Radeon HD4870 review
Author: Luka Rakamaric
Date: 09 Sep 2008

Today we are taking a look at the fastest ATI single chip card currently available, the Radeon HD 4870. Just like with the 3800 series, the 4870 is based on the same GPU as the 4850, but almost everything else is different.

The GPU

The RV770 core comprises almost a billion transistors (956 million) and is substantially larger than the previous RV670. Still, NVIDIA's GT200 GPU has more than 1.4 billion transistors and is built using a larger production process, so ATI’s GPU is still relatively small and, more importantly, has a lower heat dissipation. It is also cheaper to produce, both in terms of chips per wafer and yields of working chips.  The RV770 GPU has quite a lot of computational power, because the number of processors has been increased from 320 to a round 800, or 250%. This 800 processors number has to be taken with a pinch of salt, because it is actually 160 five stage pipeline processors. In the best case scenario, where the operations are not dependant on data from the previous ones, it can achieve its full processing power, but in most real situations, the input data for the calculation is dependent on the output data from the previous operation. That is why ATI’s number, although significantly higher than NVIDIA’s does not actually represent the same thing. It would be unfair to say you have to devide ATI’s processor count by five to get a real number, but from what we could see in 2900XT and 3800 series, it is not far from the truth. The texture mapping units also received an upgrade from 16 to 40, but the raster operator units remained at 16. The chip still retains the 4:1 compute to texture ratio. NVIDIA’s GT200 offers 3:1 and their previous generations had to do with 2:1 ratio. The texture processors received a major upgrade. In R600 and RV670 they were grouped in four groups of four. In RV770, each of the ten texture units has 4 address processors, 16 FP32 samplers and 4 texture filters. 32 bit filtering is now 250% faster and 64 bit is around 150% faster.

ATI decided to stay with the lower priced 256 bit external bus for the RV770. In the 4850 model, this means that the memory throughput is not that great, but with 4870 they decided to use GDDR5 memory which offers 4 times the effective bandwidth instead of just 2 like with GDDR3 and GDDR4. This effectively offers the same performance as using a 512 bit bus, but without the need for the complex PCB design which is why the 4870 will still be a rather affordable product.