3 ms·
Nvidia Tesla T4 GPU for Data Centers
- PaulHoule 8y agoIf anyone is going to make money off machine learning models, then the cost of inference is the real bottleneck, not the cost of training. (If the number of inferences to be done is not greater than the number of training examples required, why not just do them all by hand?) So long term there should be a bright future for inference accelerators. Some of the capability is being backported for graphics: for instance, use a neural network to touch up aliasing errors in ray-traced or rendered graphics.