NVIDIA G73
G73 is NVIDIA's new mid-range graphics ASIC. Largely engineered as half a G71, the 125mm² G73 will power a set of products to replace a wide range of GeForce 6-series products. The chip it replaces, NV43, is the best selling chip in NVIDIA's history, therefore G73 has a lot to live up to. GeForce 6600 GT is still a strong product for the price.A 128-bit memory interface still makes sense for mass-market mid-range products, making a GPU cheaper to produce. So it's no surprise to see G73 sport such an interface to its main board memory, via 2 internal memory channels. Smaller than the 149mm² ATI RV530, at decent clocks G73 has the on-paper ability to smack the top RV530 SKU around without much problem, in a performance sense at least.
I mention that G73 is half a G71 in essence, and the 5/12/8 VP/FP/ROP configuration proves that's the case, with an extra vertex processor of course. G73 only contains a single dual-link DVI TMDS, which is somewhat a shame since the TMDS is tiny in silicon and easy enough to implement. Driving two very high-end displays on a mid-range product might seem somewhat mad, but the configuration is more popular than you'd think, especially in non-obvious markets.
So while the top-end G71-based products might seem like minor updates compared to NVIDIA's own GeForce 7800 GTX and GTX 512, or something like ATI's Radeon X1900 XT or XTX, the first G73-based product is a significant step forward in the mid-range space. And that's compared to both the older NVIDIA GeForce 6600 GT and ATI's X1600 XT. Here's the formal spec to get the basics down to show you why that's the case.
| NVIDIA G73 GPU Properties | ||
|---|---|---|
| GPU | NVIDIA G73 | |
| Process and Fabricator | 90nm @ TSMC, 90GT w/ low-k | |
| Die Size | 125mm² | |
| Transistor Count | ||
| DirectX Shader Model | Shader Model 3.0 | |
| Basic Configuration (VP/FP/ROP) | 5/12/8 | |
| Vertex Shader Info | VS3.0 5D FP32, co-issue MADD, branch, tex |
|
| Fragment Processor Info | PS3.0 4D FP32, dual and co-issue MADD/MADD, branch, tex |
|
| ROP Info | 4x FX8/FX16 MSAA (2 subsamples/cycle) 2x Z-only rate, 2x colour-only rate FP/FX blender (inc. FP16) |
|
| Texture processing | 24 FP32 address units, 24 samplers Bi/Tri/Aniso (128-tap), FP16 filtering |
|
| Memory Interface | 128-bit, 2 memory channels GDDR->GDDR3 |
|
| GPU | ||
| Display output | 1x dual-link DVI TMDS, 1x single-link DVI TMDS, NVIDIA PureVidio | |
So compared to NV43, G73 sports another quad of fragment processors, with those processors the updated versions from G71/G70. Compared to the units found in NV43 and NV40/45, they can issue twice the MADD instructions per cycle and contain updated texturing performance from the tied samplers, along with a lower branching penalty due to how they're driven.
The chip can also output twice the finished pixels per cycle, and the ROPs are wider allowing them the ability to do more intermediate work per cycle, before the final colour write and display on your screen. 90nm also lets the chip clock even higher than before, so the list of improved extras get a stronger core clock rate to boot.
Like G71 and G70, G73 only falters versus the competition in terms of pixel output quality. NVIDIA's mid-range now gets transparency antialiasing, and MSAA can now be gamma corrected before scan-out from the framebuffer, but ATI's mid-range hardware has both those features, and the MSAA quality is higher at 6x (with a true 6x6 EER).
Like G71, FP16 blend performance is up, SLI AA resolve is done over the inter-card link, there's double ROP rates for Z and colour and the chip can now drive a dual-link display.
"Stop waffling, Rys", you say, "what about the theoretical rates of the new flagship mid-range product that's powered by G73?". I hear ya'. "What about GeForce 7600 GT!". Yes, yes, here you go.
Theoretical Performance
| Theoretical Rates for GeForce 7600 GT, versus 6600 GT | ||
|---|---|---|
| NVIDIA GeForce 7600 GT | NVIDIA GeForce 6600 GT | |
| Core Clock | 560MHz | 500MHz |
| Memory Clock | 700MHz | 500MHz |
| Pixel fillrate | 6.72G pixels/sec | 2.00G pixels/sec |
| Texture sampling rate | 26.8G samples/sec | 16.00G samples/sec |
| Z-only fillrate | 8.96G samples/sec | 4.00G samples/sec |
| Vertex transform rate | 0.70G tris/sec | 0.35G tris/sec |
| VP MADD issue rate | 2.80G instr/sec | 1.50G instr/sec |
| FP MADD issue rate | 13.4G instr/sec | 4.00G instr/sec |
| Memory bandwidth | 22.40 GB/sec | 16.00 GB/sec |
More than triple the FP MADD issue rate and pixel fill, and #60% more texel sampling ability, despite only a 60MHz increase in core clock, is where the bulk of the new performance will come from. Not forgetting VP rate, vertex processing performance doubles, memory bandwidth increases to feed it all as much as possible.
It outguns 6600 GT in a big way, and with 256MiB of memory will dispatch 6800 GS without issue, leaving the GeForce 7600 GT sat pretty in the middle of NVIDIA's desktop PCI Express product range.
Time for a look at the reference boards themselves, before we look at performance.
