If everything we are reading here is accurate, Intel is going to be utterly humiliated and badly.
Let's take every performance metric as absolute, 2TFLOPS sounds like a great number when you compare it to a current rasterizer. The architecture should lend itself nicely to executing graphics code in terms of extracting paralelism, but why aren't we seeing the IPS rating anywhere? Long before FLOPS mattered in the graphics industry it was all about IPS(Instructions Per Second)- the area where the majority of your actual rasterization comes into play and furthermore where is the enormous amount of eDRAM going to be on this chip and why not more details on that? On a system level they aren't going to have remotely close to enough bandwidth to handle anything resembling rasterization by today's standards which puts them into a place where we already know they are trying to go- ray tracing. Ray tracing uses a staggering amount less read/write bandwidth then rasterization and would lend itself nicely to this general layout, however, seems to me that they are several orders of magnitude short of where they need to be to consider real time ray tracing remotely close to comparable rasterization. 2TFLOPS placed on top of a monster rasterizer is a lot. 2TFLOPS as a full rendering engine is good for a nice chuckle.
Then we get to the other end of the spectrum they are going after- HPC. Intel has tried approaches like this in the past and they have run into the same problem every other CPU maker has in trying to extract the level of paralelism that they are shooting for- trying to figure out a way to get their compilers to work with their processor architecture. This seems to be the fundamental difference between the direction Intel is headed and where nV is going. nVidia seems to be taking their compiler and code base and then building their processors around what the compiler needs to operate most effectively. Don't get me wrong, I think Intel has the most talented compiler coders in the world(and I say that without exception) but anyone who has tried to deal with the problem nows how incredibly daunting it is. I know, a lot of people will rightly point out the incredible resources at Intel's disposal and the fact that they can devote nigh infinite resources to deal with the problem at least compared to their competition. This is a well reasoned and well informed point, but I would have to counter with a single word.
Itanium
If you don't know what it is, look it up
Intel has failed in catastrophic fashion twice before, once when trying to enter the discrete graphics market and once when trying to enter HPC with their explicitly paralel instruction computing- Intel is obviously a very smart company, but one can't honestly fail to wonder how it is they expect to conquer the two markets that soundly rejected them in the past
with the same part.
Edit- Just want to point out for the more well read members that I am grossly oversimplifying as I don't think this would be the ideal space to explain exactly what the difference is in terms of execution units and resources required between IPS and FLOPS and how that relates. Yes, it was intentional, just trying to get the idea across
Facebook
Twitter