Re:Footage of Australian Raven Attacking a Wing Drone in Can...
by mirker on Monday September 20, 2021 @01:33PM
(#28596347)
I mean it’s a bit murkier than that:
1. GPUs used to be much less programmable and had fixed-function graphics pipelines.
2. General purpose GPU computing was demonstrated a decade before the 2012 deep learning craze. OpenCL/CUDA was quite mature in 2012 and used by high performance computing.
3. Transistor scaling (Moore’s law) has been speeding everything up exponentially.
4. There are a lot of non-GPU solutions now.
If GPUs didn’t exist, someone still could have trained a deep model on a cluster of CPU, and then we’d end up with GPUs (or similar hardware) again. So it’s not necessarily true that gaming->GPU->deep learning implies that GPUs couldn’t be used for deep learning without gaming, because you could have had CPU->deep learning->GPU. I think this was much more of a case of the stars magically aligning than some preset goal by Nvidia/AMD.
In fact, there are plenty of cases where you had CPU->research->HW accelerator. One example (on top of gaming/rendering) is compression or parity coding, which was slow initially and then had add-on cards and/or CPU extensions.
1. GPUs used to be much less programmable and had fixed-function graphics pipelines.
2. General purpose GPU computing was demonstrated a decade before the 2012 deep learning craze. OpenCL/CUDA was quite mature in 2012 and used by high performance computing.
3. Transistor scaling (Moore’s law) has been speeding everything up exponentially.
4. There are a lot of non-GPU solutions now.
If GPUs didn’t exist, someone still could have trained a deep model on a cluster of CPU, and then we’d end up with GPUs (or similar hardware) again. So it’s not necessarily true that gaming->GPU->deep learning implies that GPUs couldn’t be used for deep learning without gaming, because you could have had CPU->deep learning->GPU. I think this was much more of a case of the stars magically aligning than some preset goal by Nvidia/AMD.
In fact, there are plenty of cases where you had CPU->research->HW accelerator. One example (on top of gaming/rendering) is compression or parity coding, which was slow initially and then had add-on cards and/or CPU extensions.