3 ms·

Going to work at a competitor to Pixar in ‘96, it was the first time I was exposed to real life computer science problems that needed more compute power than co

Going to work at a competitor to Pixar in ‘96, it was the first time I was exposed to real life computer science problems that needed more compute power than could be imagined. Between then and 2017, think about the staggering performance increases available, more so than in many fields because of the embarrassing (parallel nature) thing inherent to many of the problems. And yet there is still no end in sight to the compute power that could be leveraged. I wonder what would change if we had say, 3 orders of magnitude performance improvements across the board starting tomorrow. What might be some of the first practical advancements to be exploited and seen? Let’s assume the amount of compute power that could be leveraged effectively and practically for computer graphics is finite, however large. How much less compute power would that be, compared to what it’s going to take achieve general AI? Or even a Turing effective chat bot?

frik 9y ago
I like the 3D animation movies of Pixir with non-realistic rendering (especially from 1990s, 2000s). Though like many I dislike the CGI movies that started with the first Transformer movie (with blue screen action) and now with the dozens of super hero boring fad. Action movies of the 1980s and 1990s and early 2000s were so great. Nowadays thankfully they use more real stunts again, though this mass of boring super hero and hundreds of TV series that no one has time to see. All we want is new high value Hollywood movies.
dahart 9y ago
That's an interesting question, with an interesting premise. Will realism (graphics + physics simulation) take less compute than AI for sure? If we look at the limit case, the reality I can see out my window has more atoms than are in my brain. So, in theory perhaps, human-equivalent AI will similarly require less compute than realistic graphics & physics simulation/animation. OTOH, evolution took billions of years and presumably depends on all the animals and all the physics that has ever happened, so maybe inference will always be cheaper than graphics, and maybe AI training and evolution will always be more expensive? I have no idea, but it's a fun thought experiment. FWIW, I also worked for a Pixar competitor, and the renderfarm we had was indeed more compute than I had been exposed to before. That said, we didn't make a huge amount of effort to optimize the system, we only made sure really stupid things didn't cause overnight renders to take longer than required for dailies the next day. I'm absolutely certain that at least 1 order of magnitude is available to software render farms, if the primary focus shifts from production to optimization. I'm halfway sure that using today's and tomorrow's GPUs could get you at least another order of magnitude, but that comes with a second level of expensive engineering effort. In short, I think if you really really wanted it, and had the money & time to commit, you could probably have your three orders of magnitude right now.
- cr0sh 9y ago
  > In short, I think if you really really wanted it, and had the money & time to commit, you could probably have your three orders of magnitude right now. I'd imagine something along the lines of a datacenter, filled with racks stuffed full of these: https://www.nvidia.com/en-us/data-center/dgx-1/ https://www.nvidia.com/en-us/data-center/dgx-1/ ...but repurposed for graphics rendering instead of deep learning. In fact, it wouldn't surprise me to find out something like this actually exists (though whether for graphics or deep learning, I don't know). Also - I wouldn't be surprised to learn that a datacenter full of D-Wave machines exists somewhere (perhaps next door to the datacenter filled with DGX-1 systems - I mean, something has to make sense of the data output). Pure speculation, of course; I'd suspect we'd know about it though if it were the case, unless it was done as a government black project...

frik 9y ago

I like the 3D animation movies of Pixir with non-realistic rendering (especially from 1990s, 2000s). Though like many I dislike the CGI movies that started with the first Transformer movie (with blue screen action) and now with the dozens of super hero boring fad. Action movies of the 1980s and 1990s and early 2000s were so great. Nowadays thankfully they use more real stunts again, though this mass of boring super hero and hundreds of TV series that no one has time to see. All we want is new high value Hollywood movies.

dahart 9y ago

That's an interesting question, with an interesting premise. Will realism (graphics + physics simulation) take less compute than AI for sure? If we look at the limit case, the reality I can see out my window has more atoms than are in my brain. So, in theory perhaps, human-equivalent AI will similarly require less compute than realistic graphics & physics simulation/animation. OTOH, evolution took billions of years and presumably depends on all the animals and all the physics that has ever happened, so maybe inference will always be cheaper than graphics, and maybe AI training and evolution will always be more expensive? I have no idea, but it's a fun thought experiment. FWIW, I also worked for a Pixar competitor, and the renderfarm we had was indeed more compute than I had been exposed to before. That said, we didn't make a huge amount of effort to optimize the system, we only made sure really stupid things didn't cause overnight renders to take longer than required for dailies the next day. I'm absolutely certain that at least 1 order of magnitude is available to software render farms, if the primary focus shifts from production to optimization. I'm halfway sure that using today's and tomorrow's GPUs could get you at least another order of magnitude, but that comes with a second level of expensive engineering effort. In short, I think if you really really wanted it, and had the money & time to commit, you could probably have your three orders of magnitude right now.

cr0sh 9y ago
> In short, I think if you really really wanted it, and had the money & time to commit, you could probably have your three orders of magnitude right now. I'd imagine something along the lines of a datacenter, filled with racks stuffed full of these: https://www.nvidia.com/en-us/data-center/dgx-1/ https://www.nvidia.com/en-us/data-center/dgx-1/ ...but repurposed for graphics rendering instead of deep learning. In fact, it wouldn't surprise me to find out something like this actually exists (though whether for graphics or deep learning, I don't know). Also - I wouldn't be surprised to learn that a datacenter full of D-Wave machines exists somewhere (perhaps next door to the datacenter filled with DGX-1 systems - I mean, something has to make sense of the data output). Pure speculation, of course; I'd suspect we'd know about it though if it were the case, unless it was done as a government black project...