We have more questions sent to Futuremark and the answers are still pending, that is, how specific is the physics workload affinity to the CPU or, from a different perspective, what are chances that GPUs in the system will be able to cannibalize some of the CPU physics workload and process it on their own. None of these questions should impact the value of 3DMark 11 as a benchmark, however, it is clear that under the circumstances the Physics benchmark cannot be considered a classic CPU benchmark since other factors can have an impact of up to 16% (from the limited sample we were running) in the score spread.
How much of this is GPU dependent and how much of it is CPU limited is still up for debate but it looks like Lost Circuits has some answers at least:
In other words, the graphics card still needs to do the rendering of the visual output of the workload and, depending on the capabilities of the graphics card, this will have more or less impact, especially if more threads are running. In the case of the Gulftown and Thuban processors, six threads are running, meaning that the workload is digested quite efficiently at high frames per second, which, in the case of a lower power graphics card may have some impact. The problem in this case is that we see almost the same effect when running two GeForce GTX 480 in SLI configuration, which should have enough render power to keep up with the workload. or at least, keep up with any single RADEON card. In fact, the Physics score with nVidia cards appears to be inversely correlated to the performance of the graphics card.