AMD FirePro W8000, W9000 Challenge Nvidia's Quadro
These Results Make Little Sense
What do our performance figures indicate? There's a rather fundamental problem. Here's how the FirePro V7900 and W9000 compare in some key departments.
In SpecViewPerf, the average performance delta between the V7900 and W9000 is 11%, despite the W9000 having 60% more cores, 60% more TMUs, and a whopping 2.15X the texture bandwidth.
Now, let's look at those figures for Nvidia when comparing the Quadro 4000 and the Quadro 6000.
The average performance delta between the Quadro 4K and the Quadro 6K? 50%.
The sharp-eyed will notice that the AMD cards have the same number of ROPs (Raster Operation Pipelines) in both cards, while the Quadro 6000 picks up a further 16 units relative to the Quadro 4K. Could this design difference account for the FirePro's relatively poor scaling?
Partially, yes. There are, however, too many places where AMD's scaling fails (or even moves backwards) to pin the problem on ROP count. If ROPs were the limiting factor, we'd expect to see scaling that was significantly influenced by clock rate and fill rate. The W9000's typical speed gain of 11% is far below its 34% clock rate improvement.
In the old days, we might've said this was an OpenGL optimization problem, but the SPEC 3DS Max test runs in Direct 3D. Even if we were willing to accept that Fermi is just that much better than GCN in professional applications, that wouldn't explain why AMD's latest $4000 GPU doesn't significantly outpace its own $650 cousin. Tests like Creo clearly show the benefit of the newer architecture, but it's virtually the only one that does.
The race car used in a number of PTC Creo 2.0 tests
There's something else at play here. It's possible that AMD's pro graphics drivers need much more work to wring all of the performance out of the GCN architecture, that some of the applications we tested are older versions that don't take proper advantage of GCN, or that Nvidia has simply better exploited its relationships with the major software ISV's. As of this writing, Sunnyvale has been silent as to why its brand-new FirePro cards can't keep up with two year-old hardware from Nvidia, or what it intends to do about it. The company's silence on these issues is as troubling as the fact that they exist.
Either PTC Creo is a rare application that just happens to fit GCN's particular strengths, or it's an example of the type of performance we should be seeing, but aren't. We tend to believe it's more the latter than the former, but honestly can't tell for certain. We have to offer an evaluation of the cards as they exists today with the software and drivers at our disposal.
In SpecViewPerf, the average performance delta between the V7900 and W9000 is 11%, despite the W9000 having 60% more cores, 60% more TMUs, and a whopping 2.15X the texture bandwidth.
Now, let's look at those figures for Nvidia when comparing the Quadro 4000 and the Quadro 6000.
The average performance delta between the Quadro 4K and the Quadro 6K? 50%.
The sharp-eyed will notice that the AMD cards have the same number of ROPs (Raster Operation Pipelines) in both cards, while the Quadro 6000 picks up a further 16 units relative to the Quadro 4K. Could this design difference account for the FirePro's relatively poor scaling?
Partially, yes. There are, however, too many places where AMD's scaling fails (or even moves backwards) to pin the problem on ROP count. If ROPs were the limiting factor, we'd expect to see scaling that was significantly influenced by clock rate and fill rate. The W9000's typical speed gain of 11% is far below its 34% clock rate improvement.
In the old days, we might've said this was an OpenGL optimization problem, but the SPEC 3DS Max test runs in Direct 3D. Even if we were willing to accept that Fermi is just that much better than GCN in professional applications, that wouldn't explain why AMD's latest $4000 GPU doesn't significantly outpace its own $650 cousin. Tests like Creo clearly show the benefit of the newer architecture, but it's virtually the only one that does.
The race car used in a number of PTC Creo 2.0 tests
There's something else at play here. It's possible that AMD's pro graphics drivers need much more work to wring all of the performance out of the GCN architecture, that some of the applications we tested are older versions that don't take proper advantage of GCN, or that Nvidia has simply better exploited its relationships with the major software ISV's. As of this writing, Sunnyvale has been silent as to why its brand-new FirePro cards can't keep up with two year-old hardware from Nvidia, or what it intends to do about it. The company's silence on these issues is as troubling as the fact that they exist.
Either PTC Creo is a rare application that just happens to fit GCN's particular strengths, or it's an example of the type of performance we should be seeing, but aren't. We tend to believe it's more the latter than the former, but honestly can't tell for certain. We have to offer an evaluation of the cards as they exists today with the software and drivers at our disposal.
