4 ms·
Though, GPUs are meant for highly specialized number-crunching, they doesn't have direct access to system memory. One also gets additional latency transferring
by z3phyr 14y ago
Though, GPUs are meant for highly specialized number-crunching, they doesn't have direct access to system memory. One also gets additional latency transferring data between RAM and GPU memory.
Also, GPU is a temporary technology. Wait for some time, and see CPUs getting more GPU like.
- pjmlp 14y ago> Though, GPUs are meant for highly specialized number-crunching, they doesn't have direct access to system memory. One also gets additional latency transferring data between RAM and GPU memory. Yes they do when shared memory is used, not all GPUs have dedicated memory. Even the PS4 will have shared memory between CPU and GPU systems. > Also, GPU is a temporary technology. Wait for some time, and see CPUs getting more GPU like. This is Intel's wish because their GPUs suck, I don't think it will happen.
- z3phyr 14y agoAMD is also talking about 'GPU technology' afaik. Game devs head for it too, so that they don't have to manage a shader language. It is inevitable. This will happen.
- rdavl 14y agoHmm, I always saw each of them as an answer to completely different problem, so I have hard time contemplating this scenario. How exactly would CPU be more like GPU? What would happen to branching and out of order execution pipelines? Do you see today's complex superscalar cores so cheap and small in the future that you could have thousands of them on single die? What other way is there get vector computations to work as well as scalar computations on same hardware? Would you rather have few superscalar cores and a lot of SIMD's around it? Or do you envision a completely new architecture? I'm just trying to get a picture of how this would be done. Any reading material would be welcome.