melnary has quit [Remote host closed the connection]
melnary has joined #prjcombine
_whitelogger_ has quit [Remote host closed the connection]
_whitelogger_ has joined #prjcombine
_whitelogger_ has quit [Remote host closed the connection]
_whitelogger_ has joined #prjcombine
_whitelogger_ has quit [Remote host closed the connection]
_whitelogger_ has joined #prjcombine
_whitelogger has joined #prjcombine
_whitelogger_ has quit [Remote host closed the connection]
_whitelogger_ has joined #prjcombine
<Wanda[cis]> I have a possibly stupid question
<Wanda[cis]> is it like common practice to manufacture both a die and a perfect mirror image of a die? like literally flip the X coord on all masks
<Wanda[cis]> for the purpose of multi-chip modules where it'd convenient to have the chips touching by the common edge
<whitequark[cis]> do you mean flipping the mask?
<Wanda[cis]> yeah?
<Wanda[cis]> I'm looking at xcvp1902 geometry
<Wanda[cis]> it's a 4-die device, and it has a very unusual layout
<Wanda[cis]> all other multi-die devices just have the individual die stacked in a column bottom to top
<Wanda[cis]> while vp1902 has 4 identical die arranged in a mirrored square
<Wanda[cis]> well "identical" of course two of them would be flipped
<whitequark[cis]> which ones?
<whitequark[cis]> like the two right ones?
<Wanda[cis]> I um
<Wanda[cis]> imagine GL_MIRROR_CLAMP_TO_EDGE
<Wanda[cis]> the left top and bottom right one would effectively require flipped masks
<Wanda[cis]> (left bottom one being the "base", and top right being just rotated 180°)
<whitequark[cis]> Wanda[cis]: right, that was my thinking
<whitequark[cis]> * my thinking when asking the question
<azonenberg> Wanda[cis]: i cant say it's *common*
<azonenberg> but if you are able to fit two dies on a single physical mask
<azonenberg> i see no reason why you couldn't mirror one copy of the circuit
<azonenberg> basically a MPW with two mirror image layouts
<azonenberg> i've never heard of that actually being done though
<Wanda[cis]> that's like the literal largest die to date though
<Wanda[cis]> anyway
<Wanda[cis]> I'm wondering if the vp1902 die actually come in two enatiomers, or if the geometry database is just lying and the interposer die papers over it somehow
<azonenberg> Good question
<azonenberg> my gut feeling is that they're not physically mirrored
<Wanda[cis]> so is mine
<Wanda[cis]> it'd be strange
<azonenberg> but there's no way they'd have connections from the right edge of the left die to the right edge of the right
<azonenberg> just to simulate mirroring it
<Wanda[cis]> I'm not so sure about it
<Wanda[cis]> the SLL connections aren't concentrated on edges btw, they're kinda... spread across the whole die, I think
<Wanda[cis]> (I haven't yet written the code to properly extract the cross-die routing, but that's what I figured from some manual poking at random wires in Vivado)
<whitequark[cis]> i've always wondered how do Laguna connections work
<whitequark[cis]> like on a physical level
<Wanda[cis]> "microbumps"
<Wanda[cis]> apparently the laguna tiles just have smaller-than usual pads on the top metal layer, which is then facing the interposer die, which... I think has through-silicon vias at these points?
<whitequark[cis]> so they're like a bonding ball sandwich?
<Wanda[cis]> mhm
<Wanda[cis]> that's what the xilinx drawings imply, too
<Wanda[cis]> and the balls are spread over the 16 or so laguna columns in the top and bottom clock regions of the die (they displace CLEM columns in these regions only)
<Wanda[cis]> this is for ultrascale/ultrascaleplus; versal has uh. every CLB is a laguna connection point. 6 balls per CLB or so.
<Wanda[cis]> and the actual connections between die seemed much less regular than ultrascale too
<whitequark[cis]> wow
<Wanda[cis]> (virtex7, otoh, is very crude and has no laguna tiles or even dedicated buffers for cross-die connections; instead they did the bare minimum modification of the interconnect tiles in top and bottom clock regions, and just fucking connected the middle-points of all long lines via super-long-lines)
<azonenberg> i havent worked with xc7v but it makes sense it'd be the first generation of the tech
<azonenberg> i did some work on the vu9p a while back
<Wanda[cis]> aside of the main laguna / interconnect connections, you also have connections on the clock routing tracks
<azonenberg> yeah i wasnt sure how htose worked
<whitequark[cis]> Wanda[cis]: ohhh.
<azonenberg> but i do remember having to worry about routing around the laguna tiles launching to the adjacent SLR
<Wanda[cis]> for virtex7 it's simple: 32 balls for the 32 GCLK spines (which can be connected between die or split as desired)
<Wanda[cis]> ultrascale has distributed clock routing (1 to 4 vertical routing tracks per column), so you just have a few extra balls here and there
<Wanda[cis]> okay dealing with ultrascale was slow but dealing with versal makes me fantasize about ln2
<azonenberg> lol
<whitequark[cis]> ln2?
<whitequark[cis]> oh
<whitequark[cis]> overclocking