By: anon2 (anon.delete@this.anon.com), September 16, 2022 5:45 pm
Room: Moderated Discussions
groo (charlie.delete@this.semiaccurate.com) on September 16, 2022 11:06 am wrote:
> anon2 (anon.delete@this.anon.com) on September 16, 2022 9:00 am wrote:
> > Groo (charlie.delete@this.semiaccurate.com) on September 15, 2022 11:46 pm wrote:
> > > anon2 (anon.delete@this.anon.com) on September 15, 2022 7:04 pm wrote:
> > >
> > > > What is expensive about L1 ECC which is less costly in L2?
> > >
> > > Time.
> >
> > On which side?
>
> It takes time to compute ECC data and L1s tend to be a bit tight on timings.
>
Nah. On the generation side, ECC could queue and store into another array at longer latency, on the checking side, you only need to check parity. The exceedingly rare parity failure would flush the pipe, drain the ECC queue, and go from there.
That's close to how you can look at a store-through L2 ECC design of today (you didn't think L1D loads were going to the L2 to check parity?): instead of L2 holding data+ECC, there is an array that just holds ECC. Still has the benefit of write-back L1 because stores can be released while ECC updates are in flight.
> anon2 (anon.delete@this.anon.com) on September 16, 2022 9:00 am wrote:
> > Groo (charlie.delete@this.semiaccurate.com) on September 15, 2022 11:46 pm wrote:
> > > anon2 (anon.delete@this.anon.com) on September 15, 2022 7:04 pm wrote:
> > >
> > > > What is expensive about L1 ECC which is less costly in L2?
> > >
> > > Time.
> >
> > On which side?
>
> It takes time to compute ECC data and L1s tend to be a bit tight on timings.
>
Nah. On the generation side, ECC could queue and store into another array at longer latency, on the checking side, you only need to check parity. The exceedingly rare parity failure would flush the pipe, drain the ECC queue, and go from there.
That's close to how you can look at a store-through L2 ECC design of today (you didn't think L1D loads were going to the L2 to check parity?): instead of L2 holding data+ECC, there is an array that just holds ECC. Still has the benefit of write-back L1 because stores can be released while ECC updates are in flight.
