the cooler should perform its functions well enough.
To test ASUS PhysX P1 we assembled a standard test platform configured as follows:
We decided to go with Radeon X1900 XTX, because this graphics card boasts the today’s most advanced architecture among solutions with SM3.0 support. Besides, it performs at very high speed typical of contemporary high-end graphics solutions. We decided not to use any CrossFire or SLI configurations, because multi-GPU solutions are not that widely spread yet among the majority of end-users.
We used the following games and demos for our tests:
Testing physics accelerators is not a trivial task. The AGEIA card itself cannot speed up the already existing content. It simply allows the developers to use new special effects in the games and significantly increase the number of objects with physical features compared with the game versions without the PhysX support. As a result, it doesn’t make much sense to compare the performance of the system with the PhysX accelerator against that of the system without it: the average performance in the second case may turn out even high while the realism and image quality overall would be lower.
Nevertheless, we performed this type of testing as well. We used Tom Clancy’s Ghost Recon Advanced Warfighter (GRAW), since this game knows to work with PhysX card as well as without it, and a demo included with AGEIA drivers. The latter is a scene with a flying sphere that breaks the wall built of separate blocks. It can also be launched without the PhysX PPU in the system. These applications were tested in a usual manner with the help of FRAPS utility to record the average and minimal fps rates.
CellFactor and Hangar of Doom could run only with AGEIA accelerator installed. To make the best out of it we tested our systems in a different way in these applications: the results showed a performance curve throughout a certain period of time.
We have also taken a number of screenshots showing the quality of effects provided by PhysX solution. In case of GRAW game we could also compare the image quality with and without the PPU.
Hereinafter the screenshots on the left are obtained from a system without the PhysX accelerator, while the ones on the right – from a system with the AGEIA PhysX PPU.
Standard mode | AGEIA PhysX PPU |
Without the physics accelerator the wooden chips the bullets break off the trunk do not abide by the laws of gravity and fly along the preset trajectories. With PhysX solution there are more chips and they behave in a more realistic way. Unfortunately, it is hard to illustrate what I am talking about with a static screenshot. Also, you can see that the chips generated by the PPU are of identical shape, which takes away some of the realistic feel. However, in motion you will not notice it as clearly as on the static screenshot.
Standard mode | AGEIA PhysX PPU |
The second scene definitely looks a way better with the AGEIA card. The sheaf of sparks hit off the car door is quite impressive, especially in dynamics: all falling sparks that later jump off the road have individual trajectories and look absolutely real. Besides, there are much more of them than in case there is no PhysX processor in the system. Although you can shoot the car as long as you want: the paint will remain brand new and solid, the surface will not get deformed in any way and if you heat the gas tank, no explosion will follow.
Standard mode | AGEIA PhysX PPU |
The third scene also shows a few very evident differences. With PhysX processor the turf chunks torn out of the ground with the fire burst behave exactly the same way they would in real life. Moreover, the cloud of dust feature more diverse structure and looks more realistic than in case there is mo PPU support.
Standard mode | AGEIA PhysX PPU |
The car explosion from the thrown grenade also looks much better with the PhysX in the system: there are more fragments thrown around and their behavior is much more natural. Besides, the realistic smoke contributes to the great picture.
Standard mode | AGEIA PhysX PPU |
The same is true for blowing up a pile of boxes in the fifth scene. Note that a lot of objects on the screenshot obtained from the system with AGEIA PhysX have a kind of a shadow attached to them that looks very much like motion blur effect.
Standard mode | AGEIA PhysX PPU |
And the sixth scene looks very similar in both cases: with and without the AGEIA PhysX card. The only difference is the bigger number of smaller fragments generated by the PPU.
Standard mode | AGEIA PhysX PPU |
The fence in the seventh scene gets completely destroyed in both cases, but only in the system with PhysX you can see a realistic cloud of smoke from the explosion and a lot of panel fragments on the road. Unfortunately, they all disappear after a while, which reduces the realistic feel of the game.
All in all, Ghost Recon Advanced Warfighter looks much better with AGEIA PhysX accelerator installed, however we cannot claim that this is the new level of gaming realism. Moreover, the absence of interactive physical effects makes the additional elements look pretty weak.
Besides, the game that has been released not only for the PC platforms but also for the PS2 and Xbox360 consoles is far from perfection from the graphics standpoint. Namely, there are no high-resolution textures and the overall level of detail leaves much to be desired. In other words, it hardly makes any sense to spend extra $299 for a physics accelerator to play Ghost Recon Advanced Warfighter, even if you are a dedicated fan of Tom Clancy’s games.
However, let’s see how we can benefit from AGEIA PhysX in this game from the performance prospective. As you should remember, it doesn’t support full-screen anti-aliasing because of the deferred rendering.
We see a slight increase of the average and minimal performance: although the physical effects in GRAW are much more complex, it is the specially dedicated PPU and not the system processor that deal with them. There is no significant performance booth, however, and the 1600x1200 resolution is still not playable in both cases, although AGEIA PhysX allows to slightly increase the power reserves, especially in 1024x768+ resolutions.
The screenshots above show the effects quality provided by AGEIA PhysX in CellFactor demo. We didn’t have any critical comments about the shattering and breaking (but for some reason not deforming) pipes, boxes and barrels as well as the smoke. However, when it came to liquid effects, the results turned out quite far from ideal.
Although the liquid streaming from a broken barrel and the blood of the main character behave according to the physical laws, they look nothing like real blood or real liquid. It is especially true for blood that looks more like some sticky mess consisting of dark-red bubbles. It is not the issue with AGEIA accelerator, because it depends on the game developers’ methods, so the final version of CellFactor game may have much better liquids implementation. At the same time, the heat haze effect created with fluid objects and pixel shaders look very realistic.
