3 ms·

A bit offtopic: It would be great if there was a 10GHz CPU (single-core), for some special domains. A hardware startup company that designs a new CPU that int

by frik 13y ago

A bit offtopic: It would be great if there was a 10GHz CPU (single-core), for some special domains. A hardware startup company that designs a new CPU that interrups the status-quo would be great. (3GHz in 2004, 3.8GHz in 2006, ~4GHz in 2014) Like in the 1990s 3dfx (GPU) and Cyrix (CPU): http://en.wikipedia.org/wiki/3dfx http://en.wikipedia.org/wiki/3dfx http://en.wikipedia.org/wiki/Cyrix http://en.wikipedia.org/wiki/Cyrix

DanWaterworth 13y ago
It would be great if there was a 10GHz CPU (single-core), for some special domains. There's a reason it hasn't happened.
- frik 13y ago
  What are the reasons? Moore's law? What are the physical reasons beside heath transfer? In 2014 we have now the Intel Haswell (14 nm) with 3D tri-gate transistors yet just 3.7 GHz single-core clock rate. A 10 GHz CPU from Intel would be possible today, we just need dielectric liquid cooling (not water, but pressured Fluorinert) like the http://en.wikipedia.org/wiki/Cray_2 http://en.wikipedia.org/wiki/Cray_2 and http://en.wikipedia.org/wiki/Fluorinert http://en.wikipedia.org/wiki/Fluorinert General public cannot buy such equipment, but it does exist.
  - gradschool 13y ago
    At 10GHz, a signal can propagate no more than 300 microns in a cycle due to the speed of light. The CPU would have to be designed either asynchronously or with many more clock domains than currently. The performance probably wouldn't scale linearly by any realistic metric, and you'd still have the same memory latency.
    - knz42 13y ago
      Arguably the previous poster only mentioned he wants a 10GHz processor, not fast off-chip memory accesses. Also you can achieve a lot of performance in signal processing applications using a fast-clocked single core with on-chip SRAM scratchpads.
    - T-hawk 13y ago
      Your math is two digits off. Using Grace Hopper's "one nanosecond is one foot", one cycle at 10 GHz is 1/10 of a foot or 3 cm. That's 30,000 microns not 300. But yes, the speed of light limitation on signal propagation is a major factor in our current limit of 3 to 4 GHz chips.
  - knz42 13y ago
    There are 3 physical reasons besides heat dissipation: - size: higher frequency means you need smaller circuits, otherwise the electric impedance works against fast switching. There's a limit to how small we can build chip components right now. - molecular stability: higher frequency requires higher voltage (to provide the energy), which in turns makes it easier for the metal oxide to degrade over time (this effect is additional to degradation due to heat) - phase divergence: the digital circuits in our processors are heavily synchronized, and require components to react simultaneously to clock edges. However the propagation of the clock signal across the chip introduces errors in the signal phase. As the frequency increases, phase errors accumulate and cause circuits to "lose synchronization". To avoid this, one must increase the complexity of the clock distribution network greatly, which in turn compounds the effects above (and the heat problem). All in all the current state of affairs is summarized by a rule of thumb (so-called "Pollack's law"): the speed (instructions/sec) of a single core processor increases with the square root of its complexity. By this rule, to clock up from 3 to 10GHz would be a 9-fold increase in complexity, for which we don't have the technology ready just yet.
alecdbrooks 13y ago
In what special domain do you envision such a high clock rate helping? Performance is only loosely connected to GHz[0]. If there is a case where clock rate trumps performance, I'm genuinely curious to know. As processor architects are hitting fundamental physical limits when it comes to increasing clock speeds[1], it would probably take a team of world-class engineers, not a few disruption-minded startup founders, to create a 10GHz CPU. Such a chip would have extreme cooling and power demands and wouldn't necessarily perform well. Cyrix doesn't seem like a great example, as it mostly competed with Intel on its budget chips, if Wikipedia is correct. I suspect the greater amount of low-hanging fruit in processor design made it easier for smaller companies to compete with Intel in the 90s. [0]: http://arstechnica.com/gadgets/2011/04/ask-ars-whats-the-relationship-between-cpu-clockspeed-and-performance/ http://arstechnica.com/gadgets/2011/04/ask-ars-whats-the-rel... [1]: https://www.quora.com/Why-havent-CPU-clock-speeds-increased-in-the-last-5-years https://www.quora.com/Why-havent-CPU-clock-speeds-increased-.... http://www.reddit.com/r/askscience/comments/ngv50/why_have_cpus_been_limited_in_frequency_to_around/ http://www.reddit.com/r/askscience/comments/ngv50/why_have_c....
- frik 13y ago
  Higher clock speed, more instructions per cycle and good pipelining architecture means higher single-core raw performance. We had already complex pipelining superscalar architecture with Pentium Pro and Pentium 4 (Netburst/Prescott). The current Core Architecture has a lot simpler pipelining based on Pentium 3/M. (see http://en.wikipedia.org/wiki/Clock_rate http://en.wikipedia.org/wiki/Clock_rate ) I favor a high single-core over a slow many-core CPUs. Have you ever coded a many-core application that runs on thousands of CPUs? I have done using it using http://en.wikipedia.org/wiki/Cilk http://en.wikipedia.org/wiki/Cilk , http://en.wikipedia.org/wiki/OpenMP http://en.wikipedia.org/wiki/OpenMP (and CUDA and OpenCL on GPU), as well as traditional using operating system process and threads. You need new algorithms that work on massive parallel computers. Converting algorithms from serial to massive parallel is possible in many cases, but really hard science work (have done it). Nevertheless, for a specific domain I would need a really high speed single-core CPU. A good book about the topic is "Inside the Machine" from ArsTechnica: http://www.amazon.com/Inside-Machine-Introduction-Microprocessors-Architecture-ebook/dp/B004OEJO0A/ http://www.amazon.com/Inside-Machine-Introduction-Microproce... ...and various university lectures.
chopin 13y ago
If believe, if any, we (as in software engineers) would benefit more from faster memory and higher bus frequency. In many cases fetching from and writing to memory is much more of a choke point than raw processor speed.
sparkie 13y ago
If we throw away conventional wisdom about how processors work, we could do something magic: throw away the clock. Ivan Sutherland has done some work on a new processor architecture called Fleet, which takes this premise and rethinks how we should program from the bottom-up. Instead of instructions performing operations, the programming model becomes one of passing messages between individual units in the processor, and the problems become that of traffic management and routing, which happens to relate to functional programming, the actor model, and data-flow paradigms. Have a read of "The tyranny of the clock", and other publications by the Asynchronous Research Center: http://web.cecs.pdx.edu/~mroncken/ARC-2012-is13_TuringCentenary_IvansRemarks.pdf http://web.cecs.pdx.edu/~mroncken/ARC-2012-is13_TuringCenten... http://arc.cecs.pdx.edu/publications http://arc.cecs.pdx.edu/publications

