You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Hi All - I'm really trying to help boost my biz acumen by understanding the pros & cons of using GPU vs FPGA accelerators within an ADAS solution. My [limited] understanding is GPUs are good for parallel processing, which helps in Deep Learning, thereby helping the car identify & recognize objects on the road. Whereas FPGAs are reprogramable logic that are good at executing one specific task, and once the Deep Learning is set, FPGAs can help the car infer and make decisions with low latency.
But where my finance brain is lost... is it seems (at a high level) a CPU + GPU + FPGA + Modem are all needed to help with getting an L5 car on the streets. So, why do you have Nvidia focusing on GPUs as the answer to ADAS and Intel focusing on FPGAs? What are the pros & cons of each?
Short answer is that you build solutions around where you are already strong:
* Intel - CPUs are a non-player in deep learning (which is what a lot of recognition requires), Intel doesn't have a play in GPUs, they do have a play in FPGAs through Altera
* NVIDIA - very strong in GPU, they have nothing I am aware of in FPGA
Modems are a different story. Branching out into anyone of these areas would be a huge and risky investment. Put another way, if you have a hammer, you look for problems that look like nails.
I'll add that you are right, in that a solution for L5 autonomous vehicles will require some combination of CPU, GPU, FPGA, and connectivity. Different companies have different perspectives as to which components will more or less important, but that argument is a little moot unless you have the software and fabric to make all the pieces of the solution work well together. NVidia seems to have a bit of a head start from a software standpoint, but Intel is investing heavily here. A number of companies, including NVidia, Xilinx, Mellanox, IBM, and others, in the fabless ecosystem are working together to design a common fabric. It's a bit of a disadvantage to have a large consortium where everyone has to agree on the standards vs Intel, which can just design a fabric for its own needs.
