ConclusionWhen I first read up on the Fermi architecture last year, I was not sure how well NVIDIA could pull it off. My first thought was, “Yes, it will be a great compute architecture, but will it sacrifice too much of its transistor budget to compute applications and not enough for gaming… which will still be its primary market?” Add onto that thought the problems TSMC has been having with its 40 nm process, which the GF100 will be based on. By early December, after seeing the gaming performance of the midrange and high end AMD HD 5000 lineup, I thought that NVIDIA may have sacrificed too much to CUDA and compute functionality to make it an effective gaming chip.
After seeing what NVIDIA actually has in store, I am far more positive in my outlook for the Fermi architecture when it comes to traditional gaming, as well as upcoming DX11 gaming. Off the cuff I would guess that in current applications, the GF100 based high end card will be 15% to 20% faster overall than the HD 5870. It will not overtake the HD 5970 dual GPU card, but it looks to certainly trounce the single HD 5870. I think that NVIDIA is also doing tessellation right, and when games come out which will implement it in a more pervasive manner, NVIDIA will take a much larger performance lead.
NVIDIA has certainly done a lot of work to make the GF100 architecture as modular as possible. We can easily see that NVIDIA will produce multiple products on the basis of these SM and GPC units. The low end will feature one to four SMs and one GPC, the midrange will feature up to eight SMs and two GPCs, and of course the high end will feature the full 16 SMs and four GPCs. After looking over the architecture, it certainly seems to me that the midrange products from NVIDIA could hold a significant performance lead over AMD’s midrange products (namely the HD 5600 and HD 5700 series of cards). The highly modular setup, combined with how the new SM units are arranged, could spell some trouble for AMD in these high revenue shipment areas.
This is definitely not a “Bridge Too Far” scenario for NVIDIA, and after taking my deep dive into the architecture I truly feel that they have implemented DirectX 11, compute, and tessellation the right way. Now the big question will be if they can produce adequate numbers of their initial GF100 product, and if they can include enough performance in the mainstream and budget categories to take marketshare and mindshare away from AMD. Being this far behind the curve in shipping DX11 products, especially in light of how popular Windows 7 is becoming, will be another hurdle for NVIDIA.
When all of this is cast against the GF100 architecture though, NVIDIA has some serious advantages over AMD. We will have to see if NVIDIA is able to ship enough product in the 9 to 10 months before AMD releases its second generation of DX11 GPUs built on a 28 nm process to regain lost marketshare. It is a huge hurdle in front of NVIDIA, but they have at least given themselves a very compelling tool to address the DX11 and compute market with.
Looking for some quick performance previews of the GF100 architecture compared to GT200? Check out these two videos I recorded at the NVIDIA editor’s day earlier in the month. One looks at Far Cry 2 and the other at NVIDIA’s ray tracing engine.