by frik 13y ago

DanWaterworth 13y ago

It would be great if there was a 10GHz CPU (single-core), for some special domains. There's a reason it hasn't happened.

frik 13y ago
What are the reasons? Moore's law? What are the physical reasons beside heath transfer? In 2014 we have now the Intel Haswell (14 nm) with 3D tri-gate transistors yet just 3.7 GHz single-core clock rate. A 10 GHz CPU from Intel would be possible today, we just need dielectric liquid cooling (not water, but pressured Fluorinert) like the http://en.wikipedia.org/wiki/Cray_2 http://en.wikipedia.org/wiki/Cray_2 and http://en.wikipedia.org/wiki/Fluorinert http://en.wikipedia.org/wiki/Fluorinert General public cannot buy such equipment, but it does exist.
- gradschool 13y ago
  At 10GHz, a signal can propagate no more than 300 microns in a cycle due to the speed of light. The CPU would have to be designed either asynchronously or with many more clock domains than currently. The performance probably wouldn't scale linearly by any realistic metric, and you'd still have the same memory latency.
  - knz42 13y ago
    Arguably the previous poster only mentioned he wants a 10GHz processor, not fast off-chip memory accesses. Also you can achieve a lot of performance in signal processing applications using a fast-clocked single core with on-chip SRAM scratchpads.
  - T-hawk 13y ago
    Your math is two digits off. Using Grace Hopper's "one nanosecond is one foot", one cycle at 10 GHz is 1/10 of a foot or 3 cm. That's 30,000 microns not 300. But yes, the speed of light limitation on signal propagation is a major factor in our current limit of 3 to 4 GHz chips.
- knz42 13y ago
  There are 3 physical reasons besides heat dissipation: - size: higher frequency means you need smaller circuits, otherwise the electric impedance works against fast switching. There's a limit to how small we can build chip components right now. - molecular stability: higher frequency requires higher voltage (to provide the energy), which in turns makes it easier for the metal oxide to degrade over time (this effect is additional to degradation due to heat) - phase divergence: the digital circuits in our processors are heavily synchronized, and require components to react simultaneously to clock edges. However the propagation of the clock signal across the chip introduces errors in the signal phase. As the frequency increases, phase errors accumulate and cause circuits to "lose synchronization". To avoid this, one must increase the complexity of the clock distribution network greatly, which in turn compounds the effects above (and the heat problem). All in all the current state of affairs is summarized by a rule of thumb (so-called "Pollack's law"): the speed (instructions/sec) of a single core processor increases with the square root of its complexity. By this rule, to clock up from 3 to 10GHz would be a 9-fold increase in complexity, for which we don't have the technology ready just yet.

