SECDED on header

[kgreigTT]

Hey @fischeti or @Lore0599 - I'm considering adding SECDED on the header. There are a few places that it could be done,

Option A: Hop-by-Hop

• Mechanism: Encode at Chimney, Check/Correct at every router hop
• Trade-off: Prevents misrouting immediately, but increases latency/area per hop

Option B: End-to-End

• Mechanism: Encode at Chimney, Check/Correct only at the destination interface
• Trade-off: Lower latency/area, but risks misrouting if address bits flip in transit (I think this is ok)

Challenges

• Timing: The logic depth might require an additional pipeline stage.
• Header Space: I need to verify if the current reserved bits can accommodate the ECC syndrome without widening the link

Has anyone in your group thought about this, or do you have any strong opinions/suggesetions on how this might be done

Cheers
Kevin

[fischeti]

Hey @kgreigTT,

We have had some discussions recently about this, and @AlaraWu is already actively working on reliability with support from @micprog.

Our current approach goes even a bit further than what you proposed and is a mix between Option A and B. The plan was to have SECDED for the payload which is checked end-to-end. The headers (which are smaller) would even use TMR in each router hop.

That is if I remember correctly, but I think @micprog can tell you more since he is the expert in this field.

Cheers,
Tim

[kgreigTT]

Thanks Tim, makes sense! If Alara is looking at this already I'll wait to see what they produce! Glad its being considered