What is the meaning of second generation?
When the PCX2 came out there were already signs of obsolition. Add-on board going against integration imperative, high CPU demand, and image quality issues. A second-generation gaming accelerator in my sense of the word should be able to provide all the features that one could find in the space back in 1996 already without any nasty surprises. Does it match what "second generation" meant for PowerVR? For them, the Series2 was again a product family. There was a CLX chip powering the Dreamcast console and Naomi arcades. And then there was PMX powering R-Cade Vision 250 and a PC card Neon 250. It was still NEC and VideoLogic Technology behind the designs. Of course, this article is all about the Neon 250. Based on the PowerVR 250 chip, it was one of those highly anticipated solutions burdened by "Voodoo killer" or later "TNT killer" hype. But silly those who would wait throughout 1998 for that deliverance. Only in the summer of 1999, when it would quite reasonably be considered another vaporware, Videologic finally made the move to release the Neon for PCs. And while the Series2 proved itself in Sega's products, the PC market full of the latest greatest architectures would have no mercy for the little Neon.
While NEC's manufacturing capabilities could not satisfy Dreamcast alone, VideoLogic was making only minor tweaks to the PC core, slowly adopting the new 250 nm process and waiting for free capacity. The idea of a 3D add-on board was canceled, which was the right decision since the whole market moved to complete solutions by that time. The new accelerator should be feature-complete and on par with the competition. After one of the longest delays of 3D accelerator ever the desktop product named Neon 250 finally launched in the middle of 1999 and of course performance was so last year. To sweeten the deal, Videologic doubled the originally intended 16 MB memory capacity while still asking for $169. Still, judging by the rareness of the card alone it probably was not worth releasing after such a delay. The PCI version to be reviewed here was out only by the tail end of the year. It should be no surprise to state right away the Neon 250 did not put pressure on contemporary architectures. But was there any silver lining?
The card
This review is a bit special because we are not looking at a card from the nineties, but at one made more than 20 years later. In the test is a modern recreation of Neon 250 PCI. This board will not take advantage of built-in AGP 2x support. But back in those days, I was always more fascinated by PCI cards giving AGP one run for their money, so this was my choice.
This is a "Neon Heritage" card designed by Anthony. You can find his creations here.
This is a superbly clean layout not only because it was made some twenty years later. The original Neon cards were somewhat bigger, which helped their passive cooling. This puppy is reminiscent of PCX. But although there are also only two memory chips, they are now on a 64-bit bus, twice the size of previous PowerVR cards. And of course, there is now a display output, since the Neon 250 is an all-around 2D & 3D accelerator. Where did the number 250 come from? Probably not after the 250MHz RAMDAC. Can it be a 125 MHz chip clock + 125 MHz memory clock? The increase in frequency compared to the 100 MHz Sega chip was one of the few advantages.
Architecture
Second-generation PowerVR with full floating-point geometry and texture setup engine put CPU demand in line with other solutions. Well-aligned with DirectX and equipped with a capable 2d engine Neon 250 could have kept PowerVR in the spotlight should it arrive sooner. Rarely used infinite planes were dropped. Now, having to deal with vertexes of triangles and quads the front-end could drop old complexities. On the other hand, Neon 250 brought other extras like a programmable processor in front of rendering. Up to 4 million polygons/sec can pass through, a huge leap from PCX2. And of course triangle setup is now done on the chip in full. The primitives are then marked for appropriate tile bins according to their potential visibility. The compiled display list is passed to the Image Synthesis Processor, our old friend ISP, to render a tile of 32x32 pixels. Its Z-buffer supports 32-bit precision without performance degradation. This internal tile cache, effectively a framebuffer of sorts, is still the key to PowerVR's way of avoiding overdraw. Depth tests can be performed here, much sooner than immediate rendering methods. The texturing engine, the TSP, was also improved in comparison with PCX2. It can work with non-square textures up to 1024x1024 pixels. There is still only one texture-mapping unit, but blending can loop to perform multi-texturing in one pass. Good for newer games, however, some might miss support for palletized textures. Once the work on a single tile is finished, it is written to local video memory. PowerVR survives with low-end memory bandwidth figures because usually no multiple reads and writes on pixels happen, and there is no need for a z-buffer there. The whole pipeline claims true color precision and with its single-cycle operation provides 125 million/s pixel and texel fillrate. But the tile-based hidden surface removal is still there to multiply effective fillrate in complex scenes. Other extra features are super-sampled AA, anisotropic filtering, table fog, and emboss mapping. For newer games, there is nothing special that would allow the Neon 250 to stand out. While the Dreamcast enjoyed some uniquely rendered games, its PC counterpart had no such game to wear on its sleeve. That being said, how does it actually do in games?
Gaming experience
One of the Carmageddon 2 problems: the smoke texture is not bilineary filtered.
Since the Neon uses the same deferred rendering that caused PCX compatibility issues with open APIs, one would expect the same hurdles. Again, the drivers offer an unusually rich set of options to help the user adjust to many situations. For example, Jedi Knight: DF needs depth clear style viewport. Wing Commander Prophecy requires tweaks to get rid of lines around crosshair and black backgrounds of UI elements. But there were some mistakes that I could not tweak away. From a lower deck of a carrier in Carmageddon II the sky was not rendered, it was simply black. Transparent parts of textures in Homeworld are rendered white. In a game as basic as Formula 1 there were no animated textures. Finally, the sky texture in the Insanity benchmark did not blend with clouds. Overall, the Direct3D problems aren't that numerous, not at all for a renderer as exotic as PowerVR.
