[j-nsp] MX304 reliability

[Saku Ytti]

On Tue, 5 May 2026 at 22:05, Andrey Kostin <ankost at podolsk.ru> wrote:
>
> I can only guess that there isn't a problem with high failure rates as
> happened before with some linecards.
> Regarding Saku's note about pure single chip design, MX301 is a single
> Trio system, but it has the same number of ports available for network
> connectivity because Trio6 has two slices that act as independent
> forwarding engines, so theirs fabric lanes connect them together and no
> additional ports can be gained there.
> https://community.juniper.net/blogs/david-roy/2025/11/24/mx301-deepdive

Right, but the fabric connects to shared memory as far as I understand.

So technically we're looping the packets between package trio
instances back to the same memory, over fabric.

This seems eminently solvable problem in HW design phase, but it needs
someone to specifically want to add some additional feature of 'pick
up from memory, instead of fabric', which would be only useful for
these pizzabox designs. As in chassis design having local switching
between instances isn't useful enough to justify adding complexity.

I believe they are solving it though for next gen Trio, just not in
YT. So there may be some hope for real pizzabox design with WAN ports
on both sides.


-- 
  ++ytti