- Jan 1, 2011
- 3,266
- 169
- 106
Back when Nvidia premiered their Geforce 400 series, comparisons with the Radeon HD 5000 series was mixed, but one thing was clear: Nvidia was better at tessellation. Nvidia had a tessellator on each SM, allowing them to draw many more triangles per cycle on the Geforce GTX 480 than the Radeon HD 5870. AMD's approach had some advantages -- tessellation on Nvidia graphics cards scaled worse on the low end as they removed SMs, while low end AMD graphics cards were chugging along with mostly the same tessellator. But for the high end, Nvidia ruled tessellation.
AMD resolved to improve tessellation performance, and the first significant improvement came with the Radeon HD 6900 series. AMD decoupled geometry functions so that multiple geometry engines, including multiple tessellators, was possible. The 6900 series had two geometry engines, though theoretically AMD could add as many geometry engines as they had die space for. That wasn't enough to dethrone Nvidia, but it was certainly an improvement.
For the Radeon HD 7000 series, AMD brought the dual geometry engine design down to the mid-range 7800 graphics cards. However, they did not increase the amount of geometry engines on the 7900 cards, leaving both tiers with the same geometry engines. This had the odd effect of the higher clocked Radeon HD 7870 beating out the 7950 and even sometimes the 7970 at tessellation benchmarks.
I was browsing through some tessellation benchmarks on Anandtech, and AMD seems to be very competitive on the DirectX 11 Detail Tessellation benchmark:
http://www.anandtech.com/bench/GPU12/423
But you have to keep in mind that this benchmark was made for to test the first generation of DX11 graphics cards. It's not very stressful at all on either designer's modern high-end graphics cards.
Take a look at the tessellation-heavy Unigine Heaven benchmark and Anandtech's current tessellation benchmark, Tessmark:
http://www.anandtech.com/bench/GPU12/409
http://www.anandtech.com/bench/GPU13/596
AMD trails a bit behind in Unigine Heaven, and is way behind in Tessmark.
What do you guys think? Have AMD's tessellation improvements worked, or are the latter benchmarks accurate and AMD is still way behind? What do you think they should do to remedy that, if so? I think the obvious step -- and one I feel they should have made to the 7900 series -- is to simply add more geometry engines to the 9900 series.
AMD resolved to improve tessellation performance, and the first significant improvement came with the Radeon HD 6900 series. AMD decoupled geometry functions so that multiple geometry engines, including multiple tessellators, was possible. The 6900 series had two geometry engines, though theoretically AMD could add as many geometry engines as they had die space for. That wasn't enough to dethrone Nvidia, but it was certainly an improvement.
For the Radeon HD 7000 series, AMD brought the dual geometry engine design down to the mid-range 7800 graphics cards. However, they did not increase the amount of geometry engines on the 7900 cards, leaving both tiers with the same geometry engines. This had the odd effect of the higher clocked Radeon HD 7870 beating out the 7950 and even sometimes the 7970 at tessellation benchmarks.
I was browsing through some tessellation benchmarks on Anandtech, and AMD seems to be very competitive on the DirectX 11 Detail Tessellation benchmark:
http://www.anandtech.com/bench/GPU12/423
But you have to keep in mind that this benchmark was made for to test the first generation of DX11 graphics cards. It's not very stressful at all on either designer's modern high-end graphics cards.
Take a look at the tessellation-heavy Unigine Heaven benchmark and Anandtech's current tessellation benchmark, Tessmark:
http://www.anandtech.com/bench/GPU12/409
http://www.anandtech.com/bench/GPU13/596
AMD trails a bit behind in Unigine Heaven, and is way behind in Tessmark.
What do you guys think? Have AMD's tessellation improvements worked, or are the latter benchmarks accurate and AMD is still way behind? What do you think they should do to remedy that, if so? I think the obvious step -- and one I feel they should have made to the 7900 series -- is to simply add more geometry engines to the 9900 series.