Over the past two months I've read a series of articles about a gloomy end to the relatively predictable improvements forecast by Moore's Law. It all started with an article by the CTO of Broadcom who declared that the nanometer declarations from many companies is well a marketing gimmick. In the interview entitled The Moore's Law blowout sale is ending, Broadcom's CTO says | PCWorld "Broadcom Chairman and CTO Henry Samueli delivered some sobering news: Moore’s Law isn’t making chips cheaper anymore." The article is an interesting read because it goes on to cite particular technologies, economic choices around which ones to deploy, and ultimately the deterministic behaviors around process shrink aren't that certain any longer.
In the past two weeks I've partaken of a presentation by Marc Snir of the Argonne National Labs and the University of Illinois, Urbana-Champaign who ask the question "[Are] We On Our Way to the Next Extinction?" Where the first extinction was the shift from bi-polar technologies to the current CMOS silicon manufacturing approach (see slide 8). His material goes on to cite the lack of solutions arriving for post 7nm CMOS technologies for fundamental issues like "Equivalent Oxide Thickness," "Source-Drain Leakage," "Threshold Voltage," etc. (see slide 11). While this is a long ways off, 2024, this is surely something to pay attention to which gets me to my next article within the past two weeks.
Reading Welcome to the era of radical innovation - Computerworld discusses the money trail from the EU and the US-NSF to procure funding for advanced computational technologies like carbon nano-tubes, nano-wires, etc. Throughout the arc of the article it goes on to describe governmental research organizations recognizing that now is the time to invoke not evolutionary updates to the current model, CMOS, but it is time to initiate a revolutionary shift. Certainly various three letter monikered governmental organizations believe this to be the case and at least one company, D-Wave, has a commercially available quantum computer on the market.
For sure there have been lots of prescient statements about the end of Moore's law, so what's the difference this time? Surely the wizards who figure out how to do the unbelievable will surely come through. Well I think that the difference is that this time the High Performance Computing community is starting to plan for the demise of Moore's law and the technology it is based upon. So instead of something to file on the shelf to read about once a year, I think it is about time to start to ponder about the impacts in the enterprise and more broadly in the commercial sector (and maybe it is even time to act).
As I ponder this point I know that innovations in software will be required, but the fundamental assumption(s) there will be a new and more powerful processor elements, denser memories, higher capacity storage class memories, etc. with few tradeoffs can and should be challenged. So yes the current fad called the Software Designed Data Center is correct, but not in the way that people think. Instead I think many of us will have to become more like the long fabled Eastern European programmers who could do unbelievable things with less than 1K of memory. In some sense we're there today with GP-GPUs, what I think is a hidden Renaissance around using FPGAs, virtual machines, SDNs, etc. To really make them work well you have to look towards end-to-end integration and innovation at the entire systems level to get bona fide benefit. Add into this the sense of constant uncertainty -- a new fundamental assumption -- leading to a desire to unleash more competitive (and experimental) innovations in the markets by users directly and we're not talking about your typical COTS approaches.
So I'll close by saying that I foresee innovations in the SDDC world that take advantage of unique hardware elements to give users competitive edges in their markets including early time to market. The key to get there will be open and not closed approaches, so be wary of seemingly open approaches with "hidden backend costs," and don't be afraid of end-to-end innovations even if they show up in hardware and not software