Ann: David Perlmutter appointed to Board, page-76

Guys,


https://research.microsoft.com/en-us/projects/catapult/

Check out this video interview with Doug Burger who is a researcher for Microsoft working on the Catapult Project where they are using FPGA's to accelerate data-center services. The discussion is interesting in respect of Weebit because he talks about Moore's Law (1965: number of transistors per chip would double every 18 months to two years which held true for 50 years) and the issues about it reaching its limits. Problems with silicone scaling. Listen to the bit about Dennard scaling from 6min 22seconds (inventor of DRAM??) https://en.wikipedia.org/wiki/Dennard_scaling

Dennard scaling ran for 40 years and failed in the mid 2000's (around 2005) because the insulator on the transistor started to get so thin you couldn't shrink it any more without seeing electrons leaking across it. He is saying that Moore's law will come to an end in the next 5 years. The economics of transistors worked for so long because you could double their density on a chip but still lower the cost of that chip in the market. Once you reach the silicone scaling limit devices using transistors are going to stop getting cheaper so you will have to pay more for more functionality which he says breaks the economics of a lot of these semiconductor businesses.

According to Burger the FPGA's could accelerate cloud software by up to 100 times which gives the cloud services a scaling path for a little while.

I think ReRAM could completely disrupt the server industry because as per previous posts the RAM and storage memories will combine and be able to talk faster with the computer processor (lower costs, less power consumption, smaller physical size, faster write and erase speeds, faster communication with processors). The question is how quickly could they phase in ReRAM? You can't just throw the baby out with the bath water because you've invent something new. These companies are conservative because they have to keep delivering working solutions but if ReRAM memories prove reliable and put Moore's law back in business they will significant improve costs and power consumption in these server farms IMO and will eventually become state of the art. Also I cant see FPGA's being a solution in portable devices because they are still limited by "three terminal device" scalability issues and their functionality needs to be integrated with the conventional computer architecture through internal programming. Because FPGA's are based on transistors they share the same silicone scalability issues so their only real functionality (if ever deployed on mobile devices) will be to provide some very limited machine learning off the back of hardware which in most cases can be done through software or in the cloud servers and then sent back to the device in any case IMO. The density of gates that can be shrunk down onto a compact chip will be too small and/or far to expensive to achieve any serious learning, let alone "machine intelligence" which I don't believe are one and the same thing IMO. ReRAM is totally compatible with CMOS technologies so basically will become plug and play.

Eshmun