by FastSite
07/19/2001 | 12:00 AM
Kyro II graphics chip from STM, which licensed the PoverVR Series3 line from PowerVR, managed to win the hearts of two famous names at a time: Hercules and Videologic. Both well-known graphics cards manufacturers introduced their graphics solutions based on Kyro II - 3D Prophet 4500 and Vivid! XS. In spite of the pressure NVIDIA exerts on the manufacturers, the number of manufacturers announcing Kyro II based cards keeps on growing. The Korean SUMA company, which has become pretty beloved due to its high-quality products offered at favorable prices, couldn't stay aside. SUMA developed a new SUMA PLATINUM K2 family of Kyro II based graphics cards with 32MB and 64MB memory. <%BANNER[article]%>
This article touches upon SUMA PLATINUM K2 32MB graphics card. So, let's go!
Like the cards built on NVIDIA chips, SUMA PLATINUM K2 is shipped in a brand box made of transparent plastic. The only different thing is the color:
The package includes the card itself, a CD with drivers and a brief user's manual made as a folding brochure:
SUMA PLATINUM K2 is designed on a lacquered black PCB - a traditional solution of SUMA's. It is equipped with Kyro II chip from PowerVR/STM and 32MB graphics memory by Samsung with 5ns access time:
This card almost coincides with Hercules 3D Prophet 4500 64MB in design, but on SUMA PLATINUM K2 you'll find a different regulation circuit for the core power supply and no memory chips on the rear side of the card.
It's noteworthy that the twice as small number of memory chips by the SUMA baby in comparison with Hercules 3D Prophet 4500 64MB doesn't tell negatively on the memory bus bandwidth. SUMA PLATINUM K2 has a wholesome 128bit memory access. The card is equipped with four 2Mx32 microchips from Samsung, i.e. each of them has 2Mbit address space and 32bit access making 64Mbit (8MB) capacity. So, the graphics memory size in general equals to 32MB (8x4). Hercules 3D Prophet 4500 has 8 similar microchips providing 64MB memory (8x8), though its data bus is not 256bit (8x32) wide, but 128bit again. This trick can be explained by Kyro II's ability to work with double data banks switching over from one group of microchips to another. Accordingly, as the chip works with 64MB memory, its controller is responsible for carrying out this switch-over among the 32MB banks at the right time.
Kyro II supports 16bit memory chips too, but in this case the graphics memory is organized in a conventional way. That is, Kyro II supports 32MB configurations of 4 x (2Mx32) and 64MB configurations of 8x (4Mx16) using a single bank or 8x (2M32) using two 32MB banks.
The memory chips installed on SUMA PLATINUM K2 feature 5ns access time and 200MHz maximum working frequency:
The chip is equipped with a "deadly-fixed" cooler, which we can also see on SUMA GeForce2 MX based graphics cards.
The clock frequencies of the chip and memory are synchronized and make 175MHz.
The card's developers haven't provided any TV- and DVI-Outs with corresponding chips, but they've left enough space for them on the card PCB.
For our investigation we assembled the following testbed:
We tested in the following software:
For Kyro II based graphics cards we took the freshest driver available at that time - 7.111 for Kyro/Kyro II from Imagination Technologies. NVIDIA GeForce2 MX 400 based cards were tested with Detonator 12.41.
Since SUMA PLATINUM K2 differs from Hercules 3D Prophet 4500 64MB just by less graphics memory, both cards are supposed to be pretty close in performance except, perhaps, the highest resolution modes with 32bit color depth and textures, namely when 32MB of memory can be not enough to store the textures.
Keeping this fact in mind, we tested in Quake3 Arena (a popular OpenGL test) and in 3DMark2001 with plenty of textures (especially in the "High Details" mode).
In Quake3 Arena for 16bit color mode we selected 16bit textures, while for 32bit mode we set 32bit textures. We chose the highest texture and detail quality, enabled trilinear filtering and texture compression. All the other settings were left by default. The tests were run in demo127 from Quake3 Point Release Patch 1.27g.
In 3DMark 2001 the settings were the following: for 16bit color mode we enabled 16bit textures and 16bit Z-buffer depth; for 32bit modes - 32bit textures and 24bit Z-buffer. The graphics cards were tested in "D3D Software T&L" mode.
For a better comparison we also tested two NVIDIA GeForce2 MX 400 based graphics cards with 32MB and 64MB memory.
As you can guess, demo127 is not so rich in textures to make SUMA PLATINUM K2 32MB transferring the texture data via AGP bus lose the fight. For the most part, it's for the peculiarity of Kyro II: it doesn't store the Z-buffer data in the graphics memory, this way leaving some extra space for the textures.
With the full-screen anti-aliasing SUMA PLATINUM K2 and Hercules 3D Prophet 4500 64MB showed identical results. The reason stays the same - to carry out the full-screen anti-aliasing, Kyro II doesn't need to store enlarged frame buffer and Z-buffer in the graphics memory. All the additional strain falls upon the core, and the core clocking of the two cards is equal.
