The Interesting Tale Of AMD's FirePro Drivers

When starting with the SPECViewPerf 9.0 industry workstation benchmark, we used the 3dsmax-04 view-set. As you can see from the graph above, the FirePro V8700 performance under Ubuntu Linux crept up throughout 2009 and then in 2010 there are two prominent spikes -- early this year and again during the summer -- the periods during which AMD announced their FirePro optimizations. What's interesting though and what unfortunately was not uncovered during our second set of FirePro optimization tests is that the second set of driver optimizations are really just re-implementing the driver work from the first time around. In Catalyst 10.5, the optimizations regressed and were then fixed / reintroduced in Catalyst 10.7. The second set of optimizations has since been lost again with Catalyst 10.8. Since Catalyst 10.8 and through this month's Catalyst 10.12 driver, the performance has been relatively flat and its net gain compared to the beginning of 2009 is actually quite small for this SPECViewPerf view-set. AMD will hopefully fix the driver performance again in 2011 to bring back these prior optimizations.

With the CATIA test in SPECViewPerf 9.0, it is pretty much the same story with the FirePro performance slowly maturing and then the two temporary spikes in 2010. The weighted geometric mean for this test in the Catalyst 9.2 driver was 25.33, bottomed out at 23.68 in Catalyst 9.6, peaked at 34.96 in Catalyst 10.7, and with Catalyst 10.12 is now at 27.76.

The EnSight test within SPECViewPerf 9.0 takes a slightly different twist to the driver story. The first time the FirePro V8700 performance really improved was during their first set of driver optimizations in 2010, but like the previous two tests, it regressed for a couple months. Except the Catalyst 10.2 optimizations actually restored the performance to where it was at a year earlier in Catalyst 9.2 when the RV770-based graphics card was first introduced. Though unlike the first two tests, when the second set of optimizations came this summer, since then the performance has been maintained. However, this is not a net gain since most of 2009 this test was left in a regressed state. In Catalyst 9.2 the weighted geometric mean was 42.64 before dropping to 29~32 throughout 2009, and now with Catalyst 10.12 it is at 42.21.

The fourth test of SPECViewPerf 9 again shares a similar fate. There are two major spikes in performance during 2010 when AMD's marketing department announced the workstation driver optimizations, but these optimizations have not been sustainable. The Catalyst 10.12 driver is 13% faster than where it was at with Catalyst 9.2, but it is nearly 30% lower than where it was at with the "magical" Catalyst 10.7 driver. What is different though about this test was the small performance boost introduced last fall with the Catalyst 9.10 driver.