alecdbrooks 13y ago

In what special domain do you envision such a high clock rate helping? Performance is only loosely connected to GHz[0]. If there is a case where clock rate trumps performance, I'm genuinely curious to know. As processor architects are hitting fundamental physical limits when it comes to increasing clock speeds[1], it would probably take a team of world-class engineers, not a few disruption-minded startup founders, to create a 10GHz CPU. Such a chip would have extreme cooling and power demands and wouldn't necessarily perform well. Cyrix doesn't seem like a great example, as it mostly competed with Intel on its budget chips, if Wikipedia is correct. I suspect the greater amount of low-hanging fruit in processor design made it easier for smaller companies to compete with Intel in the 90s. [0]: http://arstechnica.com/gadgets/2011/04/ask-ars-whats-the-relationship-between-cpu-clockspeed-and-performance/ http://arstechnica.com/gadgets/2011/04/ask-ars-whats-the-rel... [1]: https://www.quora.com/Why-havent-CPU-clock-speeds-increased-in-the-last-5-years https://www.quora.com/Why-havent-CPU-clock-speeds-increased-.... http://www.reddit.com/r/askscience/comments/ngv50/why_have_cpus_been_limited_in_frequency_to_around/ http://www.reddit.com/r/askscience/comments/ngv50/why_have_c....

frik 13y ago
Higher clock speed, more instructions per cycle and good pipelining architecture means higher single-core raw performance. We had already complex pipelining superscalar architecture with Pentium Pro and Pentium 4 (Netburst/Prescott). The current Core Architecture has a lot simpler pipelining based on Pentium 3/M. (see http://en.wikipedia.org/wiki/Clock_rate http://en.wikipedia.org/wiki/Clock_rate ) I favor a high single-core over a slow many-core CPUs. Have you ever coded a many-core application that runs on thousands of CPUs? I have done using it using http://en.wikipedia.org/wiki/Cilk http://en.wikipedia.org/wiki/Cilk , http://en.wikipedia.org/wiki/OpenMP http://en.wikipedia.org/wiki/OpenMP (and CUDA and OpenCL on GPU), as well as traditional using operating system process and threads. You need new algorithms that work on massive parallel computers. Converting algorithms from serial to massive parallel is possible in many cases, but really hard science work (have done it). Nevertheless, for a specific domain I would need a really high speed single-core CPU. A good book about the topic is "Inside the Machine" from ArsTechnica: http://www.amazon.com/Inside-Machine-Introduction-Microprocessors-Architecture-ebook/dp/B004OEJO0A/ http://www.amazon.com/Inside-Machine-Introduction-Microproce... ...and various university lectures.

chopin 13y ago

If believe, if any, we (as in software engineers) would benefit more from faster memory and higher bus frequency. In many cases fetching from and writing to memory is much more of a choke point than raw processor speed.

sparkie 13y ago

If we throw away conventional wisdom about how processors work, we could do something magic: throw away the clock. Ivan Sutherland has done some work on a new processor architecture called Fleet, which takes this premise and rethinks how we should program from the bottom-up. Instead of instructions performing operations, the programming model becomes one of passing messages between individual units in the processor, and the problems become that of traffic management and routing, which happens to relate to functional programming, the actor model, and data-flow paradigms. Have a read of "The tyranny of the clock", and other publications by the Asynchronous Research Center: http://web.cecs.pdx.edu/~mroncken/ARC-2012-is13_TuringCentenary_IvansRemarks.pdf http://web.cecs.pdx.edu/~mroncken/ARC-2012-is13_TuringCenten... http://arc.cecs.pdx.edu/publications http://arc.cecs.pdx.edu/publications