By: David Kanter (dkanter.delete@this.realworldtech.com), October 1, 2007 1:13 pm
Room: Moderated Discussions
Jonathan Kang (johnbk@gmail.com) on 10/1/07 wrote:
---------------------------
>David Kanter (dkanter@realworldtech.com) on 10/1/07 wrote:
>>Do cache coherency protocols do critical word first? It seems like you would
>probably cut out a good bit of latency...
>
>A good bit but not all or nearly enough I'd say. Even if the arrival of the first
>word (4-bytes) got to the CPU that needs it for processing in the minimal time (assuming
>1 link active, that's 40+ cycles still of idle waiting for one operation, assuming
>only one memory operand is needed), unless the CPU is kept busy for another 32 cycles
>for the next word to arrive (assuming the next word is also the word to be used
>next), it will have to stall again, and then again, etc. >The number of idle cycles will add up between each word >fetch.
My query was sort of unrelated to this point about wake-up latency.
>The ultimate bandwidth of the link in low-latency mode, >low-power mode (1 link
>active) is simply too low to transfer cache information >fast enough for the CPU to process it.
>Ultimately, it's a matter of being able to go from low (or >no) bandwidth to full
>bandwidth in the shortest amount of time (in CSI's case, >one clock cycle).
One cycle isn't really achievable...but a short period of time is feasible.
DK
---------------------------
>David Kanter (dkanter@realworldtech.com) on 10/1/07 wrote:
>>Do cache coherency protocols do critical word first? It seems like you would
>probably cut out a good bit of latency...
>
>A good bit but not all or nearly enough I'd say. Even if the arrival of the first
>word (4-bytes) got to the CPU that needs it for processing in the minimal time (assuming
>1 link active, that's 40+ cycles still of idle waiting for one operation, assuming
>only one memory operand is needed), unless the CPU is kept busy for another 32 cycles
>for the next word to arrive (assuming the next word is also the word to be used
>next), it will have to stall again, and then again, etc. >The number of idle cycles will add up between each word >fetch.
My query was sort of unrelated to this point about wake-up latency.
>The ultimate bandwidth of the link in low-latency mode, >low-power mode (1 link
>active) is simply too low to transfer cache information >fast enough for the CPU to process it.
>Ultimately, it's a matter of being able to go from low (or >no) bandwidth to full
>bandwidth in the shortest amount of time (in CSI's case, >one clock cycle).
One cycle isn't really achievable...but a short period of time is feasible.
DK