NVIDIA's GeForce 7600 GT and 7900 series GPUs
...plus ATI's Radeon X1800 GTO
by Scott Wasson — 8:00 AM on March 9, 2006

THE STRETCH FROM late 2005 into early '06 has been a remarkably busy time in the world of PC graphics hardware. Most of it has been ATI's doing. Their CrossFire dual GPU technology was tardy, but it showed up to class just before the end of September. Shortly thereafter, ATI finally delivered its much-delayed Radeon X1800 GPU as part of a top-to-bottom lineup of new graphics chips, all among the first GPUs to be built on a 90nm fab process. Among other things, the Radeon X1800 XT put the company back on top in the high-end performance sweeps, however briefly. The welts had barely faded from the whooping NVIDIA laid on the Radeon X1800 XT with its GeForce 7800 GTX 512 when ATI countered with its new high-end GPU (right on schedule this time), the Radeon X1900. Not only was the Radeon X1900 XT the new fastest card on the block, but it got some help on the multi-GPU last week in the form of a new dual sixteen-lane PCI Express chipset, the CrossFire Express 3200.

Phew.

Now it's time for another round of new stuff, and the competition is tighter than ever. Today, NVIDIA is unveiling two GPUs powering three different flavors of video cards. Say hello to the GeForce 7600 and 7900 series, new mid-range and high-end graphics cards ranging from $199 to, well, very expensive. These introductions mark NVIDIA's transition in earnest to 90nm fab process technology, a place where ATI has been for a number of months.

Speaking of ATI, the folks there couldn't stand by and watch new GeForces arrive without doing something, so they've decided to unveil a new $249 Radeon card and rework their product lineup to include a range of more affordable options. We've tested the Radeon X1800 GTO alongside the new NVIDIAs.

Oh, yes. And we have the juicy details on the King Kong of PC graphics, quad SLI, where not one, not two, but four GeForce 7900 GPUs can pull together to squash all competition and raid your Swiss bank account.

The G71 graphics processor
Code-named G71, NVIDIA's new high-end GPU is the successor to the G70 graphics processor that powers the GeForce 7800 lineup. The G71 is very similar to the G70 in terms of basic specs, but it's produced using a new, finer 90nm fabrication process that squeezes a similar number of transistors into less space. Like its precursor, the G71 has six pixel shader "quads," for a total of 24 pixel shader units, plus eight vertex shaders and 16 raster operators (ROPs).

The G71 is more than just a die shrink, however. Get this: the transistor count estimate is actually down from the G70's 302 million to 278 million for the G71. Why so? NVIDIA says it has replumbed the G71's internal pipelines throughout the chip, making them shorter, because those longer pipes and extra transistors aren't needed to help the G71 achieve acceptable clock speeds—the faster-switching transistors of the 90nm process will suffice for that purpose. How's that for confidence?

Shorter pipelines typically make for higher clock-for-clock performance, and we may see some of that from G71, but this isn't a radical change. I wouldn't expect anything revolutionary on that front.

A couple other alterations to G71 will have a more tangible impact. NVIDIA's design team has modified the SLI logic so the GPU can pass sample data for the SLI antialiasing mode over an SLI link, instead of via PCI Express like on current chips. This arrangement ought to bring higher performance with SLI antialiasing, perhaps to the point where it becomes a truly useful load-balancing method for multi-GPU setups.

Also, the G71 includes two built-in dual-link TMDS transmitters, so GeForce 7900 cards can power a pair of high-def digital displays without the need for an external TDMS transmitter. ATI's Radeon X1800 and X1900 cards have this feature already, and the G71 does well to follow suit.

Beyond those tweaks, we couldn't shake loose much more info on changes in the G71 compared to the G70. NVIDIA confirms it hasn't changed the amount of register space on the GPU, and although there have been some modifications to buffer sizes and the like, the company prefers not to get into those gory details.


The G71 GPU

The final product of this die shrink and handful of revisions is a GPU that fits into 196 mm2, way down from the G70, which I measured (somewhat shakily) at 333 mm2. If things go as they should and the 90nm process doesn't lead to excessive leakage, the G71 should run at lower voltages, consume less power, and produce less heat than the G70.

Those of you who follow these things closely will recognize that this combination of elements could present an intriuging challenge to ATI's high-end part—a possibility that NVIDIA is crowing about every chance it gets. At 110nm, the G70 was pretty competitive with ATI's R580 GPU (a.k.a the Radeon X1900) in terms of die size and power consumption, with only slightly lower overall performance. The changes put the G71 at just under two thirds the size of the R580; although it's also manufactured on a 90nm fab process, the R580's die size is roughly 315 mm2. If the G71 handles well enough, it could end up offering equivalent performance to the R580 in a fraction of the physical and thermal footprint.