By: Martin Høyer Kristiansen (test.delete@this.example.com), November 18, 2010 1:00 am
Room: Moderated Discussions
anon (a@b.c) on 11/18/10 wrote:
---------------------------
>Martin Høyer Kristiansen (test@example.com) on 11/18/10 wrote:
>---------------------------
>>...And width of the registers. Intel switched back to read-after-schedule (used
>>in P4) for Sandy Bridge because supporting fat result buses would burn much more power.
>
>Read-after-schedule? What is that, a particular OoO trick?
It just means instructions schedule when data is ready. The scheduled instruction then reads source operand values from the physical register file (that's the PRF they talk about in the Sandy Bridge articles)
PPRO and derivatives up to and including Nehalem broadcast results in the ROB. Each ROB entry listens for the source operands needed. When both operands are ready the instruction can schedule, thus schedule-after-read. Broadcasting multiple wide register values to many ROB entries costs power.
Cheers
---------------------------
>Martin Høyer Kristiansen (test@example.com) on 11/18/10 wrote:
>---------------------------
>>...And width of the registers. Intel switched back to read-after-schedule (used
>>in P4) for Sandy Bridge because supporting fat result buses would burn much more power.
>
>Read-after-schedule? What is that, a particular OoO trick?
It just means instructions schedule when data is ready. The scheduled instruction then reads source operand values from the physical register file (that's the PRF they talk about in the Sandy Bridge articles)
PPRO and derivatives up to and including Nehalem broadcast results in the ROB. Each ROB entry listens for the source operands needed. When both operands are ready the instruction can schedule, thus schedule-after-read. Broadcasting multiple wide register values to many ROB entries costs power.
Cheers