Electrical Engineering Stack Exchange is a question and answer site for electronics and electrical engineering professionals, students, and enthusiasts. It only takes a minute to sign up.
Sign up to join this community
Modern CPUs today are commonly clocked in the 3-4 GHz range. How fast can typical modern ASICs be clocked?
For example if I were building a commodity ASIC for something like a disk drive or network card, how fast could I clock that ASIC while still following the normal design flow from a foundry like TSMC?
It depends on the complexity of the 'critical path' - the sequence of logic events (gates triggering, signals being propagated down the ASIC trace/metalization - capacitive and inductive integrity losses).
For very simple circuits, it can easily reach 100s of GHz - on the other hand, once you start doing something useful with your signals, it becomes more and more bothersome.
Obviously, such high frequencies can't be delivered on the IO pins - which means you need to start thinking about rather complex PLL systems on-chip (so-called frequency multipliers on CPUs), but, 40Gbit network cards are completely attainable via TSMC, and those tend to communicate with their GBICs at 5-10GHz range.
Depends on the ASIC you are referring to. One thing is for sure, is that higher clock speed of different ICs even of same designed application does not mean the higher clock rate one will be faster.
Take for example a ripple-carry 8-bit adder conpared to another 8-bit carry-lookahead adder. If you clock them, the 8 bit adder would require 18 clock cycles (8*2 +2) while the carry-lookahead requires only 5 clock cycles (lg2 8 +2).
And disk drives are very different from network cards. Also typical disk IO ASICs range from 10-1200 MHz (some HDD vs PCIe SSD), yet it is still dependent on your monetary budget, power/thermal budget, space budget etc.
If you can get your hands on some affordable 45nm process you might get better clocks and thermals, but again clocks are not everything and you might save more money going for TSMC/GloFo old 90nm or 65nm process and getting great performance for your specific need with some clever custom microarchitectural designs.
