Survey of mid-range video cards (page 1: Introduction, X1000 family)

Introduction

Several new products have recently hit the video card market. There are the cards based on the X1800 (ATI´s new high end product) and X800, and then those derived from the NVIDIA 6800 and 6600. This article is here to shed some light on this situation and see how they stack up against each other.

Cards with the X1000 architecture

For once, ATI has decided to simultaneously release an entire line of GPUs, from basic to high end products. We remind you that this architecture supports the vertex pixel shader 3.0 and works with small batches of pixels. It remains efficient even with shader branching. For more details on this subject, take a look at our Radeon X1800 test.

The Radeon X1600 and X1300 are based on the same architecture. There are, however, a couple of differences. The Radeon X1300 is more or less a fourth of a Radeon X1800. It only features two geometric processing units (instead of 8) and one shader core (instead of 4), which includes 4 pixel shading pipelines. Next to this shader core is a group of four texturing units, compared to 4 groups for the high-end card. It´s the same for the ROPs, which give the Radeon X1300 a quarter of the X1800´s processing units. The ring bus, which increases memory controller efficiency, isn´t included for chip simplification. Memory access is made in 4 x 32 bits compared to 2 x 64 bits for competing GPUs, which wastes less memory bandwidth. This chip is mainly intended for entry level products, nevertheless the "Pro" version that ATI sent us can be considered as mid-range with a price of 125€ / $120.

A special chip

The Radeon X1600 has the particularity of being the first GPU, which fully benefits form the possibility of decoupling pixel shader processing and texturing units. Because of its architecture, ATI can include more units of one type, and of course the number of pixel shader units was the first to be increased. If the previous ATI mid-range GPU, the X700, included 8 pixel shader pipelines and 8 texturing units, the X1600 includes 12 pixel shader pipelines and only 4 texturing units. ATI´s choice comes from the fact that shaders have increased more than texture accesses in complexity and in the number of mathematic instructions. In our tests, we will check if the 3:1 proportion is well adapted to current games or if 4 texturing units aren´t enough.

This difference between the X1300 and X1600 on the "pixel" part will probably also be found between the X1800 and the next GPU, which will feature 48 pixel shaders and 16 texturing units. Another specificity is that even if the 12 pipelines are physically related to the 3 shader cores, they aren´t really independent like the 4 of the X1800. It is more a single 12 pipeline shader core than 3 shader cores of 4 pipelines. It means that this GPU works with batches of pixels (threads) three times bigger, and that branching efficiency is reduced. Despite the size increase of 16 to 48 pixels, it´s still much lower than the 1024, 2048 or 4096 batches used by NVIDIA. ATI can still keep its advantage on this level.

For other operations, this GPU has 5 geometric processing units, 4 ROPs and the ring bus. Memory is accessed in 4 x 32 bits. Comparison with the X700 is a bit strange, because besides pixel shader calculation units, which are increased by half and with the same memory bus, everything else is reduced.

ATI only had a couple of cards available for tests. Production of this GPU is only beginning and cards will be available at the end of November. This chip is an "engineering sample" from the latest revised wafer tests, which has encountered several problems in frequency increase with the 90 nm process (like the R520). This has delayed production.