Core changes
NV31 core
GeForce FX 5600 Ultra is based on NVIDIA’s NV31 core, a derivative of the NV30 architecture that was first announced last year. NV31 is composed of approximately 80 million transistors, considerably less than the 130 million transistors in NVIDIA’s GeForce FX 5900. With fewer transistors NVIDIA is able to trim manufacturing costs but at the expense of chip complexity. While the GeForce FX 5900 has four pixel pipelines with two texture units per pixel pipeline (4x2), GeForce FX 5600 has four pixel pipelines with only one texture unit per pixel pipeline. This effectively gives the GeForce FX 5600 half the texturing performance of GeForce FX 5900, assuming equal clock speeds.
One interesting note about the GeForce FX 5600’s transistor count however is that it has been increased over GeForce4: roughly 65 million in GeForce4 to GeForce FX 5600’s 80 million. With more transistors, the increased production cost would normally cut into NVIDIA’s profit margins. However, since GeForce FX 5600 is built on a smaller, 0.13-micron manufacturing process, NVIDIA should be able to yield more chips per wafer, assuming good yields of course. This change should help offset the added cost of incorporating more transistors.
NVIDIA has another trick up its sleeve for the GeForce FX 5600 Ultra in particular. NVIDIA has replaced the packaging on its original design to flip chip on the Ultra “2.0” for its improved signal quality at higher clock frequencies. This allows NVIDIA to hit higher clock speeds than they were able to achieve previously, core clock frequency has been increased from 350MHz to 400MHz on the revised Ultra core. This is the same clock speed as ATI’s RADEON 9600 PRO. NVIDIA has also bumped the memory clock up to 800MHz.
With these changes, let’s see how the mainstream segment now shapes up:
| Mainstream Graphics Feature Comparison |
| GPU |
Core Clock (MHz) |
Memory Clock (MHz) |
Peak fill rate (Mpixels/sec) |
Peak fill rate (Mtexels/sec) |
Peak Memory Bandwidth (GB/sec) |
MSRP |
| GeForce4 Ti 4200-8X |
250 |
512 |
1000 |
2000 |
8.2 |
$149 |
| GeForce FX 5600 Ultra |
400 |
800 |
1600 |
1600 |
12.8 |
$199 |
| GeForce FX 5600 |
325 |
550 |
1300 |
1300 |
8.8 |
$149 |
| RADEON 9600 PRO |
400 |
600 |
1600 |
1600 |
9.6 |
$199 |
| RADEON 9500 PRO |
275 |
540 |
2200 |
2200 |
8.6 |
$219 |
| RADEON 9600 |
325 |
400 |
1300 |
1300 |
6.4 |
$149 |
| RADEON 9500 |
275 |
540 |
1100 |
1100 |
8.6 |
$149 |
Notes
With its 400MHz core clock, GeForce FX 5600 Ultra matches RADEON 9600 PRO in fill rate. This is important for many of today’s games, which tend to rely on texel fill rate. Thanks to its 8x1 architecture, RADEON 9500 PRO is still the king of the hill in fill rate, despite giving up over 100MHz to the FX 5600 Ultra and RADEON 9600 PRO.
Memory bandwidth on the other hand goes to GeForce FX 5600 Ultra. At 12.8GB/sec, GeForce FX 5600 Ultra holds a 3.2GB/sec advantage over RADEON 9600 PRO; this should give GeForce FX 5600 Ultra a performance advantage at high resolutions, and will also play dividends in situations where anti-aliasing is enabled. Of course, this is all on paper, in the real world other factors will come into play like ATI’s HYPERZ III and NVIDIA’s Intellisample technology, both of which allow the card’s to make more efficient use of their available bandwidth. Driver performance also plays a role.
As far as we’re aware, 350MHz GeForce FX 5600 Ultra cards never made it to retail from any of NVIDIA’s board partners. These cores will instead become 325MHz GeForce FX 5600 cards. If you do come across a 350MHz GeForce FX 5600 Ultra, you should return it immediately.