The ramp of T4 is completely related to customers porting their model on top of our platform. And the inference model is really complicated. This is one of the things that I've talked about in the past that on the one hand, people think that inference appears to be simple because there are so many ASICs built being talked about. The vast majority of the complexity of inference is actually in the optimizing compiler on top.
The TensorRT, fifth-generation optimizing compiler that we announced just recently took 3, 4 years to build. And then on top of that, in order to get it to scale as quickly as what people saw in Google -- Google's cloud requires us to build something called a TRT server, an inference server that allows multiple models to run on top of Kubernetes in the cloud. That piece of software is also super complicated to write. And so the pieces of technology that we're putting together have come together.
...
And at this time and as we look out into the horizon, the T4 Cloud GPU is just unquestionably the most effective. It can run models, whether it's an image model or a recommendation model or a speech synthesis model. It is the highest throughput processor in the world at 70 watts, which fits into a hyperscale data center OCP server. It is also the lowest latency of any processor at inferencing in the world at less than 1 millisecond.
