GeForce7: New Graphics Architecture?

The new graphics architecture was quite expectedly named GeForce 7, and the graphics processor which is its embodiment is called G70. NVIDIA seems to imply that this is a perfectly new architecture rather than an improved GeForce 6. Is it really so? Let’s delve deeper into the matter.

First, the new graphics processor from NVIDIA is the most complex GPU for personal computers today, consisting of as many as 302 million transistors. For note: the NV40 consists of 222 million transistors, and the ATI R480 of only 160 million. This complexity is quite natural since the new GPU contains 24 pixel pipelines and 8 vertex processors against 16 and 6, respectively, in GPUs of the previous generation. Moreover, the G70 is not just an overgrown NV40. NVIDIA has considerably revised the architecture of pixel and vertex processors to improve their performance. More about that shortly, but now let’s have a look at the technical specification of the new GPU in comparison with previous-generation models:

	NVIDIA GeForce 7800 GTX	NVIDIA GeForce 6800 Ultra	ATI RADEON X850 XT Platinum Edition
Manufacturing technology	0.11 micron	0.13 micron	0.13 micron low-k
Number of transistors	302 mln.	222 mln.	160 mln.
Clock frequency	430MHz	400MHz	520MHz
Graphics memory controller	256bit GDDR3 SDRAM	256bit GDDR3 SDRAM	256bit GDDR3 SDRAM
Graphics memory clock frequency	1200 MHz	1100 MHz	1180 MHz
Memory bus peak bandwidth	34GB/s	32.8GB/s	33.4GB/s
Maximum graphics memory size	512MB	512MB	512MB
Interface	PCI Express x16	PCI Express x16	PCI Express x16
Pixel processors, pixel shaders
Shader model	3	3	2.x
Static loops and branching	yes	yes	yes
Dynamic loops and branching	yes	yes	no
Multiple Render Targets	yes	yes	yes
Floating-Point Render Target	yes	yes	yes
Maximum number of pixels per clock cycle	24	16	16
Maximum number of Z values per clock cycle	32	32	16
Number of texturing samples	16	16	16
Texture filtering algorithms	Bi-linear	Bi-linear	Bi-linear
	tri-linear	tri-linear	tri-linear
	anisotropic	anisotropic	anisotropic
	tri-linear + anisotropic	tri-linear + anisotropic	tri-linear + anisotropic
Maximum level of anisotropy	16x	16x	16x
Vertex processors, vertex shaders
Shader model	3	3	2.x
Number of vertex processors	8	6	6
Static loops and branching	yes	yes	yes
Dynamic loops and branching	yes	yes	no
Reading textures from the vertex shader	yes	yes	no
Tesselation	no	no	no
Full-Screen Anti-Aliasing
FSAA algorithms	Ordered-grid super-sampling	Ordered-grid super-sampling	Rotated-grid multi-sampling
	Rotated-grid multi-sampling	Rotated-grid multi-sampling	temporal anti-aliasing
	super-sampling + multi-sampling	super-sampling + multi-sampling
	transparent supersampling
Number of samples	2..8	2..8	2,4,6
Technologies increasing the efficiency of the memory bus bandwidth
Hidden Surface Removal (HSR)	yes	yes	yes
Texture, Z-buffer, frame buffer compression	yes	yes	yes
Fast Z-buffer clear	yes	yes	yes
Additional technologies
OpenEXR (HDR)	yes	yes	no
Videoprocessor	yes	yes	no

So, the G70-based graphics card is called GeForce 7800 GTX. This new device looks quite imposing, being technologically head above the GeForce 6800 Ultra, not to mention the RADEON X850 XT Platinum Edition. There’s only one parameter the ATI card is superior in ?the clock rate of the G70 is lower, having increased by only 30MHz above the previous-generation NV40 chip. This is of course the consequence of the high complexity of the chip, but there seem to be no problems with production. NVIDIA says graphics cards on the new GPU will be available at the day of the announcement. That’s good, recalling that the last-year announcement of the NV40 was in fact just a marketing event, the actual silicon being unavailable in shops.

The table above does not reveal the distinguishing features of the G70 which make it a truly new-gen solution. We will dwell on these features individually. At first sight, the G70 is similar to the NV40, save for the number of pixel and vertex processors:

The number of Raster Operation (ROP) units has remained the same. There 16 of them in the chip, and the number of texture samples per pass is still 16 or 32 when only Z values are sampled. In other words, the number 24 refers to the number of pixel processors only. So, the GeForce 7800 architecture generally resembles the GeForce 6800, but there are some considerable differences, too.