NVIDIA’s GeForce 8800GTX
Specs

With the insane amount of speculation and rumors circulating around the web about the specs of this card, let’s take a look at what they really mean.

• DirectX 10 Compatible
• Shader Model 4.0 Support
• 575MHz Core Clock
• 768MB of GDDR3 memory at 900MHz
• 384-bit memory interface (256-bit+128-bit)
• 86GB/s memory bandwidth
• 128 unified shaders at 1350MHz
• Theoretical texture fill-rate at 38.4 billion pixels per second
• HDCP compliant
• Dual link DVI, VIVO, and HDTV outputs
• Dual-slot cooling

Straight off the bat it is evident that this video card is unlike any other. Breaking down these specs step-by-step we see that the 8800GTX is DirectX 10 compatible. This is significant purely for the fact that it is the first such card to the market. Should DirectX 10 drop in a big way sometime in the very near future, gamers wanting to live the DX10 experience to its utmost will have no choice but to get a G80 based NVIDIA card, as ATI’s DX10 compatible cards based on the R600 core are still a couple of months out. Among other features, DirectX 10 will have this lovely thing called “Epic’s Unreal 3 Engine uses it”. To us that is reason enough to be giddy. Next on the spec list is Shader Model 4.0 support. SM3.0 is not very old itself. First implemented fully with NVIDIA’s 6 series, SM3.0 offered considerable advancements over SM2.0. SM4.0 is quite simply the next step in the progression of shader models. No other card is known to support SM4.0, so if a game drops using SM4.0, the G80 cards will again be the only way to go.

The next two bullets on the list are kind of normal…. Wait… 768MB of memory? This is far from a commonly used memory amount. Why does it use 768MB of memory instead of 512MB or even 1024MB? We have some ideas, but we can’t know for sure. We speculate that the extra 256MB of memory on top of the 512MB will probably act as some sort of buffer for one of the unmentioned features of the card, the 128-bit HDR + 16AA capability. This feature might be what puts NVIDIA ahead of ATI in terms of image quality, but only time will tell on that one since there aren’t really any games that have code for 128-bit HDR. The whole memory bus width deal is kind of weird too. 256-bit+128-bit? Huh? We can only assume this is for some other newfangled feature on the G80, namely NVIDIA’s Quantum Physics Engine. The name itself is kind of confusing, not only because it just sounds complicated but also because the whole “Quantum Physics” title is already kind of being used by a real life science. Technical documents that have been percolating across the internets have NVIDIA claiming that QPE will be a direct competitor to AGEIA’s PhysX. This isn’t something we are entirely convinced will work out too well, not after our recent interview with AGEIA’s CEO about their processor and the PhysX SDK. Since we doubt NVIDIA went out and coded a new proprietary physics middleware by themselves, we have to think Havok has played a role with QPE. In terms of hardware, NVIDIA has apparently allowed several of the 128 unified shaders on the G80 to carry out physics calculations. The debate over whether or not a specifically designed chip like the PhysX is more effective at physics calculations than a few shaders on a GPU has been mulled over for months already. We can tell you that if history is any indication, taking a page out of the GPU book will show that a dedicated chip is usually the way to go. We’re not taking any sides though, it is WAY too soon to speculate about which will come out on top.

The next bullet down concerns the memory bandwidth of the G80. While the listed 86GB/s (really 86.4GB/s but let’s not argue semantics) of bandwidth exceeds that of the predecessor by nearly 10GB/s, that is entirely the result of the significantly higher operating frequency than anything else. The 8800GTX is still using the tried and true GDDR3 memory modules, even though GDDR4 is already out and being used with impressive results on ATI’s flagship Radeon X1950XTX. Although the bandwidth on this X1950XTX is “only” 64GB/s, bandwidth isn’t everything. Easily the coolest spec on the G80 is that there are 128 unified shaders. A unified shader is one that is capable of executing both vertex and pixel shading instructions. Whereas in the past the number of pixel shaders and vertex shaders were carefully balanced and differed from card to card, unified shaders allow manufacturers to use more where more are needed. Imagine having a set of 10 dudes on either side of a wall. All of a sudden, a whole bunch of work is dumped on one side of the wall that will take the 10 dudes 2 hours to complete. Meanwhile, the other 10 dudes are just chilling on the other side of the wall, not only because there is a wall in the way, but because they speak a different language than the other 10 dudes so they wouldn’t be able to do the work that they are doing in the first place. Now imagine all 20 dudes are bunched together and all speak a singular hybrid language. That same work load is dumped on the 20 dudes and they are able to knock it out in 1 hour. If you just read all that and it made sense to you, congratulations. That, we believe, is the premise behind unified shaders. That’s enough about specs and features we think. Sure we didn’t go over HDCP, which is kind of a mixed blessing, but we’re anxious to get into the benchmarks. Aren’t you?