S3's S4While the heart of this article is examination and review of the multimedia features provided by OmniChrome, I'll cover, quickly, what the S4 GPU that powers the initial OmniChrome products is capable of, from a specification standpoint.
|Process||130nm @ TSMC|
|Fragment Processor Setup||1 vector ALU, one texture ALU|
|Fragment Processor Precision||FP24|
|Traditional Render Setup||4 x 1|
|Basic Texture Filtering||Bilinear|
|Texture Filtering||Bilinear, Trilinear, 16X Anisotropic|
|AA Sample Type||Rotated grid, 4 sample|
|Native Bus Support||AGP8X|
|Basic Core Frequency||325MHz|
|Basic Memory Frequency||300MHz|
|Memory Bus Width||128-bit, memory crossbar (2 split)|
|Basic Pixel Fillrate||1300Mpixel/sec|
|Basic Multitexture Fillrate||1300Mtexel/sec|
|Basic Memory Bandwidth||~9.60GB/sec|
At 325MHz with four pixel pipes (each with a texture sampler), two vertex shaders and DX9.0 compatibility using 24-bit per component fragment precision, it's highly reminiscent of RV350, ATI's mid-range (and now low-end) GPU product. The differences lie in the S4's anti-aliasing ability and desktop colour precision. S4 can't do multi-sampling, anti-aliasing geometry along the edges of objects in the scene. It can only do super-sampling, anti-aliasing the image using a 2x2 rotated grid sample method before the final buffer swap in the render process, after all image output has been processed and has been through the ROPs at least once. Super-sampling AA uses more fillrate and memory bandwidth than multi-sampling (MSAA), so although it offers good AA image quality, anti-aliasing texture data too, it's at the expense of lower performance than MSAA.
In terms of desktop colour precision, S3's S4 and S8 GPUs support 30-bit basic colour precision on desktop images. Each red, green and blue colour channel gets 10-bits of precision, the remaining 2 bits used for limited alpha channel information. Almost all other hardware, with the notable exception of Matrox's Parhelia range, only supports 24 bits of precision in the main colour channels, with an 8-bit alpha channel.
When set to 32-bit colour on a Windows XP desktop, the driver automatically uses the A2R10G10B10 back buffer format. It can cause programs with screenshot capture programs that don't know how to process that buffer format, but delivers the highest quality static image output possible for RGB images with a 32-bit framebuffer, offering over 1 billion separate RGB colour variations. A 10-bit DAC partners the output quality for maximum image quality when processing 2D images. I'm not certain if Matrox's 30-bit colour plug-in for Adobe's Photoshop works with S4 and S8 (and indeed my LCD monitor lacks the colour range to display a true 10-10-10 image anyway), but it's something present in S3's hardware and will be in all future parts, too.
DDR memory support lets AIBs take advantage of low cost DRAMs, outfitting S4 and S8 boards with up to 256MB of card memory. What's also interesting is the relatively low transistor count of S4, a svelte 40M transistors. Using TSMC's 130nm node and knowledge of the clocks other 130nm 4x1 designs are hitting, we can theorise that all other things being equal, there's a little bit of overclocking headroom in the S4, clocked at only 325MHz.
The basic performance characteristics of S4 are well known by now, Tarinder taking a look at the S8 for us in the past and S4 having half the power at the same clocks. I won't cover performance in anything more than a fleeting manner, so be mindful that while it's nothing incredible and it lacks in certain areas, it's a decent GPU nonetheless.