4 ms·
CPU and GPU are good at different tasks, but transferring any amount of data across the PCI bus is pretty expensive, forcing you to use both in a suboptimal way
by wongarsu 5y ago
CPU and GPU are good at different tasks, but transferring any amount of data across the PCI bus is pretty expensive, forcing you to use both in a suboptimal way. If you have fast enough data transfer between them then you can do SIMD-like tasks much faster on the GPU, while doing a lot of tasks that naturally involve branching faster on the CPU, using whichever processor is faster and has spare capacity instead of whichever has the data.
- h0l0cube 5y agoThis is correct. CPUs are better suited for arbitrary execution (e.g., physics, game logic, networking) whereas the GPU is suited to highly organized data structures and massively parallel computation. Things that require dynamic effects can benefit greatly from reducing dispatch times between GPU/CPU. UMA, as featured on M1 and the PS4, permit almost instantaneous transfer, save for cache coherency issues. The nexus of real time physics solvers and graphics rendering I think is the next step in taking advantage of SoC and UMA
- ses1984 5y agoOk but in practice what can you do on console that you can’t do on pc?
- h0l0cube 5y agoI mean you could run a database or a spreadsheet on a console, but it wouldn't be a good fit. The benefits of a console are the price-to-performance ratio for gaming applications, whereas the trade-offs on PCs are geared toward general purpose computing and browser speed. And as I've explained a few times in other threads, SoC and UMA architectures permit certain types of processing (such as destructible/deformable environments) to become feasibly performant. Ultimately a console process everything as a PC can, and vice versa, it's just consoles are more performance tuned for gaming, and the other has to cater for the lowest common denominator.