On Wed, Aug 25, 2021, 3:33 AM <[email protected]> wrote:
> And what about this Matt ?? >> > https://www.flickr.com/photos/jurvetson/51391518506/ > What about it? You can extrapolate the graph to see Cerebras is a few years away from building it. The Business Wire article just repeats the same press release. That's assuming they can raise funding AND Moore's law continues. But you might have noticed that computer clock speeds stalled around 2010 and transistor sizes are stalling now. You can't distinguish between P and N type silicon below a few nanometers. The problem with sparse matrices is that any representation using pointers requires random memory access, which is 100 times slower than sequential access because address decoding is O(log n) in memory size. All this extra logic produces heat, which is the main reason we don't use wafer scale integration. Repairing manufacturing defects isn't the problem. Once we finalize the optimization of existing hardware over the next decade, the only way to solve the power problem is the way our bodies do it, by moving heavy, slow atoms and molecules instead of electrons. Biotechnology is advancing rapidly but it will be a few decades before it catches up to our stalled transistor technology. Google will probably buy Cerebras for cheap and build their own system. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T607f92b6c1ea30a8-Mad0e13fb1701d9241c120ab9 Delivery options: https://agi.topicbox.com/groups/agi/subscription
