Introduction or How Kyro II came into Being...
The birth of Kyro II baby is directly connected with PowerVR Company, so we think we should first say a few words about this company and the predecessors of Kyro and Kyro II.
Well, the beginning of PowerVR's history goes back to the year of 1996, when its first products, PowerVR Series 1, appeared. Together with NEC they launched NEC PX1 and NEC PX2 graphics accelerators. Five years ago there was no active 3D graphics card market, so the companies had different ideas about 3D graphics and looked for their own ways to implement them. NEC PX1 and NEC PX2 didn't gain public recognition and very soon sank into oblivion.
In February 1998 PowerVR announced Series2 products based on tile architecture, which became more successful and were welcome first of all in gaming consoles and machines. In 1998 SEGA chose the PowerVR Series2 architecture to be used in its Dreamcast gaming console, and a bit later the most advanced gaming machine of those days, NAOMI 1, was equipped with several chips from PowerVR Series2. These chips worked in parallel during the scenes rasterization, while geometric calculations were carried out by an additional geometric processor. In August 1999 PowerVR Series2 chips penetrated the PC graphics adapters market. These were NEON 250 graphics cards by VideoLogic Systems built on NEC PowerVR 250 chips, which were manufactured on the basis of PowerVR Series2.
These cards provided excellent image quality, but they suffered from grave problems with the drivers. Moreover, NEC's troubles with the chips manufacturing hampered the mass sales of NEON 250. Thus the tandem of PowerVR and NEC failed once more to enter the PC graphics card market and PowerVR turned to another manufacturer - STMicroelectronics.
The third generation - PowerVR Series3 - was announced in April 1999 and STMicroelectronics, a well-known semiconductor manufacturer, obtained PowerVR's license to use Series3 technologies in PC graphics chips and gaming consoles. In June 2000 STMicroelectronics announced its first chip based on PowerVR Series3 technologies called Kyro, and in September 2000 there were three manufacturers to announce new products built on PowerVR/STM Kyro - VideoLogic Systems, InnoVISION and PowerColor.
Kyro turned out a smart, pretty fast chip, which was unlucky to appear during a confrontation flare-up among NVIDIA, ATI and 3dfx. The price of Kyro based graphics cards made them feel utterly uncomfortable - they were more expensive then the Low-End cards offered by the competitors and slower than the latest high-performance gaming cards, which were pricing only a little bit higher. Again, tile architecture and Kyro itself were rejected by the gaming industry.
However, in March 2000 STMicroelectronics announced the next product based on PowerVR Series3, Kyro II, which was virtually a slight remake of the first Kyro enriched with a couple of modifications and made with a new 0.18micron technology (the migration from 0.25micron to 0.18micron technology allowed to bring up the core clocking from 125MHz to 175MHz). Taking into account that the performance of graphics cards based on tile architecture is strongly determined by the core frequency, this increase in core frequency was supposed to improve the performance and to make Kyro II based graphics cards more competitive.
VideoLogic Systems was the first to manufacture Kyro II based graphics cards and its pilot product was named VideoLogic Vivid!XS. VideoLogic Systems is a branch of Imagination Technologies Group. Alongside with VideoLogic Systems dealing with the production of graphics cards, sound cards and all sorts of multimedia products, Imagination Technologies Group also includes PowerVR Technologies (it develops various algorithms of image creating and processing), Metagence Technologies (this one works on broad-spectrum DSPs - digital signal processors) and Ensigma Technologies (this company develops and licenses to third companies algorithms for audio data processing).
The second company to pay attention to Kyro II became Hercules, recently bought by Guillemot concern. Hercules brand, which boasts excellent reputation in Europe, together with Guillemot's financial and marketing abilities as well as adequate pricing policy can ensure successful promotion of Hercules Kyro II based graphics cards - Hercules 3D Prophet 4500 64MB.
Hercules 3D Prophet 4500 64MB graphics card was announced on March 9, even earlier than Kyro II chip by STMicroelectronics. We were lucky to get hold of one of these cards, so here is the review.
Chip Specs and Tile Architecture Peculiarities
Key specs of Kyro II chip look as follows:
- Manufacturing technology - 0.18micron;
- Clock frequency - 175MHz;
- 2 pixel pipelines with one texturing unit each;
- Fillrate - 350Mtexels/sec;
- Tile architecture supporting "Deferred texturing" technology;
- It is capable of overlaying 8 textures per single pass;
- All the color calculations within a tile are performed in 32-bit color mode;
- Supports tri-linear and anisotropic filtering;
- Supports EMBM and Dot3 bump mapping methods;
- Supports texture compression;
- Supports full-screen anti-aliasing via 2x2, 2x1 and 1x2 supersampling;
- Supports up to 64MB 128-bit graphics SDRAM;
- Graphics memory is clocked at 175MHz. RAMDAC frequency is 300MHz;
- Supports AGP1x, 2x protocols with SideBand Addressing.
