- Page 1 nVidia GeForce GTX 280
- Page 2 GT200: Graphics Architecture
- Page 3 Counter-Strike: Source
- Page 4 Call of Duty 4
- Page 5 Enemy Territory: Quake Wars
- Page 6 Race Driver: GRID
- Page 7 Crysis
- Page 8 GTX 280: Test Setup
- Page 9 GTX 280: The Card
- Page 10 GT200: GPGPU Architecture and Other Features
- Page 11 GT200: Graphics Architecture
- Page 12 Verdict
Although GT200’s architecture is in many ways very similar to that of G80, there is a huge number of tweaks that have been applied to this new core that make it a considerably better whole. Before we get too far into the comparisons, though, let’s get back to basics.
The following is a Streaming Multiprocessor, or SM as nVidia abbreviates it to:
As you can see from the picture, an SM is an array of processing cores called Streaming Processors (SPs), with a portion of cache memory and an instruction scheduler. There’s actually a bit more to it than that, including two Special Function Units (SFU) that aren’t pictured, but we won’t go too far into the specifics. Essentially, this little lot acts like a mini eight core CPU with each core performing the calculations associated with a single pixel. The small (16KB) portion of memory is used to handle only the specific data associated with those eight pixels the SM is currently working on. This is the basic building block of nVidia’s unified shader architecture and is the same on both the G80, G92 and GT200.
(centre)”TPC from GT200”(/centre)
As we zoom out one step we see the first major dividing point between G80/G92 and GT200. In GT200, three of these SMs are combined to form a Texture/Processor Cluster (TPC), whereas G80 and G92 used two SMs per TPC. The principle’s the same, there’s just more in the new chip – 32 SPs vs. 24 on G80.
(centre)”TPC from G80/G92”(/centre)
As well as combining the SMs, a TPC also adds texture processing capabilities and again this is another area where the new chip differs from its forebears, sort of. On G80 you got four texture address units and eight texture filtering units per TPC, with G92 the number of address units doubled to eight while filtering remained at eight units. Now, with GT200 things have, well, stayed the same.
You still get eight texture address units and eight filtering units per TPC, it’s just the ratio of shaders to texturing units has changed. So, while the shader count of each TPC has increased by 50 per cent, the texturing power of each TPC has remained the same. At first this may seem like a backward step but most modern games are becoming shader dependent so the change in ratio makes sense. Moreover, when you look at the bigger picture you’ll see that the total texturing power of GT200 has actually increased a small amount.