8 ms·

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 12y 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.

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 12y 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.