Well, the chip looks a bit too modest compared with the modern graphics cards features, but its image creating principles are absolutely different, so it doesn't fit conventional performance standards like the number of megahertz, pixel pipelines and texturing units.
Kyro II is built on the so-called tile architecture. The idea is that the image is built not as a whole, as it happens by common graphics accelerators, but in tiles, that is, the screen is split into fragments of 32x16 pixels in size, which are called tiles, where parts of a scene are built one after another. Kyro II builds a scene in consecutive order: first is builds an image within one fragment (tile), then it passes over to another one and so on, tile by tile, till the scene is completed.
An image within a single tile is built in the following way:
When Kyro II receives the whole list of polygons in the scene, it defines for each tile, which of the polygons covers it partially or completely and makes individual lists of such polygons for each particular tile. After that for all the tile pixels the smallest distances (Z coordinates) are saved and the numbers of triangles, to which these distances correspond, are defined. Z-coordinates are checked not for only one pixel per clock but for 32 pixels of a tile simultaneously (Z32 comparisons per clock), so this stage of rendering takes considerably less time. Only after all the tile polygons are built, all the closest pixels (those which are to be eventually textured) are singled out with the help of the tile Z-buffer and the numbers of corresponding polygons are defined, the actual texturing of these pixels begins. After that, a ready image, which makes part of the full-screen image, is transferred to the frame buffer stored in the graphics memory.
The most exciting thing to mark is that the small tile frame buffer of Kyro II, the tile Z-buffer, the buffer with the stored numbers of the textured polygons for each tile pixel and all sorts of buffers needed to create some image within a tile - all this stuff is integrated into the chip core as common caches, and almost all the image creating operations take place in the core. Most of the data, such as the Z-buffer data, for instance, which transfer between the core and the graphics memory would slow down any conventional graphics card, also circulates within the core of Kyro II.
Thanks to tile architecture the chip addresses the local graphics memory very rarely as long as more or less significant amounts of data are transferred via the memory bus only by tile pixel texturing when the core requests textures from the graphics memory or by transferring the ready image from the tile frame buffer to the general frame buffer. We tried to illustrate our explanation in the following way:
Moreover, the peculiarities of the chip architecture save it time and trouble texturing invisible surfaces, i.e. Kyro II simply doesn't waste time on that. As a result, there are fewer textures transferred to the graphics core, which makes the data transfer rate along the graphics memory bus not so vital any more.
Of course, tile architecture is no absolute perfection. It doesn't consist only of advantages. For example, before building a scene Kyro II should obtain the data about all its polygons while a conventional graphics accelerator textures the polygons one after another as soon as it receives them. Such a peculiarity makes Kyro II suffer performance losses in some games. This way, if the developers have designed their gaming engine for ordinary graphics accelerators, then on building a scene the engine may send a portion of polygons (say, describing some room) to the accelerator, because it "expects" that the accelerator will process and texture them at once. And in the meanwhile the processor is supposed to calculate the moving objects parameters or the gaming logic and when it is through it sends another portion of polygons (say, determining the models of the game characters in that room) to the accelerator. An ordinary accelerator will process them right away and complete the scene. But Kyro II will take a different way. It will wait until it gets the entire list of polygons that determine the gaming scene and only then it will start building the image.
Another shortcoming of tile based rendering, which looks much more like an inevitable historical factor, is that game developers fairly assume that hidden surfaces texturing is not the best way to improve the fps rate. So, they try hard to decrease Overdraw value as greatly as possible (just in case you forgot: Overdraw is a measure of how many times a pixel is drawn for each frame rendered). As a result, they remove the hidden surfaces on the software level already. It affects the performance when we speak of ordinary 3D graphics cards, but tile based accelerators lose the efficiency of their hardware hidden surface removal technology.
Another point that can negatively affect the performance of tile based graphics cards is the necessity to define a list of polygons that refers to each particular tile. We can't disregard this fact because the average number of polygons in a contemporary gaming scene is constantly growing.
Moreover, alongside with all the highs and lows of tile architecture, Kyro II chip itself has a number of both: pleasant and upsetting traits. We'll tackle them in the section titled "3D Image Quality". And now we suggest speaking a bit about Hercules 3D Prophet 4500 64MB graphics card, which is the key figure of the today's review.
