FPGAs versus GPUs in datacenters
The document compares FPGAs and GPUs for use in datacenters. Bill Dally, an expert from Nvidia, says GPUs are best for arithmetic and memory-intensive problems because they have high bandwidth and are programmable. FPGAs have more overhead to implement algorithms and are difficult to program. Desh Singh from Altera says FPGAs can be specialized for efficiency by implementing custom hardware from software, but this requires multiplexing resources over time. Both agree the right tool depends on the specific workload and its demands.