Other than that, even this single-level of the CellFactor demo looks quite impressive and shows very well the advantages of physics accelerators. You cannot get anything like that having only the computational power of your CPU: there are too many objects with physical features to be calculated. Moreover, heavy objects like containers and machinery have different physics than lightweight pipes and boxes.
Since we couldn’t analyze the performance difference in CellFactor with and without the PhysX processor, we decided to offer you a time diagram. It shows how greatly the gaming performance may drop because of massive effects.
Having obtained the performance results for each given moment of time we managed to calculate the average performance level. As you see, if there is a single Radeon X1900 XTX graphics card in the system, the overage performance will be quite low, less than 40fps. In fact, we expected something like that with the massive special effects enabled.
The momentary performance diagrams are not that smooth, which indicates that the performance in CellFactor may vary within a very wide range. If there is an effect involving a lot of objects, you may notice a significant performance drop at this time. It is especially typical for liquid effects. It could be the performance of the graphics card pixel processors that affects the results so badly. I believe that appropriate optimization of the gaming engine or graphics card driver could help in this case.
In common situations, when the player simply runs over the level without making much mess, the performance is well above 30-35fps, especially in 1024x768. CellFactor is a new generation game, so you shouldn’t really expect the today’s hardware to ensure superior performance in this game.
This demo is included on the same CD disk with AGEIA drivers and is a mini-game where you have to shoot down the enemy’s airplanes from a permanently placed antiaircraft gun. In the best case you will be rewarded with a walk around the barrack shooting the boxes, barrels and air-bombs that will explode when shot and throw all other objects around. Looks like Hangar of Doom is based on a modified Unreal engine. The demo runs in 1024x768 and the resolution cannot be changed.
The main task of Hangar of Doom is to demonstrate the ability of AGEIA PhysX to calculate the behavior of multiple objects. And this demo does it fairly well, however all boxes, barrels and bombs act as if they were all of the same weight, which takes away some of the realism in the scene.
We can also see quite noticeable performance fluctuations from 10fps to 40fps, while the average performance level will be only 20fps with maximum quality settings. The major performance drops occur, when the bombs explode and the fragments of surrounding boxes and barrels get thrown away by the shock wave.
The average performance gain resulting from the PhysX PPU is relatively small, less than 10%, however the minimal performance increases by well over 40% in this case. It could result from the parallel architecture of the PhysX chip that allows processing the parameters of several objects at the same time without losing or losing very little of its performance.
Since there were almost no games supporting physics co-processor when we worked on this article, we believe it would be fair to call it AGEIA PhysX technology preview. It would be a mistake to make any final conclusions about the efficiency of the new technology judging only by the results of one single game and a few demos. However, there are a few things that we feel quite confident saying.
First of all, it is very clear that AGEIA’s success depends solely on the game developers’ support. If the support is relatively weak or there is none, this theoretically promising solution may follow in the footsteps of Nvidia NV1. I allowed myself this comparison because, in fact, there is no unified standard for creating physical effects. There are a few engines that compete for the prestigious title of the standard setter: AGEIA, Havok 4, Havok FX and later there will be another one – DirectX Physics from Microsoft.
The latter two allow using graphics processors supporting FP32 to calculate the physics model. Since they are very widely spread in the today’s market already, they have better chances to become more universal solutions than AGEIA engine that requires special hardware, which is not so widely available in the market yet.
Note that multi-core processors are also getting more and more popular, so the game developers may go with Havok 4 engine for physical effects, because this engine is optimized for multi-core processors, including Cell of PlayStation 3.
AGEIA needs at least one really well-known and popular game to popularize its approach to realistic gaming physics. Right now there is no game like that, and Tom Clancy’s Ghost Recon Advanced Warfighter is definitely not the one. Although there is one candidate we all are waiting for: the multi-player Unreal Tournament 2007 shooter.
The effects provided by PhysX processor are not very diverse yet, and in Ghost Recon Advanced Warfighter they are not even realistic enough. There are a lot of practically identical fragments that disappear after a while, which is far from what the real life situations would look like. Moreover, it even gives you that strange artificial feel from the action in the game. The CellFactor demo looks a much better from this prospective, however there is also a pretty big fly in the ointment: liquid effects are not just far from real, but also cause significant performance drops in corresponding scenes. Of course, the new PhysX API versions will acquire new effects. In particular, CellFactor already has support for cloth effects. The use of “cloth” type of objects will allow game developers to create naturally dressed characters and much more other things.
Let’s sum up a few things. The AGEIA PhysX physics co-processor does have the claimed features and can really increase the realism of the gaming environment.
However, since there are almost no games supporting PhysX out there, it doesn’t make much sense to invest money into a card based on this new processor, such as ASUS PhysX P1. For extra $299 you will get a chance to play Tom Clancy’s Ghost Recon Advanced Warfighter with a slightly higher level of realism in the game. The performance improvement in GRAW is not worth this money from our standpoint.
When new games with PPU support come out, the situation may change, however I wouldn’t recommend buying PhysX based cards in advance, because the current modifications use the PCI bus. New generation games using a lot of PPU resources may stumble over the insufficient bandwidth of the PCI interface in the first-generation PhysX cards. By the time there are enough games using AGEIA’s physics engine, they will most likely have a new revision of the physics accelerator using faster PCI Express bus. So, hold your horses and wait a little bit longer.
