Rendition Verite V2x00

Vérité 2

Second Vérité was supposed to appear in the summer of 1997, but it wasn't ready until fall. Upon arrival it failed to impress and interesting story came out of industrial sabotage causing the chip not reaching planned clocks. Whether this is true, or speculation coming from the surprise of Rendition not retaking leadership I cannot say, but it sure is one of the slowest clocked 0.35 um graphic chips. Fact is second Vérité entered the market at very low prices and even then had a hard time competing with other high performance chipsets like Voodoo, Riva or Rage Pro. There were two variants of the chip differentiated only by integrated DAC speed: V2100 with 175 and V2200 with 220 MHz. But card makers had some freedom in clocks. First came out Diamond's Stealth II with surprisingly low 40 MHz core clock, but thanks to ultra low prices it sold a lot. Later Diamond provided an updated bios boosting the clocks up to 55 MHz, the top clock I've seen on V2x00 cards going through my hands. Fair amount of card vendors picked second Vérité, but despite decent availability it wasn't exactly a hit. I chose a model which is kind of middle way among cheaper V2200s.

The Cards

Up here is notorious Stealth II with V2100 chip and 4 MB of RAM.

And down below are AGP and PCI versions of V2200 cards with 8 MB of memory. The logo on bios chips is Manli for what I know. One has to choose from different clocks, memory capacity and buses. I decided to go with the Vérité 2200 8 MB PCI. These are late produced cards from 1999/2000. Just imagine selling anything based on Vérité so late, must have been offered for ultra low prices. Rendition sorted their V2200 chips into two grades, A and B. Most of the dealers went with slower B grade and A grade was maybe exclusively provided to Hercules, which was shipping a full featured 62 MHz Thriller 3D. But collectors dream is Bonnie and Clyde from Jazz Multimedia offering both PCI and AGP interfaces on single board. Well, I own cheap cards and those will be tested. Chip is running at 50 MHz and SGRAM memory at 100 MHz. Not the top clocks, but 8 MB can give it some breathing room in newer games. Bus width remained at 64 bits, providing 800 MB/s bandwidth. There shouldn't be much difference between PCI and AGP cards, because second Vérité features classic PCI controller capable of 66MHz operation and is only electrically compatible with AGP 1x. There were driver releases with AGP specific inf enabling(?) AGP texturing, but I could never get it to work. If you want to try it for yourself just search in your registry for agpflag and set it to 1. Chip:memory ratio of 1:2 is typical, but programmable clock generator accepts different values. At least 2:3 ratio should work as well. However, anything else than 2:1 caused my PC to freeze, so I doubt clock dividers can be changed on the fly. This may have contributed to uninspiring overclock among my V2000s. On the tested V2200 PCI core seems to top before 60 MHz and memory around 110 MHz despite 7 ns rating. The AGP card with 8 ns RAM did much better, with memory artifacts appearing at 140 MHz. Compared to V1000 we have a chip ticking at similar speed, only memory frequency doubled. But Rendition made many improvements in the architecture.


Grand Prix Legends is the highlight of RRedline library.

Inside V2x00

Rendition stayed with the concept of combining RISC CPU with fixed function pixel pipeline, though both are much improved. The processor is supported by both data and instruction cache and can do the triangle setup and other tasks as usual. New are graphics specific instruction, single cycle float to integer conversion and bigger register file. Instructions for next stage are issued per cycle in VLIW3 format. Such a processor could help Vérité create interesting special effects, it can quickly perform branching and program jumps which were weak point of recent programmable shading architectures. Direct control over pixel pipeline allows scissoring and other elaborate masking. But again, just like with first Vérité, developers were not interested. Between the RISC engine and pixel pipeline is now extra triangle engine, allowing triangle scan line and rasterization in parallel execution. Fed by three effective instruction per clock, the triangle engine outputs with all parameters one span or pixel per cycle. It also prepares edge antialiasing. Then finally comes the pixel engine, again strongly enhanced to render diffuse and specular shading, bilineary filtered texturing, z-buffer comparisons, fog and alpha blending, all at one cycle per pixel. Problematic 2d operations are also boosted by dedicated hardware for shape filling and hardware cursor.
Obviously single cycle operation is the keyword underlining V2000 architecture. Rendition catch up with the big bullies in clock for clock performance. Gone are the big performance Z-buffer hit and choking on blending. But competition did not stood still and feature wise it does not stand out like V1000 did. It is still a high quality renderer, but support for per-pixel mip mapping could help to stay on the top. It seems 32 bit depth was working only in OpenGL for a long time, I could not get any 32 bit in D3D until last beta release from summer 1999. 2D and video also received needed improvements, even TV-out is supported. I found 2D speed satisfactory in all situations. V2000 is still using the wonderful asynchronous DMA engine, however I got her quite stable on two different motherboards.

Experience


There should be shining lamp in the corner, older driver is fine though.

