The Future of Microprocessors

[ygolo]

The shape of the future of Microprocessors is rather murky, and so rather exciting for those of us who like to be where there is change a foot.

I wanted a take on what the general public thinks about the latest developments in Microprocessors.

I know this is not the general public, but I don't believe there will be too much correlation between MBTI enthusiasts and microprocessor architects and designers.

Even if there is, thoughts and ideas outside of the field always brings perspective.

Warning: this is a pdf.

With the multicore era undeniably upon us, more talk is turning to the future implications of multicore processors. Of course, software development remains a big challenge, even provoking a recent article in The New York Times, of all places. (See NYT 12/17/07,Leaving Programmers In Their Dust,” by John Markoff.) But the discussion is equally spirited on the hardware side.
One debate is about symmetric versus asymmetric multiprocessing. Should all the cores on a multicore chip be identical, or should some be specialized for different tasks? Another debate questions the value of core-level multithreading. How many threads make sense? In many ways, these debates echo the classic RISC versus CISC arguments of the 1990s—simplicity versus complexity, efficiency versus expediency.
[...]

[LostInNerSpace]

[quote=ygolo;168361]The shape of the future of Microprocessors is rather murky, and so rather exciting for those of us who like to be where there is change a foot.

I wanted a take on what the general public thinks about the latest developments in Microprocessors.

I know this is not the general public, but I don't believe there will be too much correlation between MBTI enthusiasts and microprocessor architects and designers.

Even if there is, thoughts and ideas outside of the field always brings perspective.

Warning: this is a pdf.

With the multicore era undeniably upon us, more talk is turning to the future implications of multicore processors. Of course, software development remains a big challenge, even provoking a recent article in The New York Times, of all places. (See NYT 12/17/07,

We are reaching the limits of what we can do with silicon processors. Happily there are other uses for silicon. The future is in quantum computing.

Quantum computing making 'tremendous progress' - quantum-world - 29 November 2002 - New Scientist

[ygolo]

Originally Posted by LostInNerSpace

We are reaching the limits of what we can do with silicon processors. Happily there are other uses for silicon. The future is in quantum computing.

Quantum computing making 'tremendous progress' - quantum-world - 29 November 2002 - New Scientist

This seems like the distant future to me. We still haven't made our transistors go 3D. On the other hand, quantum-dots seem extremely promising.

The International Technology Roadmap for Semiconductors releases a fairly accurate "Executive Summary" each year (tends to be conservative actually).

I think we have till 2013 (32nm node) at least. There is stuff on the roadmap till 2022.

Here is a "summary of the summary" (also a .pdf and also a fairly big size):
http://www.itrs.net/Links/2007Winter...TC_2007_JP.pdf

Attempt to download the actual executive summary at your own risk (linked right off their main page). It is almost 2MB.

Thanks for the input. Quantum computing does seem like something that captures the imagination of people.

I was hoping to glean some insight on multi-cores in particular, since it more immediate, but any input is welcome.

[Architectonic]

I kind of like the idea of 3d carbon nanotube microprocessors - but there are major hurdles involved in actually doing that.

Originally Posted by LostInNerSpace

The future is in quantum computing.

Very distant future maybe.
Quantum computing does seem like a more natural form of computing, but the problem is we still have no idea how to keep complex quantum structures stable for long enough to do useful computations. A state of the art ion trapping rig might be a bit too expensive for the average household... Endohedral fullerenes seem pretty cool too.

[aeon]

I like the idea of higher-clock main processors that are largely symmetric with some additional specialized cores for dedicated processing. While this is faster to implement and not as efficient as a well-considered asymmetric design, I think it will be easier to bring to market for two reasons: 1) it scales easily, satisfying demand for brute-force power, and 2) it better addresses the issues brought about by ever-smaller traces on the wafer.

For dedicated servers, processors with extensive threading make sense. On the other hand, for general-purpose processors in end-user systems, I do not think much benefit is to be had from extensive threading - current compilers do not deliver in such a way that the cores can remain fed enough to stay active per clock. Also, issues of cache coherency mean that extensive threading will limit clock advances because memory subsystems will not keep pace with the I/O needs. I think threading across cores with advanced predictive branching will yield better results than threading within the cores, and their associated cache issues.

That said, I think the gains to be had from small-trace, high-clock symmetry will only scale to a point. What is also needed is advances on the compiler side. Of course, writing closer to the iron would also help, but I do not think we can count on that.

Tighter integration of compiler, predictive branching, and small-trace, high-clock symmetry will be the way - for now.


my 2 cents,
Ian

[ygolo]

Originally Posted by Architectonic

I kind of like the idea of 3d carbon nanotube microprocessors - but there are major hurdles involved in actually doing that.



Very distant future maybe.
Quantum computing does seem like a more natural form of computing, but the problem is we still have no idea how to keep complex quantum structures stable for long enough to do useful computations. A state of the art ion trapping rig might be a bit too expensive for the average household... Endohedral fullerenes seem pretty cool too.

You see things very similar to the way I do. Are you looking to be (or already) a Microarchitect?

I am asking because, you call yourself Architectonic. Welcome.

Originally Posted by aeon

I like the idea of higher-clock main processors that are largely symmetric with some additional specialized cores for dedicated processing. While this is faster to implement and not as efficient as a well-considered asymmetric design, I think it will be easier to bring to market for two reasons: 1) it scales easily, satisfying demand for brute-force power, and 2) it better addresses the issues brought about by ever-smaller traces on the wafer.

SMP does scale really well, but additional specialized cores are not part of the usual model. I would be interested to know what you had mind for those uses.

Originally Posted by aeon

For dedicated servers, processors with extensive threading make sense. On the other hand, for general-purpose processors in end-user systems, I do not think much benefit is to be had from extensive threading - current compilers do not deliver in such a way that the cores can remain fed enough to stay active per clock. Also, issues of cache coherency mean that extensive threading will limit clock advances because memory subsystems will not keep pace with the I/O needs. I think threading across cores with advanced predictive branching will yield better results than threading within the cores, and their associated cache issues.

Unfortunately, this much is true. Especially, if what you meant by the last statement is that we generally program in a way to keep the data local to the caches of particular cores with little communication between the cores.

Having tried to program for this with even moderately complex systems, the task for programmers is incredibly difficult.

Transactional Memory, especially hardware implementations seem rather promising. The system would guarantee ACI of ACID of the shared memory. Allowing the programmer to program completely in "transactions."

Originally Posted by aeon

That said, I think the gains to be had from small-trace, high-clock symmetry will only scale to a point. What is also needed is advances on the compiler side. Of course, writing closer to the iron would also help, but I do not think we can count on that.

Tighter integration of compiler, predictive branching, and small-trace, high-clock symmetry will be the way - for now.


my 2 cents,
Ian

Compilers are always way behind the hardware. They are after all working on problems that tend to be NP-Hard(scheduling and optimization problems) or undecidable (program analysis problems). So approximations need to be made yielding sup-optimal results.

Affine Partitioning is a technique that seems rather powerful for automatic the determination of parallelism, but it seems-like we need a way to do it for time too.

Anyway, thanks guys for delving into it. Thoughts about processor (and processor based system) architecture have some to occupy my mind more and more lately.

[Architectonic]

Originally Posted by ygolo

You see things very similar to the way I do. Are you looking to be (or already) a Microarchitect?

Sorry to disappoint, but I'm an undergrad student with an interest in nanoscience.

I've had this username for about 4 years, but it wouldn't the be first time one of my usernames has foreshadowed one of my interests...