Following the GeForce GTX 460, it’s now the GeForce GTS 450’s turn to be reworked by NVIDIA. As with the other reworked cards in the GeForce 500 range, the resultant GeForce GTX 550 Ti enjoys higher clocks and higher performance. Does this mean NVIDIA now have a more competitive model to face down the Radeon mid-range? We’re going to try and answer this by giving our analysis of the DirectCU TOP model from Asus.
The reference GeForce GTX 550 Ti.
GF116 and heterogeneous memory
The GF116 is to the GF106 what the GF114 is to the GF104. In other words, it’s not a new GPU but rather a new revision of the chip designed to reduce current leakage and facilitate clock increases. To recap, here are the main points of the internal architecture of the various GPUs in the GeForce 400 and 500 ranges:
- GF110 = 4 GPCs x 4 SMs x (32 processing units + 8 texturing texturing units)
- GF100 = 4 GPCs x 4 SMs x (32 processing units + 4 texturing units)
- GF104/114 = 2 GPCs x 4 SMs x (48 processing units + 8 texturing units)
- GF106/116 = 1 GPC x 4 SM x (48 processin units + 8 texturing units)
The GF106/116 is basically half a GF104/114, except that it has a 192-bit memory bus as against a 256-bit bus. While the GeForce GTS 450 had to make do with a cut down 128-bit bus, the GeForce GTX 550 Ti benefits from a 192-bit bus in total… but with 1 GB of memory. How can this be? In principle a 192-bit memory bus should function with 768 MB or 1.5 GB of video memory.
NVIDIA says that it has slightly reworked its memory controller here to support asymmetric memory density. More or less all the GPUs run with one or more 64-bit memory controllers. In the case of the GF116, three of these controllers share access to the video memory. The innovation comes from the fact that it is now possible to use chips with varying density for each controller. Here’s the typical configuration for the GeForce GTX 550 Ti:
- memory controller 0: 2x 128 MB
- memory controller 1: 2x 128 MB
- memory controller 2: 2x 256 MB
The GeForce GTX 550 Ti therefore has four 128 MB chips and two more 256 MB chips. In practice, its GPU thus has a ‘high speed’ 768 MB memory that’s accessible at 192 bits and a ‘slow’ additional 256 MB accessible at 64 bits. We therefore suppose that NVIDIA have worked mainly on the software side and that the correct functioning of the system is based on the ability of the drivers to put the most performance dependent data in the part of the memory that is accessible at 192 bits.
You’ll note that NVIDIA has decided to revisit its naming structure for this new model. Why they’re using the Ti suffix still isn’t very clear, but we imagine that it will be reserved for the biggest models in each series should other models see the light of day.