EVGA GeForce 7900 GTX Preview - PAGE 2Tom Karpik - Thursday, March 9th, 2006
By now, most enthusiasts will be very familiar with the GeForce 7800 GTX's G70 architecture. Purported to be a complete ground-up redesign, the G70's key design goals were the following:
- Parallelism
- Shader arithmetic density
- High dynamic range lighting
- Power efficiency
This time around, it seems that NVIDIA has expanded the focus to encompass another rising trend -- high-resolution gaming, or as NVIDIA is calling it, X-HD Gaming ("Extreme High-Definition Gaming"). With the 7900 GTX, NVIDIA is pushing 2048x1536 and widescreen gaming more than ever. They are obviously confident that this new beast has the horsepower to drive gaming at such resolutions.

But just what drives this beast? Both the GeForce 7900 GTX and 7600 GT brandish new GPUs, which bring with them a few architectural improvements over the current G70 -- as well as massive increases in clock speeds. Let's compare the 7900 GTX to the 7800 GTX 512, 7800 GTX, and ATI's Radeon X1900 XTX:
| |
7600 GT |
7800 GTX |
7800 GTX 512 |
X1900 XTX |
7900 GTX |
| Architecture |
G7x |
G70 |
G70 |
R580 |
G71 |
| Manufacturing Process |
90 nm |
110 nm |
110 nm |
90 nm |
90 nm |
| Transistor Count |
177M |
302M |
302M |
384M |
278M |
| Pixel Pipelines |
12 |
24 |
24 |
48 |
24 |
| Vertex Pipelines |
5 |
8 |
8 |
8 |
8 |
| Render Outputs |
8 |
16 |
16 |
16 |
16 |
| Core Clock (MHz) |
560 |
430 |
550 |
650 |
650 |
| Fillrate (Gigapixels) |
4.48 |
6.88 |
8.8 |
10.4 |
10.4 |
| Memory Clock (MHz) |
1400 |
1200 |
1700 |
1550 |
1600 |
| Memory Interface |
128-bit |
256-bit |
256-bit |
256-bit |
256-bit |
| Memory Bandwidth (GB/s) |
22.40 |
38.40 |
54.40 |
49.60 |
51.20 |
| Memory Size (MB) |
256 |
256 |
512 |
512 |
512 |
Holy moly! The GeForce 7900 GTX and Radeon X1900 XTX look more closely related to each other than the 7900 GTX does to its own older brother. Coincidence? -- We think not. The X1900 XTX is within a hair of being a identical to the 7900 GX specs-wise -- its only disadvantage is a 50 MHz lag in memory clock speed. It does have a pixel shader unit advantage of 100%, but this raw number has not proven to be the end-all of pixel shader performance comparisons. Given these facts, it is obvious that the competition is going to be between implementations and drivers. Never have I seen two completely different cards from different manufacturers spec'ed so identically.
Looking at the raw design specifications, the 7900 GTX's GPU (code-named "G71") merely seems to be a die-shrinked G70 on steroids, but when you do a double-take, you realize one important fact -- the G71 is missing 24 million transistors all while retaining the same number of pixel shader units, vertex shader units, and ROPs. The jury is still out on this one as far as I know, but I am free to speculate, aren't I?
When the G70 was released, the enthusiast community erupted in talks of the G70 possibly being a 32 pixel pipeline part, with 8 of the pipelines disabled because yields were supposedly not efficient enough to produce 32-pipeline GPUs in mass quantities. The idea was that NVIDIA would release a 32-pipeline 7800-series card once those yields improved -- or maybe they were merely selling the fully-functional G70s to Sony for use in the Playstation 3? No one really knew. In any case, this "7800 GTX Ultra" was never released, and instead we have a 7900 GTX with 24 million missing transistors.
Maybe those yields really did get good enough that NVIDIA could use 100% of the GPU's silicon -- and now we have no more wasted space and improved production costs because those 8 extra pixel pipelines have been removed? I really don't know whether 3 million transistors per pipeline sounds accurate, but to me this is the most plausible explanation at the moment.
What is the actual die size of the 7900 GTX, and what does it mean for production costs? The die shrink to 90nm, as well as trimming the fat, results in a GPU that is only 196 mm^2 -- a substantial 41% reduction over the G70's 334 mm^2. NVIDIA's investors must be throwing a party, because the last time I checked, such a huge reduction in die size yields many, many more cores -- and that means more happy gamers opening up their wallets.
The 7600 GT also deserves a mention at this point. Obviously it is a reduced G71 design, with roughly have the various units as its bigger brother. Performance will be nowhere close to the 7900 GTX, but it will definitely serve the 6600 GT its butt on a silver platter. Do you want to know the real kicker, though? Well, do you? I'll tell you: Peak power consumption of the 7600 GT is a measly 67W. That's right -- low enough that a PCI-Express power connector is not required. It is the same as the 6600 GT it is meant to replace in this respect, and I'm sure this will be very well-received by the community. The 6600 GT has been a best bang-for-the-buck favourite, and now we'll have an even better bang for our buck, along with the virtues of the 7XXX architecture.
Enough of that. The question on everyone's minds must be "What does the G71/7900 GTX actually bring me?" The answer really is quite simple: Everything that the G70/7800 GTX brought, but with more pizzazz, and at the exact same average power consumption of 120W. This essentially means mind-blowing shader performance, HDR, PureVideo, decently cool and quiet operation, and NVIDIA's mature SLI technology, and the very same $499 USD MSRP that the GeForce 7800 GTX launched at -- among other things. The combination of a die shrink, transistor reduction, and undoubtedly a few other process tweaks have resulted in a GPU that delivers much higher clock speeds all while consuming the same amount of power as the previous generation.
The gamer's life just got better. I'll bet most of you want to see the board now. With no more delay, behold ...