I already hinted second Vérité does not stand out in the image quality as much as V1000, because competition moved on while Rendition did not aim higher. Maybe some corners were cut in performance hunt, because I saw some quite noisy textures in several games. Nothing big but I felt a bit disappointed after V1000. It turned out I mixed beta releases and instead of latest 3.0 beta 5 used an earlier driver from Rendition SDK. Because of that text in Falcon 4.0 cockpit is broken, just like fence in Viper Racing or missing dynamic lights in Heavy Gear. Ultimate Race Pro could not run at all. And when one tries to lower the resolution he usually finds out 512x384 is out of specs and 400x300 completely missing. Then one Vogoner (thanks idspispopd) bugged me to try another driver so I gave it a shot, with Hercules one. And what do you know, all these bugs went away with the exception of Falcon 4.0. It took two drivers, but compatibility is nearly perfect, good job Rendition. The strange thing is Unreal cannot work via Direct3D, only OpenGL works. While OpenGL drivers of V2000 are doing great in Quake games, Unreal was far from playable. There are still few unsupported effects in Quake 3 though. Check out the gallery. Then there is almost OpenGL specific 32 bit depth support. I was shocked to see Quake framerates dropping below half of 16 bit results. Worried about running out of video memory I disabled triple buffering, via environment variables because register entries did not work for me as readme for the ICD 1.1.5 suggested. All I got was another speed drop. At least I gathered enough data for graph just for OpenGL.


Who knows why is the 32 bit penalty so high in Quake games. Unreal did not lose that much but runs too slow to begin with. You can also see how little V2000 improves VQuake performance. Since this renderer was strongly optimized for low fillrates of V1000 it should be very much dependant on CPU speed. Then again VQuake scaling is diminishing even with CPUs multiple times faster than Pentiums of the time.
Because of my sloppy job I missed the opportunity to test the last beta release. It contains newer OpenGL which did not gave me any improvements, but a quick run of Direct3D tests showed some enhancements. I don't want to discard data gathered so far, after all this is what V2000 used for most of its lifetime. Instead, I will later add V2200 AGP results clocked at usual 55,5/111 with this last driver to get even more numbers. The last driver has its bugs as one would expect from beta. There are some random gray polygons blinking in the sky of Falcon 4.0. Shots in Forsaken in 32 bit mode are not rendered properly, yet speed is the same as in 16 bit depth. Screen in Mortal Kombat 4 is mostly black. Alpha in Shadows of the Empire is broken. So is texture filtering in Turok, which may explain the performance jump. Dungeon Keeper fails completely. Incoming looks awful without sky and objects rendered. Neither driver handles true color under Direct3d reliably. Despite this there are some speed ups worthy trying this 3.0 driver first. It is a complete rewrite from May 1999, too bad we did not get to see final version. It deserves it's own gallery.

How fast is it?

I saw the Vérité not catching up with Voodoo Graphics so I chose Rush for comparison. To add more variability AGP Vérité is using gold minigl 3.0 instead of the ICD.

PCI Vérité at 50 MHz turned out pretty even with Rush in average framerate, and slightly loses in lowest framerates. AGP Vérité with 55 MHz clock and latest beta driver takes a slight lead in average framerates, even after removing the suspicious Turok result and multiplicity of Quake demos. Minimum framerates improved significantly and now is the Rush actually loosing by 10%. MiniGL driver has no advantages over ICD, at least as long as fast CPU is used. There is no doubt V2000 retailed with great price/performance ratio, problem is Rendition had to take lead in other areas as well. Top models with higher clocks, 8 MB, AGP and slower CPUs of that time could challenge Voodoo Graphics, but Vérité 2000 is one year younger. Rendition's market share remained rather small and it would take a faster product to become again a brand gamers dream about. Once favourite company was in a trouble.

Exitus

Not even second Vérité provided earth shattering performance, it has however led in gate efficiency. Such features and speed divided by mere 1.8 million transistors show a strong side of Rendition's approach. V2x00 could have one more interesting incarnation. Rendition teamed up with IBM for manufacturing and Fujitsu to build first consumer Transform and Lighting accelerator. Fujitsu's DSP-based geometry/lighting processor FGX-1 should've been combined on single board with V2200. There are pictures of Hercules Conspiracy prototype board on the web. But no other product of Rendition was allowed to reach gamers. Of course, there was next chip in development, the high-end V3300 RRedline for early 1999. Performance was supposed to be 3M triangles/second and 200 megapixels trilinear fill rate, but since competition was announcing higher performing solutions Rendition did not see it as strong enough. The decision was made to skip V3 and go for fourth Vérité architecture right away. Looking back it was probably unnecessary pessimism since competitors overestimated capabilities of new 0.25 um manufacturing and failed clock targets en mass. On the other hand, since it was memory specialist Micron that bought Rendition, new interesting technological fusion was going to propel Vérité 4 to new performance levels: embedded memory. But this project was never completed, at least not in the form of discrete graphics. The technology was used only on Micron motherboards, Rendition as a gaming graphics brand became a history and today is used only for Micron's memory.