New Chip Design Lowers Power Consumption

[Arthur Hanson]

Nineteen-Year-Old Takes on Intel in a Bid to Make Computers Less Power-Hungry | MIT Technology Review

This is definitely not in my area of expertise, but I think many of the community will find it an interesting approach to lower power consumption and maybe get people thinking more outside the box. Pros and cons welcome on this since it's so far outside my area of expertise. Another very interesting note is that the developer is skipping college and becoming very productive, very early. This not only benefits the individual, but society by the person having a substantially longer career. I feel careers should start earlier in most cases and education should be on a subscription basis to not only avoiding learning things that are obsolete or have a very short half life left, but keeping one up to date on an ongoing basis. I feel education has not even come close to keeping up with the accelerating change that's here. The incumbent education system is becoming more about protecting its own interest over the interest of society in many cases. I feel much time and expense could be rung out of the education system while delivering a superior product.

 

[hist78]

It seems to go into production in 2017.

"Rex Computing will be sampling its first chips in the middle of next year and will move to full production silicon in mid-2017 using TSMC’s 28 nanometer process."

Supercomputer Chip Startup Scores Funding, DARPA Contract

 

[Daniel Payne]

Arthur,

This is yet another new computer architecture that requires software developers to re-write their existing golden source code or use a new compiler, both can be big obstacles to overcome but it's possible to replace the classic Intel processor in certain industrial segments with the promise of drastically lower power numbers. All the best to Rex Computing in their quest.

 

[hist78]

The CEO and co-founder of Rex Computing Thomas Sohmers will only turn to age 20 by February 29, 2016! The following link gives more technical detail about his Neo processor and an interview with him. It might be hard to understand all those technical aspects for many people, but it's fascinating! If he can successfully implement his ideas, he will make many today's computing approach and processors like old vacuum tube technology.

The Tiny Chip That Could Disrupt Exascale Computing

https://www.youtube.com/watch?v=6OUUC2qoQRg&feature=player_embedded


This is a great example about how the fabless model works. To any person with an excellent idea, there is an eco system to allow him/her to try to realize his/her vision. For Thomas Sohmers, he dropped out of his high school to pursue his dream to create a power efficient super computing processor. I don't think a person with such unique talent and academic background (high school dropout) can even survive in the Intel or AMD or Qualcomm's corporate environment!

 

[Li Yisuo]

Brilliant genius. The young man is truely great!

High performance computation mostly deals with finite volume, finite element and finite difference. By simplifying big problem with small pieces and solve partial differential equations on them we solved many huge problems. Mostly we get a sparse matrix and then play with it. General purpose graphics processing unit, GPGPU, jumps on the HPC stage less than 10 years ago even though parallel computation is not new.

As shown above, with many GPU cores and a decent parallel data cache, matrix operation gets smoother. AMD's Fury X brings 3D high bandwidth memory, HBM into market. Fury X2 has 17 TFLOPS/s and ~375W, 45GFlops/W. Rex's data ,128 Gflops/W, is more promising.

“Caches and virtual memory as they are currently implemented are some of the worst design decisions that have ever been made,” Sohmers boldly told a room of HPC-focused attendees at the Open Compute Summit this week. “A lot of people are thinking about the HPC problem incorrectly. Caches are great when you’re thinking of things from a pure memory and latency point, but as time went on, so many extra features were layered into the cache systems, especially virtual memory, that while making the programmers lives easier (especially before compilers), it added a lot of new inefficiencies.”

It depends on the scale of our problems. Let's use FINFET as an example. A decent mesh of FINFET needs 20k nodes. We couple poison equation, electron/hole continuity equations, hydro dynamics equations(Etemperature, Htemperature) and the lattice temperature together. There are 6 equations. So 64bit x 20k nodes x 6 equations x 6 unknowns, ~ roughtly we have a 46 M matrix size. This is just for a medium scale problem. If you would like to optimize a VLSI circuit, which has multiple million gates, the matrix size might be much bigger. The illustrated 128k x 256 core=32M in the article might not be enough to operate the matrix efficiently. Caches and virtual memory are inefficient for small problems, however it is extremely useful for handling medium and big scale problems.
Indeed we need energy efficiency for HPC. Nevertheless we firstly hope to find a solution. Accordingly I hold the belief, the "Caches and virtual memory" will be still the pillar of HPC for certain time period.

My humble advice:

Please consult Daniel Payne or any other great blogers in semiwiki. The idea is super great. It will have some beneficial application area for sure. And I wish them to achieve their big succes soon.

 

[Arthur Hanson]

As I have written before, the foundry business model is the superior business model for expertise, flexibility and economics. Basically it's based on doing what your good at and contracting out what your not good at. Bio foundries are the next up and coming set of foundries. Quintilles, Charles River and Wuxiapptec are the worlds bio foundries and are in their infancy. 3D printing and design programs are also creating a whole new set of foundries. One thing foundries have in common is that they offer access to cutting edge technologies that most normally couldn't afford or get access to.

 

[hist78]

One thing foundries have in common is that they offer access to cutting edge technologies that most normally couldn't afford or get access to.

I agree with you 100%. In the same article I mentioned above:
Supercomputer Chip Startup Scores Funding, DARPA Contract

Rex Computing will spend $250,000 to produce their first 100 prototype chips. It's not cheap. But if they want to do it by building their own fab, I'm not sure how many '0' need to be added to this $250,000!