Transcript
0:07 so we talked about three things just now we talked about the state-of-the-art of computer graphics and how it's consuming and how it's integrating Computing simulation physics visualization all into one coherent experience uses a lot of computational horsepower you guys are going to see so many different scientific papers around discoveries that are made possible because of the computational capability
0:38 of the GPU massive data realtime data is being searched through and combed through we need faster gpus for all of that and so today i' like to show you the next click of our road map you know that just to refresh your memory in 2008 we introduced firery or we introduced Tesla it's not the brand Tesla but the codeen name Tesla Tesla was our first GPU that incorporated
1:09 Cuda also called G80 internally became Tesla and brought Cuda to the world two years later we introduced fery fery so because we didn't rehearse this the the support team the support team thought that I had made a mistake and they thought that they would like the slide to be Advanced
1:41 meanwhile I didn't want to tell you tip you because you will stop listening to what I'm going to say and so I left that slide right here in my back pocket and I was going to tell you about fery because fery was the first GPU to introduce fp64 and then recently Kepler Dynamic parallelism the GPU can now generate its own work you can now call libraries from within a kernel you could have recur we
2:12 could support recursive algorithms Much More Much More elegantly you can now have data driven execution data dependent execution from within the GPU kernel the programmability that's made possible enhancement and programmability of Kepler is really quite fantastic Kepler also brought to the world a much more energy efficient GPU well the next two gpus from us is
2:45 called Maxwell Maxwell has a unified virtual memory unified virtual memory makes it possible for GPU operation to read and see the CPU memory for CPU code to be able to see the GPU memory all memory is visible to all the processors it'll just make it a lot easier for you to program notice we're trying to improve the per the performance each generation we're trying to improve its Energy Efficiency and we're also trying to improve its programmability and then the next one
3:16 after that is called Volta we all like Volta Volta is a I guess he's been credited with the invention of the battery I love that name because it would suggest that Volta is going to be even more energy efficient even more energy efficient not to mention we really like his inventions Volta is going to have a brand new technology called stacked memory to show you what that means this is what a Kepler looks like if I were to take that graphics
3:46 card take the thermal management unit system out of it okay what you'll have is Kepler like this in the center is the GPU around it are the memories Volta Volta is going to solve one of the biggest challenges of gpus today which is access to memory bandwidth the memory bandwidth on a GPU is already several times higher than a
4:16 CPU but we never seem to have enough we just never seem to have enough and so with Volta we're going to take a giant leap forward we're going to prevent from getting off of our chip onto this PC board which is physical in size and therefore a lot of system parasitics that reduces the signal speed not to mention the power that it takes to get off the chip and get onto a dam chip what we're going to do with Volta
4:48 is to stack the dam right on the same silicon substrate so it looks something like this so the G the GPU Bolta will be integrated on top of another silicon substrate that silicon substrate will carry a whole bunch of dams that are stacked one on top of the other then we Pierce through all the dams using through silicon vas cutting a hole literally through the
5:21 Silicon from the bottom to the top connecting each one of the Stacked memories on top of each other as a result of this as a result of this first of all these are to scale that just gives you a sense of how much smaller Volta will be not to mention how much more performance it we'll have but one of the most significant things about Volta is we're going to achieve one terabytes per second of bandwidth one
5:50 terabytes per second of bandwidth just really really unbelievable stuff now just to give you a perspective for for the audience that that would enjoy likeing doing some math 1 terabytes per second that's like being taking a full Blu-ray that is stored in our memory and within 150th of a second 1 15th of a second move the entire Blu-ray through our chip let's just put that in perspective so that is our GPU road map