However, the OpenGL situation is not that good. There was never any ICD, only a miniGL library translating into PowerVR's proprietary SGL API. It did not seem robust enough to venture outside of Quake based games. Further, texture mipmaps in GLQuake creep up too close to the camera. So much so that I wanted to force trilinear filter for the tests, but this option the controls do provide only for Direct3d. I am afraid it isn't even implemented. The trilinear filter by the miniGL in Quake 3 only pushed the mipmaps further, without proper transition between them.
Actually it can be worse. despite the colors indicating mip-maps at further distance when trilinear filter is on, I failed to see a difference in game. The silver lining for the Neon can be, that the texture LOD is of such quality the transitions are hard to spot anyway. There is an option to force an anisotropic filter, but the performance drop, as one should expect from 1999 hardware, is too punishing. For QuakeGL it was more than 40% loss, although it might be the worst case. In the end, benchmarks were taken at default settings, so keep in mind the OpenGL texturing quality does not have to be top-notch. And when going through screenshots, be cautious. The one from Sin shows many obvious color dithering artifacts, but I can't see them during play. So I suspect the miniGL might alter also the frame capture.
Table fog working in Star Wars: Shadows of the Empire. Rare without 3dfx.
But the majority of games use Direct3d and I have only minor complaints with its support in general. The driver may default to texture formats of 4444 or 5550 causing color banding especially on the green component. And this happens although even 32-bit RGBA textures should be supported. However, Neon 250 offers a few image quality perks of its own. Just like PCXs, it can force mipmaps smartly enough to avoid grain at a distance. Further, I do recommend forcing trilinear filter, because it indeed did not lower the performance at all. Despite many chips claiming virtually free trilinear filter, the Neon is the first reviewed card to actually make it a no-brainer.
Z.A.R. fails with OpenGL renderer, and even under Direct3d there are some transparency glitches.
View Neon 250 screenshot gallery
32-bit color resolution works as intended. On top of that the Neon 250 offers also middle-of-the-road 24-bit modes, where some speed can be gained in exchange for potentially rougher transparency handling. Such was the situation with the last driver version 3.2.1.2, released two years after launch.
Results
The Neon 250 is not a slow card. Not among the cards I have tested so far. Whence why SiS 300 was picked for comparison, to see which one will be on the provisional top spot.
There are some notable disappointments among older games. A second long reproducible hiccup in Resident Evil ruins minimal framerate. Daytona USA and Hellbender are very simple games, yet the Neon cannot move them as fast as it should. The difference between 640x480 and 1600x1200 in Myth 2 is telling. There is not much occlusion in the scene and the Neon 250 at very high resolutions performs like some generic low-end 64-bit card would.
Conclusion
Neon is short in average framerate by a measurable amount of at least 10%. The difference in minimal framerate widens to around 20%. It is well-known fact the Neon 250 was slower than the likes of Voodoo3, TNT2, and Rage 128 Pro. And now I have to place it behind the little-known SiS 300 as well. Speed is not a strong side of the Neon. PCX had its claim to fame in this regard, but the delays buried the PC incarnation of the second generation of PowerVR. The doubling of memory to 32 MB means little to nothing for a card with these characteristics. It can render 3d up to 1920x1440x16, but realistically, the fun ends at 1024x768. Even if it came a year earlier at the same price it would not be an easy sell. No wonder the Neon 250 became so rare. It is a bit surprising the driver support lasted for two years. There is always some room for improvement, but I believe they got most of what they could out of the hardware.
A word on the Direct3D full screen anti-aliasing. It does it's main job of smoothing out jaggies. Unfortunately, the loss of texture detail suggests this is a sort of "blur filter" implementaion, rather than super sampling. The performance cost is still quite prohibitibe: it cuts framerates roughly in half.
Rest of the story
The manufacturing constraint would push the designers to try a new business model by licensing the intellectual property to other vendors who would find their own ways of making PoverVR products. VideoLogic reorganized and changed its name to Imagination Technologies and continued working on the third generation. NEC left the PowerVR boat, but a new manufacturing and marketing partner was already signed. ST Microelectronics took that opportunity and delivered a new PowerVR product at the beginning of the year 2000 and received a lot warmer welcome. It was a dual pipeline deferred rendering architecture named Kyro. Having the tile cache on-chip played well for some new techniques. Kyro maxed out single-pass texture layers and processed super-sample antialiasing without higher memory usage. Also, stencil buffer operations were extremely fast thanks to tiles. Because of its very low frequency, Kyro was rather a budget product, but Kyro II with a big clock bump became one of the top performers. Now, only the absence of Transform and Lighting was severe hardware deficiency. This was to be addressed in a new four pixel pipeline core, but ST Micro decided to end its video chip manufacturing early in 2002. PowerVR cards had to face limited distribution and quirks of tiled rendering often required tweaking which only technojunkies did not mind. When properly executed some chips showed big advantages of deferred rendering and engineering talent of VideoLogic crew. Despite that Kyro sales were not so good, the fourth series was canceled and the PC market was abandoned. The fifth series of PowerVR products targeted mobile devices and there they finally achieved big success. The deferred tiling doing a lot of memory operations on-chip shines within space and power consumption constraints. Imagination Technologies was doing great in this rapidly growing market. So good that some of us are still keeping hopes for PowerVR getting back to PC in any form. But cutthroat competition remains.