Stacking up against the incumbent competition, plus future models
Graphics cards | NVIDIA GeForce GTX 295 1,792MB | NVIDIA GeForce GTX 285 1,024MB | NVIDIA GeForce GTX 280 1,024MB | NVIDIA GeForce GTX 275 896MB | NVIDIA GeForce GTX 260 896MB | ATI Radeon HD 5870 1,048MB | ATI Radeon HD 5850 1,024MB | AMD Radeon HD 4870 X2 2,048MB | AMD Radeon HD 4850 X2 2,048MB | AMD Radeon HD 4890 1,024MB | AMD Radeon HD 4870 512MB | AMD Radeon HD 4850 512MB |
---|---|---|---|---|---|---|---|---|---|---|---|---|
PCIe | PCIe 2.0 | |||||||||||
GPU(s) clock | 576MHz | 648MHz | 602MHz | 633MHz | 576MHz | 850MHz | 725MHz | 750MHz | 625MHz | 850MHz | 750MHz | 625MHz |
Shader clock | 1,242MHz | 1,476MHz | 1,296MHz | 1,404MHz | 1,242MHz | 850MHz | 725MHz | 750MHz | 625MHz | 850MHz | 750MHz | 625MHz |
Memory clock (effective) | 1,998MHz | 2,484MHz | 2,214MHz | 2,322MHz | 1,998MHz | 4,800MHz | 4,000MHz | 3,600MHz | 1,986MHz | 3,900MHz | 3,600MHz | 1,986MHz |
Memory interface and size | 896-bit (2 x 448-bit), 1,792MB, GDDR3 | 512-bit, 1,024MB, GDDR3 | 512-bit, 1,024MB, GDDR3 | 448-bit, 896MB, GDDR3 | 448-bit, 896MB, GDDR3 | 256-bit, 1,024MB, GDDR5 | 256-bit, 1,024MB, GDDR5 | 512-bit (2x 256-bit), 2,048MB, GDDR5 | 512-bit (2x 256-bit), 2,048MB, GDDR3 | 256-bit, 1,024MB, GDDR5 | 256-bit, 512MB, GDDR5 | 256-bit, 512MB, GDDR3 |
Memory bandwidth | 223.8GB/s |
159GB/s |
141.7GB/s |
130GB/s |
111.9GB/s |
153.6GB/s | 128GB/s |
230.4GB/s |
127.1GB/s |
124.8GB/s |
115.2GB/s | 63.5GB/s |
Manufacturing process | TSMC, 55nm | TSMC, 55nm | TSMC, 65nm | TSMC, 55nm | TSMC, 55nm | TSMC, 40nm | TSMC, 40nm | TSMC, 55nm | TSMC, 55nm | TSMC, 55nm | TSMC, 55nm | TSMC, 55nm |
DirectX/ Shader Model | DX10, 4.0 | DX10, 4.0 | DX10, 4.0 | DX10, 4.0 | DX10, 4.0 | DX11, 5.0 | DX11, 5.0 | DX10.1, 4.1 | DX10.1, 4.1 | DX10.1, 4.1 | DX10.1, 4.1 | DX10.1, 4.1 |
Vertex, fragment, geometry shading (shared) | 480 FP32 scalar ALUs, MADD dual-issue + MUL (unified) | 240 FP32 scalar ALUs, MADD dual-issue + MUL (unified) | 240 FP32 scalar ALUs, MADD dual-issue + MUL (unified) | 240 FP32 scalar ALUs, MADD dual-issue + MUL (unified) | 216 FP32 scalar ALUs, MADD dual-issue + MUL (unified) | 1,600 FP32 scalar ALUs, MADD dual-issue (unified)) | 1,440 FP32 scalar ALUs, MADD dual-issue (unified) | 1,600 FP32 scalar ALUs, MADD dual-issue (unified) | 1,600 FP32 scalar ALUs, MADD dual-issue (unified) | 800 FP32 scalar ALUs, MADD dual-issue (unified) | 800 FP32 scalar ALUs, MADD dual-issue (unified) | 800 FP32 scalar ALUs, MADD dual-issue (unified) |
Data sampling and filtering | 160ppc address and 160ppc bilinear INT8/80ppc FP16 filtering, max 16xAF | 80ppc address and 80ppc bilinear INT8/40ppc FP16 filtering, max 16xAF | 80ppc address and 80ppc bilinear INT8/40ppc FP16 filtering, max 16xAF | 72ppc address and 72ppc bilinear INT8/36ppc FP16 filtering, max 16xAF | 72ppc address and 72ppc bilinear INT8/36ppc FP16 filtering, max 16xAF | 80ppc address and 80ppc bilinear INT8/40ppc FP16 filtering, max 16xAF | 72ppc address and 72ppc bilinear INT8/36ppc FP16 filtering, max 16xAF | 80ppc address and 80ppc bilinear INT8/40ppc FP16 filtering, max 16xAF | 80ppc address and 80ppc bilinear INT8/40ppc FP16 filtering, max 16xAF | 40ppc address and 40ppc bilinear INT8/20ppc FP16 filtering, max 16xAF | 40ppc address and 40ppc bilinear INT8/20ppc FP16 filtering, max 16xAF | 40ppc address and 40ppc bilinear INT8/ 20ppc FP16 filtering, max 16xAF |
