How does GeForce 4x0 line up
Here are GeForce GTX 480 and 470 laid out in table form.| Graphics cards | NVIDIA GeForce GTX 480 1,536MB | NVIDIA GeForce GTX 470 1,280MB | NVIDIA GeForce GTX 295 1,792MB | NVIDIA GeForce GTX 285 1,024MB | ATI Radeon HD 5970 2,048MB | ATI Radeon HD 5870 1,048MB | ATI Radeon HD 5850 1,024MB |
|---|---|---|---|---|---|---|---|
| General clock | 700MHz | 607MHz | 576MHz | 648MHz | 725MHz | 850MHz | 725MHz |
| Shader clock | 1,401MHz | 1,215MHz | 1,242MHz | 1,476MHz | 725MHz | 850MHz | 725MHz |
| Memory clock (effective) | 3,696MHz | 3,348MHz | 1,998MHz | 2,484MHz | 4,000MHz | 4,800MHz | 4,000MHz |
| Memory interface and size | 384-bit, 1,536MB GDDR5 | 320-bit, 1,280MB GDDR5 | 896-bit (2 x 448-bit), 1,792MB, GDDR3 | 512-bit, 1,024MB, GDDR3 | 512-bit (2 x 256-bit), 2,048MB | 256-bit, 1,024MB, GDDR5 | 256-bit, 1,024MB, GDDR5 |
| Memory bandwidth | 177.4GB/s | 133.9GB/s | 2 x 111.9GB/s | 159GB/s | 2 x 128GB/s | 153.6GB/s | 128GB/s |
| Manufacturing process | TSMC, 40nm | TSMC, 40nm | TSMC, 55nm | TSMC, 55nm | TSMC, 40nm | TSMC, 40nm | TSMC, 40nm |
| DirectX/ Shader Model | DX11, 5.0 | DX11, 5.0 | DX10, 4.0 | DX10, 4.0 | DX11, 5.0 | DX11, 5.0 | DX11, 5.0 |
| Vertex, fragment, geometry shading (shared) | 480 FP32 scalar ALUs, MADD + MUL | 448 FP32 scalar ALUs, MADD + MUL | 480 FP32 scalar ALUs, MADD + MUL | 240 FP32 scalar ALUs, MADD + MUL | 3,200 FP32 scalar ALUs, MADD + MUL | 1,600 FP32 scalar ALUs, MADD + MUL | 1,440 FP32 scalar ALUs, MADD + MUL |
| Single-precision GFLOPS (single-issue) | 1,345 | 1,088 | 1,192 |
708 |
4,176 | 2,720 |
2,088 |
| Texturing | 60ppc
bilinear 30ppc FP16 15ppc FP32 |
56ppc
bilinear 28ppc FP16 14ppc FP32 |
160ppc
bilinear 80ppc FP16 40ppc FP32 |
80ppc
bilinear 40ppc FP16 20ppc FP32 |
160ppc
bilinear 80ppc FP16 40ppc FP32 |
80ppc
bilinear 40ppc FP16 20ppc FP32 |
72ppc
bilinear 36ppc FP16 18ppc FP32 |
| ROPs | 48 |
40 |
56 |
32 |
64 | 32 |
32 |
| GPixels/s throughput | 33.6 |
24.28 |
32.26 |
20.74 |
46.4 |
27.2 | 23.2 |
| GTexel/s bilinear | 42 |
33.99 |
92.2 |
51.84 |
116 |
68 |
52.2 |
| Board power (max) | 250W | 215W | 289W | 183W | 294W | 188W | 170W |
| Multi-GPU | Three-way SLI | Three-way SLI | Two-way SLI | Three-way SLI | Two-way XFire | Four-way XFire | Four-way XFire |
| Board length | 10.5in | 9.5in | 10.5in | 10.5in | 12in | 11in | 9.5in |
| Connectors (native) | 2x
dual-link DVI Mini-HDMI |
2x
dual-link DVI Mini-HDMI |
2 x dual-link DVI, HDTV-out,HDMI | 2x
dual-link DVI HDTV-out |
2x
dual-link DVI Mini-DisplayPort |
2x
dual-link DVI HDMI, DisplayPort, |
2x
dual-link DVI HDMI, DisplayPort, |
| Etail price | £440+ | £299+ |
£350 |
£285 |
£550 | £310 | £225 |
Analysis: GeForce GTX 480
Most pundits expected GeForce GTX 480 to be composed of 512 cores. NVIDIA is disabling one SM unit (32 cores) and the associated texturing, L1 cache, and PolyMorph engine that goes with it. We don't know exactly why GTX 480 loses out an SM, but the logical reason has to do with ensuring a quality of yield that makes it economically feasible. Will we see a 512-core GTX card? The chances are it will come into being when yields have significantly improved at TSMC, and that kind of timeframe falls in line with the release of the professional part, Tesla.
Clocking in at 700MHz and 1,400MHz for the general and shader clocks, respectively, and knowing the number of texturing units and ROPs, the GeForce GTX 480's peak numbers aren't particularly impressive for a 3bn-transistor chip. The Radeon HD 5870 has a decided advantage in shading and texel filtering, but it gives way on the ROP-led pixel throughput. Twin-GPU Radeon HD 5970 wins every performance segment, we note.
Peak numbers don't take efficiency into account, and this is one factor that NVIDIA is banking upon with GTX 4x0, that is, sustaining high throughput at all times. Looking out from the ROPs to the frame-buffer and NVIDIA's 384-bit bus is fed by 1,536MB of GDDR5 memory that's clocked at a rather lowly 3.7GHz. The upshot is class-leading bandwidth for a single-GPU solution, due to the ample width, but we imagine that NVIDIA will increase the GDDR5 frequency over time.
Put it all together and you realise why NVIDIA has been somewhat conservative with the frequencies. The company has to maintain acceptable yields and thermals. Board power is rated to 250W - 33 per cent higher than the HD 5870 - and increasing frequencies will bring fundamental thermal obstacles to the fore.
Our best-guess pricing indicates that etailers will charge £440 for NVIDIA's finest GPU, probably more on launch day, pushing its price 40 per cent higher than Radeon HD 5870's.
Analysis: GeForce GTX 470
NVIDIA knows that GF100 needs to be distilled into a cheaper GPU from the get-go. GeForce GTX 470 loses an SM unit (32 cores), eight ROPs, and significant frequency for the core, shader, and memory. Back-of-the-envelope notes indicate that it will perform at around 65 per cent of the level of GTX 480, putting it, NVIDIA hopes, around the Radeon HD 5850/5870 in performance terms.
GeForce GTX 470 is still a hot beastie, based around the same 3bn-transistor GPU, and it pulls around 30W more than AMD's flagship single-GPU card, based on company-provided TDPs. It carries the same dual-link DVI and mini-HDMI outputs as the GTX 480; DisplayPort being conspicuous by its absence.
The peak numbers would give the victory, easily, to AMD's range of Radeon 5-series GPUs on a pound-for-pound basis. Benchmarks will show if this is true...
