IC CAD AND TRANSACTIONAL MEMORY
The Virtual World
We now live in a Virtual World. GSEDA has a virtual office and we do much of our research using the Web. The disadvantage of the Virtual World is loss of peripheral information. That’s the stuff that comes from adjacencies that produce out-of-box thinking; which by the way is the best kind. When you do a Web search you direct your inquiry to your specific topic. When you read your newspaper or magazine on the Web it is far too easy to skip articles that you would normally read in hard copy. Those articles are the peripheral information that allows breakthroughs.
I will always remember a conversation I had at Gartner-Dataquest. I was at the coffee bar at our office talking to Bryan Lewis, our ASIC expert, when the then newly hired Nathan Brookwood, our Microprocessor expert, joined the conversation. At the end of the conversation Nathan remarked that he had learned more in the last fifteen minutes than he usually got out of weeks of research. That’s the power of peripheral information.
IC CAD 2011
About the only place you can now get into those kind of conversations is at a trade show and I was lucky enough to get into one at IC CAD. Obviously IC CAD has nothing to do with Transactional Memory; in fact IC CAD covers the exact opposite end of the design spectrum. However there was a CEDA sponsored lunch that featured Chris Malachowsky from NVIDIA. I try not to miss Chris’ talks as he is a friend and an expert in parallel computing. Not only did Chris give a great talk but during lunch, prior to his talk, I found out that IBM had used Transactional Memory in its new Super Computer.
I have been following Transactional Memory for a few years now. It has promised to be the major breakthrough in parallel computing. Unfortunately a little over a year ago I found out that IBM had solved the speed problem (until then Transactional Memories were dog slow) but had run into a different problem; they were power hogs. So I lost interest and in fact started using the term Concurrent Memories instead of Transactional Memory. It looks like I now need to find out how IBM solved the power problem. We may have the breakthrough after all.