SM: What was significant about your time at AMD for both yourself and the industry?
FW: It was a time that brought about the absolute consolidation and predominance of microprocessors. In the 1980’s the RISC processors and single chip processors where gaining importance but there were still mainframes in multiple architectures. By the mid 90’s the x86 became the important instruction set.
AMD emerged later with an interesting decision. People realized that the instruction set is not terrible important, however what is important is micro-architecture. A lot of the concepts of RISC micro-architecture were really important and were incorporated into x86 processors. The capability of processors just absolutely bloomed to replace everything else. Now that has gone to the next level where we are now at a system on a chip level where it is not just microprocessors on a chip but whole multiprocessor systems on chips.
Aside from delivering a good processor on time, there were two reasons the AMD Opteron was so successful. First there was the consolidation of the instruction set. Intel went off into Itanium and AMD stuck with x86 which turned out to be the right move. I described it then like this; Esperanto is a better language than English since it is easier to understand and easier to learn, but nobody speaks it so English wins. That turned out to be true for AMD and the Opteron. The second reason is we got serious about multiprocessing, not only multi-core but the glueless multiprocessor really made it practical with good scaling. It was the perfect time for that technology.
SM: How has that emerged as a foundation for your current work?
FW: A couple of years ago I left AMD after I saw that the processor road map was going to be straight forward. There is still a lot to invent, and it is interesting stuff, but in servers and workstations multicore will go forward for the next decade and a half. It was going to drive innovation get processors back on Moore’s law. There was no question that multicore processors were going to work, and work well, on servers and workstations.
If you look at what Sun has done with Niagara, they have proven that hundreds of threads look great. That is what the next decade is going to look like. It takes me back to my roots at Kendall Square Research, which was more about memory than it was about processors. Once the processors go multicore, multi-threaded applications are on a very fast roadmap. The problem then moves to memory and not to the processor.
There is going to be areal bottleneck in memory systems, and DRAM is not keeping up with what the processors can do. I saw an opportunity to put in a lot more memory initially and then over time put in higher performing memory into the systems so they can keep up with the processors.
About two and a half years ago I founded MetaRam, which was a small startup. That is another reason I wanted to do it. I never intended to be in a 13,000 person company. I am a person who likes to work with a small group of people, make decisions, and do things. At MetaRam, two years later, we have our product to market with only 34 people in the company. It is much quicker to decide what we want to do and do it.