The NVIDIA GeForce2 MX 400 based cards with 32MB and 64MB didn't show any performance difference either.
7.111 driver gave us an opportunity to involve trilinear approximation (we have already supposed that it exists by Kyro II in our first review of Kyro II based graphics card). However, this function turned out rather weird: in Quake3 trilinear approximation could be enabled only with enabled texture compression and it works only for those textures, which were compressed. Here are the screenshots to illustrate what we mean:
 |  | |
| Bi-linear Filtering | ||
 |  | |
| Tri-linear Filtering | ||
 |  | |
| Tri-linear Approximation | ||
Maybe, the fact that tri-linear approximation works for compressed textures only is a result of S3TC textures being stored and transferred in 4x4 texel units, 4bit per texel, giving 64bit per unit on the whole.
On the other hand, let's remember how the texture MIP levels are organized: the color of each texel from a more "remote" MIP level is created by blending the color of a 2x2 texel unit from the "closer" MIP level. So, the 4x4 texel unit corresponds to the 2x2 unit at the more "remote" MIP level. That is, we have all the data needed to perform bi-linear filtering at the current MIP level as well as at the next one. Consequently, with a 4x4 texel unit we can run tri-linear filtering. We should only figure out correct color coefficients when blending the sample colors. Formally this filtering will not be called tri-linear in the full meaning of this term and in Quake3 we won't see any smooth transitions from one color to another if we "color-mark" different MIP levels. Although from the math1ematical point of view, it won't be none other but pure tri-linear filtering.
Of course, this method works without texture compression as well, but it proves unreasonable in terms of memory in this case. That's why the developers from Imagination Technologies resorted to the common tri-linear filtering then.
In terms of the consumed computing capacity, this method is also not that perfect, but to get the color one can use 8 or even 4 texture samples instead of 16. It will shorten the calculation process, though the image quality will be poorer. Kyro II is very likely to act like that when tri-linear approximation is enabled.
If we compare visually the quality of tri-linear approximation carried out by Kyro II based graphics card and NVIDIA TNT/TNT2 based solutions, we will have to admit that in static and slowly changing scenes Kyro II performs much better. As for NVIDIA TNT/TNT2 graphics cards, there is even a special term for their tri-linear approximation: "Lod-dithering". Anyway, despite the excellent quality of static images provided by Kyro II, its tri-linear approximation in dynamic scenes bears some unpleasant artifacts. It looks as if the number of MIP levels were doubled against the common bi-linear filtering. The number of MIP bandings doubled as well, but they became far less noticeable.
At the following diagram we estimated the performance losses caused by tri-linear filtering (to be more exact, by the mixture of tri-linear filtering and tri-linear approximation) in Quake3:
The performance drop in case of this "mixed" mode are quite acceptable and make about 12%, but as we wrote in the previous review, "real" tri-linear filtering with enabled texture compression leads to a similar performance drop.
One of the reasons why we decided to compare the performance of SUMA PLATINUM K2 with that of Hercules 3D Prophet 4500 64MB is that 3DMark 2001 tests show better results when run on graphics cards with a hardware T&L unit. Therefore, it wouldn't be so very exciting to compare SUMA PLATINUM K2 with NVIDIA GeForce2 / GeForce2 MX based cards, while Hercules 3D Prophet 4500 64MB features the same tile architecture as SUMA PLATINUM K2 but has more graphics memory onboard. Besides, we have already made a comparison of Kyro II and graphics cards with conventional architecture in the article devoted to Hercules 3D Prophet 4500 64MB.
In "Game3-Lobby", the easiest test with a moderate amount of textures and polygons, SUMA PLATINUM K2 and Hercules 3D Prophet 4500 64MB go abreast. Fearing to overload the article with heaps of diagrams, we confined ourselves just to the results obtained in "Game1 - Car Chase" and "Game2 - Dragothic":
As you can see, only in 32bit modes SUMA PLATINUM K2 falls behind Hercules 3D Prophet 4500 64MB because of a growing frame buffer and larger textures.
The two NVIDIA GeForce2 MX 400 based cards (with 32MB and 64MB graphics memory) run pretty evenly, there is a very slight performance difference between them.
SUMA PLATINUM K2 provides splendid image quality. We noticed no blurring up to the resolution of 1600x1200.
Although SUMA PLATINUM K2 is equipped with only 32MB graphics memory, it falls behind Hercules 3D Prophet 4500 64MB only in the hardest modes, say at high resolutions in 3DMark 2001 gaming tests. It is a great alternative to NVIDIA GeForce2 MX 400 and even NVIDIA GeForce2 GTS based graphics cards.
Thanks to the reduced memory size, graphics cards of the kind will cost not so much as Hercules 3D Prophet 4500 64MB. Relatively low price gives SUMA PLATINUM K2 cards on PowerVR/STM Kyro II another advantage over its competitor and ensures a positive customer's decision.
Highs:
Lows: