OpenPiton – The Open Source Princeton Piton Processor
parallel.princeton.edu
parallel.princeton.edu
1) The end of the free lunch of density/frequency scaling is forcing us to pursue novel architectures to continue progress. If x86 were doubling in performance every 18 months, would this project have ever been started? Even the relatively benign trend towards GPGPU is a symptom of this.
2) These architectures are going to have us revisiting the way we write software. Sufficiently Smart Compiler has long been a joke, but we're going to see something pretty damn clever in the next few decades; taking a core algorithm, written almost in pseudocode (to make as few assumptions as possible about the underlying architecture), and optimizing it for different architectures with wildly different performance constraints.
There's no use in building garbage collected or transactional memory, memory with built in processors, runtime reconfigurable FPGAs, ridiculously NUMA machines, inexact computing if we have to rewrite everything every time we want to use them - the companies will go out of business first.
If this doesn't seem plausible - consider if you had to lead a team to write such an SSC, or build a CPU that beats Skylake with the commercial constraints that come with it (it has to be performant and bug free on a deadline), which would you rather tackle? We're already performing ridiculous engineering feats, but we set the bar lower in software.
3) Our current transistors are stone age compared to what is possible. We're probably a lot closer to physical limits on an exponential scale than we were when we were using vacuum tubes, but in 100 years we're not going to be fabbing the same silicon transistors on a cheap, ultra mature process technology, because they're good enough.
The combination of all 3 is going to yield computing power of unimaginable comparison to what we have today, in our collective lifetimes. We're in a dark place of stagnation now, but we're not done yet as an industry.