GPUs are an efficient, low-cost [both production AND development costs] & low risk way of increasing throughput in applications and especially for embedded systems.
Many times applications cheat or discount on the real solution:
- Use simpler algorithms because following the actual mathematics is too costly in terms of computation time
- Use heuristics in lieu of mathematical theory to save computation time
- Use integers or short floats instead of better precision floating point computation
- Use fewer sensors to reduce processing time requirements
As experts in the use of GPUs and parallel computation we provide workshops, training and development of GPU-based systems.
Question: What is your expectation for the throughput that can be achieved by using, say 1024 GPU cores in parallel? What does it take to realize high throughput? What are the pitfalls that can prevent GPUs from realizing high throughput ?