FPGAs have been the next big thing for at least 15 years, during which time the consumer GPU has made huge performance increases, and is turning into a general purpose parallel vector unit. Still seems to me the GPU is on track to fill this role and the fpga will stay niche, has anything changed to change this?
What's changed is the definition of GPU. 15 years ago, the most programmability a GPU had was a few registers to switch between a handful of hardcoded modes of operation. For the past several years, GPUs have been something like 80% composed of Turing-complete processors that can run arbitrary software, and they've more recently gained features like a unified address space with the host CPU, and the ability to partition the array of vector processors to run more than one piece of code at the same time. These features are 100% irrelevant to realtime 3d rendering.
GPU architectures don't exactly resemble FPGAs when viewed up close, but for the right kind of problem, an application programmer can now deal with a GPU in the same manner as an FPGA. (In fact, with OpenCL, you can use the exact same APIs.)
> FPGAs have been the next big thing for at least 15 years
It's a tough equation. They are certainly not aimed at or geared towards a commodity co-processor slot next to the CPU in a consumer PC.
What I am arguing or proposing is that if Intel acquired Xilinx and got Microsoft and Apple behind the idea of integrating an FPGA as a standard co-processor for the i-whatever chip we could see vast performance improvements for certain applications.
Per my example, I can take a $100 FPGA today and process 1.5 billion 32 bit pixels per second through an FIR pipeline. That same hardware is instantly reprogrammable to implement a myriad of other functions in hardware, not software. The possibilities are endless.
Many years ago I was involved in writing some code for DNA sequencing. We wrote the software to gain an understanding of what had to happen. Once we knew what to do, the design was converted to hardware. The performance boost was unbelievable. If I remember correctly, we were running the same data sets 1,000 times faster or more.
GPU's are great, but I've seen cases where a graphically intensive application makes use of the GPU and precludes anything else from using it. Not having done any GPU coding myself I can't really compare the two approaches (FPGA vs. GPU). The only thing I think I can say is that a GPU is denser and possibly faster because it is fixed logic rather than programmable (it doesn't have to support all the structures and layers that an FPGA has).
What I'd love to see is an FPGA grafted onto a multicore Intel chip with a pre-defined separate co-processor expansion channel via multiple MGT channels (Multi-Gigabit Transceivers). This would allow co-processor expansion beyond the resources provided on-chip. If the entire thing is encased within a published-and-supported standard we could open the doors to an amazing leap in computing for a wide range of advanced applications.
Let's see, Intel market cap is around $120 billion and Xilinx is somewhere around 8 billion. Yeah, let's do it.
