Gigabyte Radeon HD4850 review
Author: Luka Rakamaric
Date: 16 Jul 2008

It all started with the pretty unsuccessful X2900XT, which tried to compete with NVIDIA’s flagship but failed to get close. It was leaking current like crazy, overheating, some things didn't work at all and left a bad impression overall. Since then, AMD has been offering us great value for money products, and the 4850 follows the same philosophy. Its bang for the buck should make users think twice before considering any other cards because - quite simply - it's dirtcheap for its performance. But let's check out what HD4850 has to offer...

The GPU

The RV770 GPU is the base of the the new 48xx series of ATI’s cards, which at first will comprise two models, 4850 and 4870, just like it was with the 38xx. The GPU itself is quite large, although not even coming close to NVIDIA’s GT200.  Still, ATI managed to squeeze in 800 unified shaders, which is a substantial increase from the 320 found in the RV670. It required putting almost 1 billion transistors on the 55nm die, up from 666 million in the RV670. With the increase in stream processors came the increase in texture units, 40 compared to 16. The number of ROPs stayed the same, on 16. To put things into perspective, let’s look at the table:

Because of different architectures, 800 is not more than 240, it is actually quite a bit less. There are situations when the shader code is optimized, so almost all of the 800 processors are fully utilized, but more of than not that is not the case. In theory all 800 processors could be loaded if you were doing calculations whose inputs were not dependent on the result of previous calculations. However that happens very rarely, and is one of the reasons ATI cards could not reach the performance of NVIDIA’s GPUs which had fewer processors but all of them were independent.  In ATI’s case, five processors are grouped together, and only one of them can do complex operations while the other four can only do MAD instructions. For all intents and purposes each of the groups functions as a processor with a five stage pipeline. In the worst case scenario, only one fifth of the chip will be operating, which lowers the number of shaders to 160, still substantially more than 64 that we had on the RV670 GPU. ATI also states that the load on the GPU is around 40% more efficient, which combined with 250% more processors makes for a very interesting chip.

On the other hand, NVIDIA’s 80 Texture mapping units don’t bring a 100% advantage over ATI’s 40 in the RV770 as it might seem on first glance. Although the number of ROPs stayed the same, they are now twice as effective so the RV770 can process 64 Z values per clock instead of 32 in RV670. The next new thing ATI brings us is the abandonment of the ring bus. It now uses a hub switch that improves the efficiency to almost 100%, which is quite more than the 85% with the ring bus. The reason for that is that not all components of the GPU require a 512 (R600) or 256 (RV670) bit connection.

ATI has had a ratio of 4 computing to 1 texture mapping unit for a couple of generations now, while NVIDIA has just increased theirs from 2:1 to 3:1 in the last generation. With such a powerful GPU as the RV770, there shouldn’t be a problem on either side, unlike the RV670 which did experience performance drops at certain scenarios.