With AMD's recent acquisition of ATI, everyone has been wondering how the industry is going to be affected. In reality, we aren't going to see any major differences until at least 2008, but obviously plans could well change.
The launch of the X1950 can be marked out as the final true ATI launch, as it was taped-out and ready prior to AMD's buy out. Quite honestly, it couldn't have made a better card to end on as an independant company.
ATI has now launched several cards – the X1950 XT-X, the X1900 XT (256MB Edition), the X1650 Pro, and the X1300 XT. Although I'll only be looking at the X1950 XT-X today, these are all important launches and we will be looking at them all at a later date.
Pictured above you can see the reference X1950 XT-X (below) next to a Sapphire X1900 XT-X. The cooler is a lot fancier looking on the X1950, although there is something about chromed plastic that reminds of cheap children's toys. - I'd sooner have had it without. The X1950 is quite a bit heavier than the X1900 due to its all copper cooler, but heavy is good, heavy is reliable. You'll notice the fan has been moved to the rear of the cooler too, and is of a slightly different design.
From the back, it's hard to tell the two apart as the PCB is almost identical. In fact, this isn't far from the truth. The core itself is still an R580 and still operates at 650MHz, with 48 pixel shader processors, 16 texture units, eight vertex shaders and 16 pixel output engines.
There are two key differences. This is a slightly newer revision of the R580 with some minor improvements (which may or may not filter their way down to the X1900 range). The final difference is memory, which is now GDDR4 instead of GDDR3 and running at an effective clock speed of 2GHz.
GDDR4 is a welcome improvement over GDDR3 and is what the R580 core was designed to be paired with from the start. Aside from the improved frequencies, it is also considerably more efficient. There are several reasons for this, but one of the key ones is down to some clever 'bit flipping' - if you'd rather not understand how this works, skip to the next page. If you do stay here.
As you probably know, there are eight bits in a byte, and this can be represented as 00100110. It requires more power to represent a zero, than it does a one. So in that example, you would flip the bits and set a flag so the memory controller knows the data is flipped. So the memory controller would receive 11011001 and then flip it back to 00100110. If you think about it, using this technology, the maximum number of zeros ever being transmitted will be four, whereas before it could be as high as eight. That's quite a clever way of reducing power consumption!