RSS Feed

HCW Tech Blog

For the latest info on computer hardware, tech, news, video games, software tips, and Linux, check out our new improved front page: HCW Tech Blog

Reviewed by: Bryan Pizzuti [01.10.01]
Manufactured by: Hercules
MSRP: $59.99 (!)


Back in Time

We at HCW recently had a look at the Kyro2-based Hercules 3D Prophet 4500. Now we'll be having a look at its baby brother, the Hercules 3D Prophet 4000XT. This card is based on the Kyro1 chip, rather than using the newer Kyro2. As we explained in our Kyro2 article, the Kyro series of graphics processors are designed with finesse in mind rather than raw power. Instead of indiscriminately rendering everything, they will only render what will eventually be visible to the user. This allows the Kyro processor to get by with slower GPUs and, more importantly, less RAM bandwidth. While GeForce2s and Radeons get along with high speed DDR ram (Rated 300 MHz or faster) the Kyros can poke along on simple single data rate SDRAM running at the GPU's speed. For the Kyro2, this was 175 MHz. For the Kyro1, it's 115 MHz.

So, what's that mean, you ask? To do that will require a bit of explanation, and I'll be as concise as possible.

Conventional 3D cards, when receiving orders to draw 3D polygons on the screen and apply textures to them, simply do that. The program (be it a game or whatever) sends information about a scene, including all polygons contained in it, to the video card to render. This includes polygons, and parts of polygons, that are hidden from view by other polygons (Such as a character hiding behind a wall). Now, for a long time, this technology was more than sufficient, and to many extents, still is. However, this is what causes differences between advertised (also known as "theoretical") fill rates, and visible (what we tend to term as "actual") fill rates. As games get more complex, more and more polygons are generated that are hidden from view. Hence, as time has marched on, the difference between a card's theoretical and actual fill rates has grown larger.

Another factor affecting conventional 3D cards is a memory bottleneck. The RAM on a video card can only transmit so much data per second, depending on what speed and technology of memory chips they use, and how many bits "wide" the path between the graphics processor and the video RAM is. The crippled so-called "budget" video cards, tend to use SDR memory, as compared to high-end cards, which use DDR, which offers twice as much data transfer rate. Efforts are being made by companies to optimize the use of this limited pipe between the memory and the video processor (such as ATI's HyperZ technology), but the guys who made the Kyro had other ideas.

STMicro asked themselves, "Hey, how about taking all the stuff that the user isn't going to see out of the loop? Probably wouldn't need such fast memory, and the graphics processor wouldn't have to do so much work either!" So some of their researchers came up with the idea of receiving the scene from the program, and then splitting it up into a grid. Each tile in the grid is examined individually to see what will and will not be visible when texture rendering is complete, and the things that won't be visible are simply thrown away. Then the graphics processor computes the polygons and applies the visible textures to the visible surfaces and displays them. This saves so much memory bandwidth that, even with the fast-falling prices of DDR RAM chips, the chip is designed to use SDR RAM.

To illustrate this, we've asked our good friend Kevin Bacon to help us out (Kevin and Geoff the Goat go way back).

First, let's take a scenario where we have two Kevin Baconheads, one beside the other.  The Kyro II, and every other video card, will render each Baconhead in full form.

Now, let's say one Baconhead is partially obscuring the other... Most video cards will render all of Baconhead 1, and Baconhead 2, in full.  That's a whole lotta Bacon.

The Kyro gets away with rendering all of Baconhead 1, but only part of Baconhead 2.  This is how deferred rendering works... The less Kevin Baconheads it has to render, the more bandwidth it has for more important things.  Thanks Kevin!

Hercules is working closely with STMicroelectronics and PowerVR, the creators of the Kyro series of chips. They are offering both the Kyro and Kyro2 based cards, and are currently working to finish development on the Kyro2Ultra and Kyro3 graphics processors, which we hope to cover soon.

Kyro Chip features:
.25 micron process
115 MHz clock speed
115 MHz 128 bit SDR interface to max 64 MB SDRAM
Tile-based architecture
Environmental bump-mapping
2 pixel pipelines
8 layer multitexturing
DXTC Texture compression
32-bit Z-Buffer
Full-Scene Anti-Aliasing
250 megapixel/s fill rate
Motion Compensation
Internal 32-bit rendering

The fill rate and clock speed probably don't look that exciting, but remember, since the Kyro doesn't render what you don't see, it doesn't need to have as high a fill rate as some other cards. Their fill rate is theoretical, meaning it could render that many polygons, but they probably aren't all visible to the user. The Kyro's fill rate is what is visible to the user, because that's all it ever renders. Also notable is the lack of a hardware transform and lighting unit. This will hurt performance some, because it forces the main CPU to do those calculations, and it may or may not be able to do them as efficiently. A hardware T&L unit is rumored to be a pert of the upcoming Kyro3, but nothing official has been announced yet.

Next Page: (2)