ROPs | 56 | 32 | 32 | 28 | 28 | 32 | 32 | 32 | 32 | 16 | 16 | 16 |
Peak GFLOPS | 1,788 | 1,063 | 933 | 1,011 | 805 | 2,720 |
2,088 | 2,400 | 2,000 | 1,360 | 1,200 | 1,000 |
Peak fillrate Gpixels/s | 32.256 | 20.736 | 19.264 | 17.724 | 16.128 | 27.2 |
23.2 |
24 | 20 | 13.6 | 12 | 10 |
Peak Gtexel/s (bilinear) | 92.2 | 51.84 | 48.16 | 45.576 | 41.472 | 68 |
52.2 |
60 | 50 | 34 | 30 | 25 |
Peak Gtexel/s (FP16, bilinear) | 46.1 | 25.92 | 24.09 | 22.788 | 20.736 | 34 |
26.1 | 30 | 25 | 17 | 15 | 12.5 |
Board power (max) | 289W | 183W | 236W | 219W |
182W |
188W | 170W |
286W | 220W |
190W |
160W | 110W |
Connectors (native) | 2 x dual-link DVI, HDTV-out,HDMI | 2x dual-link DVI HDTV-out |
2x dual-link DVI HDTV-out |
2x dual-link DVI HDTV-out |
2x dual-link DVI HDTV-out |
2x dual-link DVI HDMI, DisplayPort, HDTV-out, Eyefinity* |
2x dual-link DVI HDMI, DisplayPort, HDTV-out |
2x dual-link DVI HDMI, DisplayPort, HDTV-out |
2x dual-link DVI HDMI, DisplayPort, HDTV-out |
2x dual-link DVI HDMI, DisplayPort, HDTV-out |
2x dual-link DVI HDMI, DisplayPort, HDTV-out |
2x dual-link DVI HDMI, DisplayPort, HDTV-out |
Current price | £330 | £230 |
N/A | £160 |
£120 |
£299 |
£220 |
£250 |
£200 |
£125 |
£95 | £80 |
Analysis
NVIDIA's current high-end range on the left, running from fastest to slowest, and AMD's line-up on the right.
Take a minute to digest the voluminous information and you'll note that the Radeon HD 5870/50 are the fastest single-GPU cards around. The visceral shading power of the HD 5870 should, on paper, make it benchmark at around the same levels as the dual-GPU Radeon HD 4870 X2 and GeForce GTX 295. As noted earlier, performance may suffer in bandwidth-intensive situations.
*An aside, AMD has an interesting display-output technology on the new GPUs. Called Eyefinity, read more here.
Future parts
AMD will launch a dual-GPU version of Radeon HD 5870 at the end of the year. Known by the codename Hemlock, its arrival is slated to trumpet NVIDIA's DX11 'GT300' GPUs. It's made possible by the frugality of the single-GPU card, which is around Radeon HD 4890 levels. We imagine AMD's engineers will further fine-tune the ASIC for power regulation, leading to a sub-300W load for the dual-GPU model.
Cheaper derivatives of the 5-series cards will be released to coincide with Windows 7's availability. The cut-down parts, codenamed Juniper, will share much of the architecture, we imagine, but will be pared with respect to shaders, bandwidth, and power, much akin to how the Radeon 4700-series lines up against 48x0.
Appreciating that the HD 5870 is reasonably similar to the HD 4870 X2 in terms of statistics, why has AMD waited this long to introduce the design? Do Radeon 5K card have other tricks up their sleeve? The answer to both questions is inextricably linked with the feature-set, led by DX11.