We encourage you to visit our GeForce GTX 680 page on NVIDIA.com, but we wanted to take the time to highlight why this new GPU should be the heart and soul of your game development platform. Let's look at some raw data first, comparing to the GTX 580 (short version: We've tripled the number of cores, dramatically increased fill rates, AND decreased power consumption):

GPU

Fermi Family

GTX 580

Kepler Family

GTX 680

Transistors
3.0 billion
3.54 billion
CUDA Cores
512
1536
Graphics Core Clock
772MHz
1006MHz
Shader Core Clock
1544MHz
n/a
GFLOPs
1581
3090
Texture Units
64
128
Texel fill-rate
49.4 Gigatexels/sec
128.8 Gigatexels/sec
Memory Clock
4008 MHz
6008MHz
Memory Bandwidth
192.4 GB/sec
192.26 GB/sec
Max # of Active Displays
2
4
TDP
244W
195W

 

The GTX 680 offers several new architectural features that can be leveraged in the years to come. Here are a few:

  • The GTX 680 continues the Graphics Processing Cluster (GPC) architecture of the Fermi series GPUs, using four GPCs and delivering 32 pixels per clock. However, the Kepler series has a next-generation Streaming Multiprocessor (SM) called SMX. Where the GTX 580 contained 16 SMs, the GTX 680 uses 8 SMXs, with substantial improvements in throughput both in terms of clock speed and bandwidth (in some cases as much as 6x higher).
  • The SMX units also feature the new PolyMorph Engine 2.0, the workhorse of DX11 tessellation, can yield as much as 30% more performance, at higher expansion factors, than our GTX 580. We expect more and more adoption of tessellation as consumer demand for visual content increases, and the GTX 680 is designed to meet, and exceed, those demands.
  • The faster 512Kb L2 Cache has had significant increases in throughput as well, up to 9x more with atomic operations to a single common address.
  • Bindless Textures allow the GTX 680 to use millions of textures in a single scene, and eliminates the need for binding tables. The added benefit of this feature is that CPU utilization is dramatically reduced. At this time, bindless textures is only exposed in OpenGL. But in the future it's possible this could be exposed in DirectX via NVAPI, or future versions of DirectX.

For a detailed description of all our new features, please read the NVIDIA GeForce GTX 680 Whitepaper.