AnandTech: NVIDIA's Tiny 90nm G71 and G73: GeForce 7900 and 7600 Debut

Introduction

Today marks the launch of NVIDIA's newest graphics cards: the 7900 GTX, 7900 GT and the 7600 GT. These cards are all based on an updated version of the original G70 design and offer higher performance for the dollar. Today we will see just how much faster the new NVIDIA flagship part is. But first let's take a look at what makes it different.

At the heart of this graphics launch is a die shrink. The functionality of the new parts NVIDIA is introducing is identical to that of the original G70 based lineup. Of course, to say that this is "just a die shrink" would be selling NVIDIA a little short here. In the future, if either NVIDIA or ATI decide to move to TSMC's newly introduced 80nm half-node process, all that would be involved is a simple lithographic shrink. Sure, things might get tweaked a little here and there, but the move from 90nm to 80nm doesn't involve any major change in the design rules. Moving from 110nm to 90nm requires NVIDIA to change quite a bit about their register transfer logic (RTL), update the layout of the IC, and verify that the new hardware works as intended.

The basic design rules used to build ICs must be updated between major process shrinks because the characteristics of silicon circuits change at smaller and smaller sizes. As transistors and wires get smaller, things like power density and leakage increase. Design tools often employ standard components tailored to a fab process, and sometimes it isn't possible to drop in a simple replacement that fits new design rules. These and other issues make it so that parts of the design and layout need to change in order to make sure signals get from one part of the chip to another intact and without interfering with anything else. Things like clock routing, power management, avoiding hot spots, and many other details must be painstakingly reworked.

In the process of reworking the hardware for a new process, a company must balance what they want from the chip with what they can afford. Yield of smaller and smaller hardware is increasingly affecting the RTL of a circuit, and even its high level design can play a part. Making decisions that affect speed and performance can negatively affect yield, die size, and power consumption. Conversely, maximizing yield, minimizing die size, and keeping power consumption low can negatively affect performance. It isn't enough to come up with a circuit that just works: an IC design must work efficiently. Not only has NVIDIA had the opportunity to further balance these characteristics in any way they see fit, but the rules for how this must be done have changed from the way it was done on 110nm.

After the design of the IC is updated, it still takes quite a bit of time to get from the engineers' desks to a desktop computer. After the first spin of the hardware comes back from the fab, it must be thoroughly tested. If any performance, power, or yield issues are noted from this first run, NVIDIA must tweak the design further until they get what they need. Throughout this entire process, NVIDIA must work very closely with TSMC in order to ensure that everything they are doing will work well with the new fab process. As microelectronic manufacturing technology progresses, fabless design houses will have to continue to work more and more closely with the manufacturers that produce their hardware in order to get the best balance of performance and yield.

We have made quite a case for the difficulty involved in making the switch to 90nm. So why go through all of this trouble? Let's take a look at the benefits NVIDIA is able to enjoy.