Product Listing

NVIDIA GeForce 7800 GTX (G70)

By Vijay Anand - 22 Jun 2005

302 Million Transistor GPU @ 430MHz

302 Million Transistor GPU @ 430MHz

If you thought the 222 million transistors in the GeForce 6800 Ultra was already mind boggling, then try 302 million transistors in the new GeForce 7800 GTX. Now that's a third more transistors than the best of theNV40, but that isn't surprising at all when you consider that the core processing engine has been enlarged with more pipelines and processing units as pointed out above. At 302 million transistors, this is the most complex mass commercial GPU out there and to intimidate some others, NVIDIA did some math for us to put things in better perspective. If you were to combine the CPU and graphics processing units of all the current generation consoles (XBOX, PS2 and Game Cube) and throw in an Athlon 64 FX-55 processor as well, the combined transistor count stands in at only 300.4 million. That's still shy of the 302 million transistors that the GeForce 7800 GTX boasts and a testament to its complex design and processing prowess.

This complex GPU is manufactured at TSMC's fab with a 0.11-micron process manufacturing technology. Some of NVIDIA's existing GPUs are fabed there as well with the same process technology such as the GeForce Go 6800 GPUs. Hence, we predict it won't take TSMC and NVIDIA long to refine the process for the GeForce 7800 GTX GPUs before we see volume quantities of them fast. Perhaps for the long run, it would have been great if NVIDIA already started adopting a 90nm process technology such as the competitor's upcoming VPU, but that would also mean a longer cycle time before getting settled and attuned to that process technology. There are pros and cons to adopting such cutting edge silicon manufacturing process technologies, but looking back at how NVIDIA met with issues during their bold moves in the NV30 days, we guess it's better for them to churn out the GPUs in the needed quantity now at 0.11-micron process while slowly preparing for the 90nm transition in the near future and perhaps re-spin them when ready. More likely, we foresee a variant of the G70 design in future to adopt the newer process technology. In any case, the current decision by NVIDIA is a safer and surer move if you ask us. Here's a comparison of the GeForce 6800 Ultra die and that of the GeForce 7800 GTX:-

The GeForce 6800 Ultra's die measures 16.5 x 18.5mm for a die size of 305 millimeter-square. Using a Flip-Chip BGA packaging, the entire chip is 40 x 40mm in dimensions.

The all-new GeForce 7800 GTX has an even larger die at 18.5 x 18.5mm for a die size of 342 millimeter-square - huge even at 0.11-micron process. The packaging however, seems to have shrunk a little and is 37.5 x 37.5mm. For a clearer photo without the thermal interface material, check our <a href="http://www.hardwarezone.com.sg/articles/view.php?cid=3&id=1603">snapshot preview</a> .

As for the GPU's operating clock speeds, it hasn't changed much from the 400MHz GeForce 6800 Ultra. The new GeForce 7800 GTX has a clock speed of 275MHz in 2D mode and when entering any 3D API, it throttles up to 430MHz. At this point, some of you might be thinking it's a measly 30MHz improvement. Well don't forget that this GPU is a massively parallel internally with far more processing engines and pipelines to accomplish tasks faster than its predecessor by a good degree even if there were no clock speed improvements. Historical advances in GPUs have showcased this many times (for example, NV40 versus NV35 versus NV30 and so on) and it's no exception this time either.

On another note about the G70 architecture, Sony's upcoming PlayStation 3 will be using NVIDIA's RSX processor that is based on the G70 architecture. A few months ago, some NVIDIA representatives mentioned it would be as powerful as the GeForce 7800 GTX. However during the G70 launch briefing, NVIDIA's CEO - Jen-Hsun Huang, made a statement that the RSX graphics synthesizer is using the G70 architecture, but is one generation ahead. That could either mean more features, or more parallel processing pipelines, or a combination of both to give the RSX the needed power to see through it's useful lifetime in the future when it's released. Whatever the case, this sounded a lot more logical to us. After all, the GeForce3 was a fresh product when the XBOX was launched and it had a GPU more powerful than the desktop GeForce3 products.


1.2GHz Graphics-DDR3 Memory On 256-bits Memory Interface

Even Memory bandwidth hasn't improved much over existing generation of hardware simply because good shader codes taking advantage of true programmable shading, which can produce fantastic and realistic image quality as NVIDIA has showcased in the past using the third generation Unreal Engine demo. This reduces the memory bandwidth required if data were to be repeatedly written to the local frame buffer (multiple passes) as was the case with many complex texture effects executed in the past using DirectX 8 and early DirectX 9 standards (and hence such effects were used sparingly those days to prevent overwhelming the graphics card). The newer Microsoft Shader Model 3.0 supported on all GeForce 6 and 7 series GPUs and supported on DirectX 9.0c has opened the doors to vastly complex shader effects with easy programmability. This allows the graphics shader hardware to do complex calculations and branching (given that the code is written to take advantage of SM 3.0) to derive the final output, vastly cutting back on the write backs required to the frame buffer (which is a problem when more 'primitive' code paths are used to accomplish complex tasks). With almost all new graphics hardware to support SM 3.0, it's only a matter of time before more and more games are designed to take advantage of the more efficient code paths and flexibility, putting the GPU to good use without overly impacting on the memory bandwidth required.

This is the reason why the memory interface on the G70 architecture is still 256-bits wide and the GeForce 7800 GTX in its first incarnation ships with 256MB of memory. Of course as we mentioned in a recent review, game developers will push the limits of realism further requiring even larger buffers to compute all sort of effects in addition to the growing size of textures and maps. Hence more memory and bandwidth required in future is inevitable, but at least with SM 3.0 and more efficient processing and effects generation methods, it has prolonged the need for bigger frame buffers and higher memory throughput needs right now (which would have escalated hardware costs exorbitantly, as if they weren't costly enough already).

Join HWZ's Telegram channel here and catch all the latest tech news!
Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.