The NVIDIA GeForce GTX 680 Tech Report

In an industry where 6 month product cycles are considered a norm, the new NVIDIA Kepler architecture took a while to get here. The last NVIDIA microarchitecture, Fermi was launched about 18 months ago, or roughly 3 product cycles ago. In the preceding 6 months, NVIDIA had only launched two new Fermi-based graphics cards - the GeForce GTX 560 Ti Limited Edition (with 448 CUDA cores), and the budget-level GeForce 510. As such, the launch of its latest Kepler architecture is truly a breath of fresh air.

As is NVIDIA's habit, they mark the launch of a new microarchitecture with their flagship model of the family. So here we are, with the tech report on NVIDIA's latest graphics behemoth, the NVIDIA GeForce GTX 680. Let's take a look.

Key Features

The NVIDIA GeForce GTX 680 will feature 3 key areas of improvements, namely better performance, a smoother display and a richer experience. How NVIDIA delivers that is via a number of new features and improvements in features already present in Fermi. We will go into the details in coming slides.

Let's start with how NVIDIA ensures that the GeForce GTX 680 delivers much better performance than their previous GeForce GTX 580 graphics card.

The New SMX

In the previous Fermi architecture, the basic building block is the streaming multiprocessor unit, otherwise known as the SM unit. It consists of control logic with 32 CUDA cores. In the new Kepler architecture, NVIDIA introduces an enhanced SM unit, which they now call SMX. The difference between the two is significant - NVIDIA reduced the amount of control logic and increased the number of CUDA cores 6-fold to 192. This increases the efficiency (performance per watt) of the Kepler SMX unit to twice that of the SM unit in Fermi.

However, the loss of control logic comes at a price - NVIDIA moved some of the scheduling logic to the compiler. This means Kepler pushes some of the work to the compiler and therefore, the CPU. With modern multi-core CPUs so fast and virtually idling in most applications, this extra work shouldn't make much of a difference, and makes a lot of sense for NVIDIA.

The NVIDIA GeForce GTX 680 will have eight of these SMX units, giving it a total of 1536 CUDA cores. The other difference in the Kepler architecture is that the CUDA cores no longer run at twice the speed of the graphics clock - they now run at the same speed. However, NVIDIA is able to raise the graphics clock from 700-800 MHz to 1 GHz.

Support Tech ARP!

If you like our work, you can help support out work by visiting our sponsors, participate in the Tech ARP Forums, or even donate to our fund. Any help you can render is greatly appreciated!