4 ms·

What practical application is your chip best for? What can we do with your chip that we won't be able to do with the current line of chips coming in the next fe

by kevin 11y ago

What practical application is your chip best for? What can we do with your chip that we won't be able to do with the current line of chips coming in the next few years if they stay on the current path? The biggest problem for hard technical problems is the storytelling. I went through your site and I couldn't find the sentences I would need to make someone else excited about what you're making in a conversational way. It should feel like you're making something science fiction a reality through your technology alone. That storytelling is what you'll need to convince/inspire customers, investors and future employees.

trsohmers 11y ago
Hi Kevin, Thanks! I really need to update the website as it is based on our (unsuccessful) fundraising pitch... thankfully, we had some great feedback a couple of months ago on storytelling, and it really improved our pitch and we were able to close a round. I'll make updating our website a priority this weekend. Right now, it is best for signal processing applications and dense matrix (although it is not limited to that). To give you some actual applications, think mobile base station processing (Everything from your wifi router to big cell antennas, small cells, etc), computer vision/image processing, video rendering, machine learning (both convolutional neural networks and recurrent NNs), plus a bunch more to long to list. The new abilities that we are providing is that we can fit the performance of a high end system with ~300 watt GPU (which also needs a ~100 watt CPU attached to it), for say machine learning, in a 4 watt power budget. The simple story that we've come up with is this: 30 years ago when existing computer architectures were designed, the power cost of moving data and processing it were roughly equal, but things have changed. Today, it takes 100 picojoules to do a double precision floating point operation, but 4200 picojoules (42 times!) more energy to move that data from RAM to your registers. We're solving that (among many other problems overlooked by the rest of the industry)problem through a new design rather than a radical new technology in order to achieve over a 10x energy efficiency increase over existing systems. I typically have a more succinct oral pitch with accompanying presentation... hopefully this came through clearly. Any feedback would be apppreciated!
- ScottBurson 11y ago
  Today, it takes 100 picojoules to do a double precision floating point operation, but 4200 picojoules (42 times!) more energy to move that data from RAM to your registers. This is a fascinating observation. As one of pg's essays asks, "what do you know that know that no one else knows?" [quoted from memory]. It sounds like this is your answer. Interestingly, it's not literally that no one else knows it -- it's that no one else (AFAIK, of course) has factored it into their designs, as doing so probably requires more-or-less a clean sheet.
  - trsohmers 11y ago
    @Scott When designing the architecture over a year ago, we used that fact as the basis of our design paradigm, and while known to some very smart people in the industry, no one has really dared to do any real redesigns in large companies. When I re-read this post (http://blakemasters.com/post/22866240816/peter-thiels-cs183-startup-class-11-notes http://blakemasters.com/post/22866240816/peter-thiels-cs183-...) and thought about what our "secret" really was, I decided to include that fact in our presentations and it finally started to "click" for people... technical folk and investors alike.
    - ScottBurson 11y ago
      That post is probably what I'm remembering. Apologies for the misattribution.

by kevin 11y ago

trsohmers 11y ago

Hi Kevin, Thanks! I really need to update the website as it is based on our (unsuccessful) fundraising pitch... thankfully, we had some great feedback a couple of months ago on storytelling, and it really improved our pitch and we were able to close a round. I'll make updating our website a priority this weekend. Right now, it is best for signal processing applications and dense matrix (although it is not limited to that). To give you some actual applications, think mobile base station processing (Everything from your wifi router to big cell antennas, small cells, etc), computer vision/image processing, video rendering, machine learning (both convolutional neural networks and recurrent NNs), plus a bunch more to long to list. The new abilities that we are providing is that we can fit the performance of a high end system with ~300 watt GPU (which also needs a ~100 watt CPU attached to it), for say machine learning, in a 4 watt power budget. The simple story that we've come up with is this: 30 years ago when existing computer architectures were designed, the power cost of moving data and processing it were roughly equal, but things have changed. Today, it takes 100 picojoules to do a double precision floating point operation, but 4200 picojoules (42 times!) more energy to move that data from RAM to your registers. We're solving that (among many other problems overlooked by the rest of the industry)problem through a new design rather than a radical new technology in order to achieve over a 10x energy efficiency increase over existing systems. I typically have a more succinct oral pitch with accompanying presentation... hopefully this came through clearly. Any feedback would be apppreciated!

ScottBurson 11y ago
Today, it takes 100 picojoules to do a double precision floating point operation, but 4200 picojoules (42 times!) more energy to move that data from RAM to your registers. This is a fascinating observation. As one of pg's essays asks, "what do you know that know that no one else knows?" [quoted from memory]. It sounds like this is your answer. Interestingly, it's not literally that no one else knows it -- it's that no one else (AFAIK, of course) has factored it into their designs, as doing so probably requires more-or-less a clean sheet.
- trsohmers 11y ago
  @Scott When designing the architecture over a year ago, we used that fact as the basis of our design paradigm, and while known to some very smart people in the industry, no one has really dared to do any real redesigns in large companies. When I re-read this post (http://blakemasters.com/post/22866240816/peter-thiels-cs183-startup-class-11-notes http://blakemasters.com/post/22866240816/peter-thiels-cs183-...) and thought about what our "secret" really was, I decided to include that fact in our presentations and it finally started to "click" for people... technical folk and investors alike.
  - ScottBurson 11y ago
    That post is probably what I'm remembering. Apologies for the misattribution.