Hercules 3D Prophet 4500: Closer Look
Hercules 3D Prophet 4500 graphics card is shipped in a retail package designed in Hercules's traditional way: with typical colors and design elements. On the front side of the box there is a "handsome" colored face of some Viking or an Indian. A nice guy, anyway:
By the way, it's no new design - we saw similar faces on the boxes with 3dfx's products. So, as long as 3dfx doesn't exist any longer, such box design can be viewed as a contrast to NVIDIA. Whatever the reason, we like a picture like that much better than various airplanes and other technical stuff.
The package includes Hercules 3D Prophet 4500 graphics card and a CD with drivers. The card's PCB is made of blue-green textolite and is rather huge, we should say:
Hercules 3D Prophet 4500 has a Kyro II chip and 64MB 128bit graphics memory (8 chips made by SAMSUNG):
The core and the graphics memory have synchronous clocking equal to 175MHz. The core is equipped with a brand Hercules cooler resembling a bit to the Blue Orb by Thermaltake.
It is our considered opinion that the cooler by Hercules is not so powerful as Blue Orb, though it produces notably less noise. In fact, we have noticed that the holes for cooler fastening are located too close to the core, making it impossible to install Blue Orb. But it won't be needed anyway: the Hercules cooler is quite enough to dissipate the emitted heat.
The PCB also has a special spot for a chip displaying the images in TV-format via S-Video or a composite Out and one more spot for another chip working as a TMDS transmitter to the DVI-Out. However, Hercules 3D Prophet 4500 doesn't have these output ports, so the chips aren't installed and there is no corresponding layout for them.
For our investigation we assembled the following testbed:
- AMD Athlon 1.2GHz (133MHz FSB) processor;
- ABIT KT7A (VIA KT133A based) mainboard;
- 256MB NCP PC133 SDRAM;
- Fujitsu MPE3084AE 8.4GB HDD.
We used the following software:
- Windows 98 SE build 4.10.2222 A;
- Quake 3 Arena v1.27;
- 3DMark2001 build 200;
- Unreal Tournament + patch 4.36;
- Serious Sam v1.00b ("Serious Sam: first blood" by Сroteam);
- Giants: Sitizen Kabuto + Pre-Patch v1.396;
- Homeworld: Cataclysm v1.00.
Here are the drivers we used in our tests:
- For graphics cards based on NVIDIA chips we took Detonator version 11.01;
- For ATI RADEON DDR 64MB VIVO we selected the driver version 7089;
- For Hercules 3D Prophet 4500 we chose version 7.89 driver by Imagination Technologies.
The driver by Imagination Technologies for Kyro/Kyro II lets adjust some properties of the graphics card:
Apart from the general card settings, you can create profiles for each application:
The driver stores the settings for each application in the registry:
We left all the cards settings by default except Vsync: graphics synchronization was disabled. In Quake3 Arena for 16bit color modes we took 16bit textures and for 32bit modes the textures quality was set to 32bit. Texture and details quality was set to the maximum, while all the other settings were left by default. We tested in demo127.dem which is included in 127g point release patch.
For Unreal Tournament we selected top texture and skin quality, enabled "Show Decals" option and dynamic lighting. In the additional Direct3D settings we enabled volumetric fog and reflecting surfaces. All the other settings were left default. We tested in utbench.dem.
For Serious Sam we tried two test modes: with "Speed" and "Quality" graphics and all other default settings. We tested in DemoSP03.dem.
In Giants we left the settings for all graphics cards at default. We used the benchmark integrated into Giants game.
Tests in 3DMark2001we run with default settings. For 16bit color modes we enabled 16bit textures and 16bit Z-buffer depth, and for 32bit modes we set 32bit textures and 24bit Z-buffer. Graphics cards with a T&L unit were tested in "D3D Hardware T&L" mode. For all the rest we set "D3D Software T&L" mode.
For a better comparison of the performances we opposed Hercules 3D Prophet 4500 to ATI RADEON DDR 64MB VIVO with 183MHz core and 183MHz memory frequencies, SUMA Platinum GeForce2 MX400 based on NVIDIA GeForce2 MX400 with 32MB 128bit SDRAM and SUMA PLATINUM GeForce2 GTS 64MB with 64MB 128bit DDR SDRAM and 200MHz core and 333MHz graphics memory frequencies.
In addition to these graphics cards we also pro9vide the results obtained for Creative 3D Blaster TNT2 Ultra graphics card based on NVIDIA TNT2 Ultra in some tests.