AMD's Flagship FirePro W9100 Graphics Card Boasts 16GB of GDDR5 Memory



+ Add a Comment


I know I was a bit off in the ballpark saying they were probably $1,500 but I was still in the range of $3000. So I don't know why they deleted my previous comment.

Also before anyone asks, NO, they're lousy for GPU-mining scrypt cryptos. In fact, GPU-mining litecoin derivatives or any kind of scrypt-based crypto is inefficient now, you're better of getting the much cheaper and much more efficient Gridseed scrypt ASIC miners.



I'm going to be so jealous, as my friend is probably going to get one (or more) of these from his company for free. Meh. :P



Beforehand, I would like to apologize for the long post. I got a bit carried away :)

AMD really needs to get on PCIe3.0 support above the APU level. With these workstation level cards completely capable of saturating 3.0 during compute intensive workloads, most builders are going to be looking at Intel systems for the CPU half of the equation since Opterons are still stuck on PCIe2.0. Some would argue that the bottleneck wouldn't matter 90 percent of the time, and I can agree. Unfortunately, the people buying the system don't know that. They simply know that 3.0 is XYZ times faster than 2.0 and that the longer it takes to render this movie, the longer they wait to get paid. This is a shame for AMD because professional software is miles ahead of consumer hardware when it comes to multithreading and proper instruction set extension targeting, making the Opteron line a reasonable X86 choice for anyone looking to build a workstation.

Also, while I do get that the 16GB framebuffer will help alleviate some of those bottleneck issues for that ever important 10% of the time, there is still the fact that not everything can be shipped off to the GPU beforehand and kept there until the work is done. Sometimes, things just happen on the software side and there is nothing you can do but ship off new data to the GPU. Eliminating this bottleneck can have huge implications for performance, as demonstrated by HSA acceleration in LibreCalc:
(fifth page, bottom chart)

I believe the performance delta would be even more pronounced with PCI due to latencies involved vs memory, but I'm no expert, so I wouldn't really know. Regardless, it is what it is.



If the CPU is going to be completely maxed out when rendering a 3D model and the GPU is just going to sit there doing nothing, making the CPU do 95% of the work, I don't see the purpose of these FirePro and Quadro cards. The CPU that's used is irrelevant because even the i7-3970x gets maxed.



I'm pretty sure if you have one of these, you're not running an i7 3970X but one or two Xeon cpu.