The Next Big Innovation in Microprocessors: Anant Agarwal (Part 12)

SM: How much do people need to learn to be able to optimize programs on the Tilera chips? AA: That depends on the applications and the domain. The good news is you have gotten something working and running. Once you have done that, you can then try various optimizations. You don’t have to read 50 page manuals to figure out how to begin to do things. The streaming approach is very simple, and it is a set of API calls that is inspired by sockets. Because the socket API is well known, this makes the learning curve lower for people.

By the way, so far from what you have read and what you have heard, what are your first impressions about this chip and this company?

SM: I think it is very powerful. I also think that a lot of concepts are familiar concepts from many years ago. I spent the latter part of the nineties doing 3 internet startups, and did not really track the Massively Parallel Computing movement. However, it seems to me that much of what you have implemented in Tilera – we were already playing with in Alewife. That makes it a lot of fun for me! AA: Yes, they should be familiar to you. When people say “How could you guys deliver a working chip with so little money in less than two and a half years”, I have to explain it was way more than that.

SM: That is because you have been working on it for 20 years! AA: Exactly, there is 20 years of experiments behind it. In many respects we were very fortunate. We started the Raw project in 1997, and we had a chip in 2002. When we started that project, it was multicore. We were fortunate enough that the whole market is now moving in this direction.

SM: That is not unusual – markets take time to develop. A good example is digital cameras. That technology has been around since the 70’s, so it took decades. I don’t remember the exact dates, but I do know it has been around for a while. AA: We have been working on this for a decade, and we have solved many of the hard issues and we were waiting for the market to develop.

SM: Yes, looks like industry is just catching up with innovations that were in the labs a decade or more back. Everything in the business world is about timing. I’m glad you didn’t try to make a company out of this earlier, because being early in a market is also a problem. You wait around for the market to develop, waste a lot of venture capital, get diluted, and waste precious energy.

[To Be Continued]

[Part 11]
[Part 10]
[Part 9]
[Part 8]
[Part 7]
[Part 6]
[Part 5]
[Part 4]
[Part 3]
[Part 2]
[Part 1]

This segment is part 12 in the series : The Next Big Innovation in Microprocessors: Anant Agarwal
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15