GeForce 9600 GT - what is it?
GeForce 9600 GT 512?
NVIDIA's current DX10 line-up covers the major price-points with various SKUs. However, on closer inspection, there exists a performance and, more importantly, price-gap between the GeForce 8600 GTS (£85) and GeForce 8800 GT 256/512 (£129+) SKUs that's currently bridged by the bastardised GPU that is the GeForce 8800 GS.Adding in another full-flavour GPU between the two would mean liberally taking from the GeForce 8800 GT(S) architecture, chopping bits off in the name of cost, and that's exactly what NVIDIA has done with the GeForce 9600 GT.
Now, knowing this, the G94-based 9600 GT isn't a series' evolution, as indicated by the nomenclature. Rather, NVIDIA felt that it would, well, sound better than GeForce 8700 (desktop).
Our NVIDIA-comparison table tells all...
| Graphics cards | NVIDIA GeForce 9600 GT | NVIDIA GeForce 8800 GT 512 | NVIDIA GeForce 88800 GTS 512 | NVIDIA GeForce 8600 GTS 256 |
|---|---|---|---|---|
| API | DX10 (SM4.0) | |||
| Manufacturing process | TSMC, 65nm | TSMC, 65nm | TSMC, 65nm | TSMC, 80nm |
| Transistors | 505M | 754M | 754M | 289M |
| PCIe | PCIe 2.0 | PCIe 1.1 | ||
| GPU clock | 650MHz | 600MHz | 650MHz | 675MHz |
| Shader clock | 1,625MHz | 1,500MHz | 1,625MHz | 1,450MHz |
| Memory clock (effective) | 1,800MHz | 1800MHz | 1,940MHz | 2,000MHz |
| Memory interface, size, and implementation | 256-bit, 512MiB, GDDR3 | 256-bit, 512MiB, GDDR3 | 256-bit, 512MiB, GDDR3 | 128-bit, 256MiB, GDDR3 |
| Memory bandwidth | 57.60GB/sec | 57.60GB/sec | 62.08GB/sec | 32GB/sec |
| Stream processors | 64 FP32 scalar | 112 FP32 scalar | 128 FP32 scalar | 32 FP32 scalar |
| Peak GFLOP/s | 312 | 504 | 624 | 139.2 |
| ROPs | 16 | 16 | 16 | 8 |
| Peak GTexel/s (bilinear) | 20.8 | 33.6 | 41.6 | 10.8 |
| Hardware-assisted video-decoding engine | NVIDIA's PureVideo HD - partial VC-1 decode and full H.264 decode | |||
| Display outputs (GPU) | DL-DVI (HDCP), HDMI, DisplayPort, mini-DIN | DL-DVI (HDCP), HDMI, mini-DIN | DL-DVI (HDCP), mini-DIN | |
| Reference cooler | single-slot | single-slot | dual-slot | single-slot |
| Retail price (default-clocked model) | £119 | £149 | £199 | £85 |
What you see is that the GeForce 9600 GT (G94) is far more like the higher-priced GeForce 8800 GT and GTS (G92) SKUs than the GeForce 8600 GTS that it effectively replaces as mid-range GPU, with the latter moving on down to the budget sector.
Interestingly, the G94 shares the same manufacturing process and core speed as the G92 GTS. It, however, packs in half the shader processors - 64 vs. 128, but clocked in at GTS 512 speeds - affirming that NVIDIA is leveraging higher-end technology for a mid-range SKU. The shader-blocks for the G94 drop to four, down from the seven and eight present on the G92 GT and GTS GPUs, respectively, making it cheaper to produce.
The render back-ends remain at a GT-matching 16 and the transistor count drops to 505M, keeps in line with the exercise of making mainstream GeForce 8800 more affordable. Eight texture processors per block remain intact, as per the G84/92 and double the G80-based GPUs.
Memory interface (256-bit) and size (512MiB) are identical to the G92 GT's, though, as is the provision for a single-slot-taking cooler.
NVIDIA's PureVideo HD engine is carried over, naturally, helping to hardware-accelerate high-definition content decode, including support for the computationally-expensive H.264's CABAC (context-adaptive binary arithmetic coding) decode. We note that Huffman Bitsream processing is still missing from VC-1 decoding, though.
The feature-list includes HDMI. Its implementation still isn't as elegant as ATI's, which is UAA (universal audio architecture) compliant and routed through the motherboard. Here, users will still need to connect sound from their motherboard's S/PDIF header to the card via a supplied cable.
GeForce 9600 GT boasts one new feature - GPU-integrated support for DisplayPort, which we first saw on ATI's Radeon HD 3650. Innate DisplayPort provision is good but we won't see the standard adopted by many of NVIDIA's card partners until more actual displays crop up with the requisite connector support. Dell's 30in 3008 is one of only a handful that currently supports it, we note.
NVIDIA mentions a power-draw of 95W for the reference card. That should infer a whisper-quiet cooler, even when the GPU is placed under load
As a pure specification-to-specification comparison against the GeForce 8600 GTS, the 9600 GT is fundamentally better in every respect. Think of it as half a GeForce 8800 GTS in terms of rendering but a full GeForce 8800 GT with respect to memory bandwidth. Finger-in-the-air predictions indicate that performance should be around 75 per cent GeForce 8800 GT's - it has a significantly lower GFLOP rate, due to the lopping off of stream processors, but matching memory bandwidth.
Initial pricing for default-clocked models will be around £119 and we expect NVIDIA's partners to release a slew of pre-overclocked models, too. The pricing for a stock model is such that it matches a Radeon HD 3850 512MiB's and should come in at around £20 cheaper than a Radeon HD 3870.
Update: 20/02/08 - AMD has dropped prices for its Radeon HD 3800-series cards with immediate effect. The Radeon HD 3850 256MiB can now be purchased for £90, the Radeon 3870 512MiB for £125, and Radeon HD 3870 X2 for £235, all including VAT.
Summary
NVIDIA's effectively bridged the price/performance gap between its GeForce 8600 GTS and GeForce 8800 GT GPUs. The GeForce 9600 GT's architecture represents the only logical way of doing so, by harnessing a higher-end architecture and reducing transistor counts, and by extension cost, by removing costly stream processors.ForceWare 174.12 drivers
NVIDIA's ForceWare 174.12 drivers, used to test this card, introduce new features for the PureVideo HD engine.NVIDIA is adding dynamic contrast and colour enhancements, run via the GPU's shaders, for better-looking images, it says. We'll be comparing the output to ATI's Radeon HD 3800-series soon.
Further, the driver supports dual-stream decode, which hardware-accelerates the decoding of two video streams running concurrently. There are only few instances where we can see this being useful - picture-in-picture commentary, for example.
Lastly, Microsoft Windows' Vista's Aero interface has traditionally been disabled when decoding and displaying high-def. content. The new driver removes this limitation and gets rid of that annoying 'Windows Vista is running in basic mode' alert. Again, having Aero available during playback isn't a make-or-break deal, sure, but it makes for an incrementally better experience.
