NVIDIA GeForce GTX 580 1.5GB Review and SLI Testing - GF110 brings full Fermi
GF110 is a new GPU
What should have been
It might be cliche to say it as this point, but despite the launch of NVIDIA's GeForce GTX 480 and 470 cards offering the best performing single GPUs available at the time, there was a lot about the GF100 GPU that didn't go as planned. First, it was late and left NVIDIA with a 6+ month deficit in DX11 market share. Secondly, and most pestering to consumers, the GTX 480/470 GPU was hot and noisy.
The GPU itself never actually met the specifications that NVIDIA set out to reach 4-5 months prior to its launch either. The GPU was capable of including as many as 512 shader cores according to NVIDIA's own information, but when the GTX 480 launch in only enabled 480 of them in order for the company to meet temperature and yield demands. The oft-discussed "full" Fermi GPU was never seen. Until today.
The GF110 GPU
As you would expect, not only do we get additional CUDA/shader cores in the GF110 GPUs but also additional texture and tessellation performance include with the addition of that extra SM. Texture unit count goes from 60 on the GTX 480 to 64 on this new GF110 GPU and this new iteration sees an additional bump thanks to more Polymorph Engine power. The ROP count remains the same though with a count of 48 since our memory bus also remains the same at 384-bits wide.
The GF110 is still a 3.0 billion transistor GPU (despite rumors to the contrary) and has not dropped features like ECC to reach these performance levels. The secret to NVIDIA's improved performance and yield comes from a monitoring and rearranging of the transistor paths that caused the most leakage on GF100 in real-world environments. NVIDIA's engineers then rearranged many segments of the die to better improve efficiency and power consumption finally allowing them to enable the full power of the architecture as GF110.
There have been some other architectural enhancements made on the GF110 GPU as well including a full speed FP16 texture filtering addition and improved Z-cull efficiency. These clock-for-clock improvements in the GPU design offer performance improvements of anywhere from 5% in Metro 2033 to 13% in Dirt 2 based on NVIDIA's own testing.
According to NVIDIA's chart above, the GTX 580 will be anywhere from 15-30% faster than the GTX 480 thanks to the combination of faster clocks, more processing cores and architectural changes. We'll have to see how these claims work out in our testing of course.
While I have no problem with NVIDIA protecting their hardware in this way I am curious to see how long it will take for a "real-world" application to hit this current draw limits in some way and initiate these performance drops on its own. FurMark doesn't really do anything special to the GPU - it simply attempts to work as much of the GPU core as possible all at one time. In many people's view, this is only testing the top efficiency possible out of the GPU - something that software developers should really be striving for anyway.