Even though it goes after the same market as the last Pro Duo, the new Polaris-based Pro Duo is a much different card in a number of important ways, virtually all of which have to do with the differences in features and design between the previous Fiji GPU and the newer Polaris 10 GPU. Fiji was a high-end GPU that had to be reined in a bit to fit on a dual GPU card – and even then AMD needed a closed loop liquid cooler – whereas the Polaris 10 GPUs that go into the modern incarnation have a much more workable TDP from the start. Coupled with that, Polaris 10 doesn't have the fixed VRAM capacity of Fiji, which means AMD can significantly increase the total VRAM. As a result, the new Pro Duo strikes a very different balance than the old card.
By the numbers then, the Polaris-based Radeon Pro Duo is actually slower than the older Fiji version. It has a clockspeed advantage, but this isn't enough to overcome Fiji's total stream processor advantage. All told, at peak rates the new Pro duo is rated for 11.45 TFLOPs of FP32, while the Fiji version was rated for 16.38 TFLOPs. In practice however I don't expect that the gap is quite so large, as Fiji would need to throttle more often, and Polaris enjoys at least a small architectural advantage.
What the new Pro Duo loses in throughput it more than picks up in VRAM. With each Polaris 10 GPU sporting a 256-bit GDDR5 memory bus, AMD has outfit each GPU with 16GB of VRAM, quadrupling their VRAM capacity. AMD's photos don't show the bare PCB, but I believe it to be a safe bet that AMD is running 8Gb modules in clamshell mode, meaning the board is truly maxed out. AMD argued that the smaller VRAM capacity of the Fiji based card was not a major liability last year, and while there's admittedly no reason not to fully equip a Polaris 10 card since the company is back to GDDR5, I do think it's an important step in making the new card more flexible.
