Are processor dies too "square"?

[imported_miahallen]

With the exception of the on die mem (ie L1 & L2) is there any other reason processor cores are rectangular? I was just thinking about everyone's roadmaps...Intel showing 100s or 1000s of cores per processor in thier 10 year projections, and then there's the cell with lots of cores per proc.

What about a hexagonal shaped core? They would be able to implement a bus system between the cores on more sides, more real-estate for inter-core communication, so-to-speak. Plus, if the real estate avalible was maximized, the per-wafer-yield would increase (not by much). Or how about 1 hex of L2 in the middle, with 3 or 6 hex procs around the sides, sharing the L2?

Is this possible, or are there higher level concepts I'm missing?

 

[Calin]

If you use a rectangular package for a processor, you are advantaged to use a rectangular die. Also, having a hexagonal die in a hexagonal package will decrease the perimeter of the die/package for a given surface. This is no good, as the pins on the processor are packed togheter really really close

 

[f95toli]

It is definitly possible but as Clin points out it makes packaging a bit more tricky.
In non-commerical designs we often use other shapes (the chips are always square, that is because we dice them with a saw).
I know a guy who succesfully used a flower-shaped design for a number of projects (he only needed a small munber of active elements, about 10 000 or so).

 

[harrkev]

Simple experiment. Get a piece of wood and a saw. See how well you can cut hexagons.

The reason (I suppose) that they use squares is that:

1) It is a no-brainer to cut.
2) No real advantage that I know of in using any other shape.

Whatever shape they use, it has to be able to tile a plane. Otherwise, you throw away silicon. This leaves the following shapes: square, triangle, hexagon. Of these, the square and triangle can be made by a series of cuts on the wafer going from end-to-end. A hexagon cannot be cut with a saw, but would have to be routed.

So, the choices are now triangle and square. The square seems easiest.

 

[Calin]

Triangles cannot be built in the same orientation (unlike rectangles or hexagons, that are just translated). Triangles must be mirrored. (even if I don't know exactly if this would be a problem)
Cutting a very thin sheet of silicon without cracking is hard enough even on continous lines, so indeed hexagonal shapes might be out of discussion

 

[ArjSiv]

Latency-wise, I'd think a sphere would be optimal, but manufacturing costs would be astronomical, not to mention how it would fit onto the motherboard.

I think eventually though, processors might be cubes if we go multi-core, which would simply be several wafers that are stuck together on top of eachother.

 

[f95toli]

A couple a days ago a saw a presentation at a conference where a guy from NTT showed a movie of a 60 micrometer globe of the earth made using 3D e-beam lithography (they had recorded a SEM session).
Everything was to scale and they could zoom in and see even rather small islands, it was easy to identify the countries. If I remember correctly the smallest details were just under 10 nm across. Total time for the lithography was 2 minutes.

So 3D lithography is definitly possible.

 

[SuperTool]

Because most EDA tools support 90 degree angles. Occasional tool supports 45 degree angles, and I can't think of a tool that supports 60 or 120 degree angles. There may be custom tools that do that on experimental basis, but you would be hard pressed to design anything complex with them.
So even if you make hexagonal die, you would then need hexagonal or triangular elements to put on that die. You would have no tool support for that, and on top of that, most modern manufacturing processes allow only one orientation of transistor per chip, which means practically speaking, you don't have much flexibility to rotate your blocks. So what you would end up is a hexagonal die with square elements on it, which means lots of wasted space.

 

[jagec]

Originally posted by: f95toli
A couple a days ago a saw a presentation at a conference where a guy from NTT showed a movie of a 60 micrometer globe of the earth made using 3D e-beam lithography (they had recorded a SEM session).
Everything was to scale and they could zoom in and see even rather small islands, it was easy to identify the countries. If I remember correctly the smallest details were just under 10 nm across. Total time for the lithography was 2 minutes.

So 3D lithography is definitly possible.

That is incredibly awesome.

 

[f95toli]

I found a link with pictures of the globe
http://www.ntt.co.jp/news/news04e/0402/040202_1.html

 

[soccerballtux]

Originally posted by: ArjSiv
Latency-wise, I'd think a sphere would be optimal, but manufacturing costs would be astronomical, not to mention how it would fit onto the motherboard.

I think eventually though, processors might be cubes if we go multi-core, which would simply be several wafers that are stuck together on top of eachother.

[trekkie with lisp]In season 4 episode 15 scene 3 take 5, Picard cleary shows why spherical shape must be used to overcome the light/processing differentiation. [/trekkie with lisp]

 

[white]

it's difficult to cut <100> silicon into hexagons. i think <111> silicon can be cut into that shape because of its symmetry, but one reason <100> silicon is chosen because you can grow a higher quality oxide on it than on <111> silicon.

if you take a <100> wafer and shatter it with a point in the middle, it'll split into four pieces (90 deg angles) from your point. simply put, it comes from the atomic symmetry.

 

[gbuskirk]

Also, routing on adjacent layers of a die is often kept perpendicular to minimize crosstalk between layers. This leads to orthogonality, which leads to square elements and square die.