Comments Locked

28 Comments

Back to Article

  • imaheadcase - Thursday, October 11, 2018 - link

    I'm thinking out loud here, but does anyone else thing CPU are going to venture away from square designs. I mean i read a long time ago that humans just do that so things fit into other things, but its more efficient to make hardware differently...just standards make it so they can't do own thing.
  • A5 - Thursday, October 11, 2018 - link

    All of the wafer tools are designed for rectangular dies, so until that changes it probably doesn't make any sense to move away from rectangular/square packaging.
  • PGFan - Thursday, October 11, 2018 - link

    The squarish designs also maximize the surface area for pins. A long time ago, chips that used DIP were rectangular, but were also limited on the number of pins because of the shape (and the pins were wide mounted and not bottom-mounted). Think 8080-80286. My last DIP-based CPU, if I recall correctly, was a V20 replacement for a 8088.
  • sing_electric - Thursday, October 11, 2018 - link

    In theory, if you wanted to maximize space while minimizing surface area for a chip (say, on a motherboard), round would be best. (It also doesn't hurt that fans have to be round, and every desktop processor is dependent on active cooling, typically fans.)

    However, two main issues: In the end, the question is how many chips you can fit on a wafer, since that's the biggest driver of cost. Putting a bunch of circles next to each other would result in a LOT of dead space. There might be some kind of polygonal array that'd theoretically minimize wasted area on the outside, but it would be dependent on chip and wafer size.

    Given some chips have MUCH bigger dies but fit into the same package (e.g. a quad core i5 and an 8 core i9 use the same socket), it'd be very hard to plan that out.

    Second, knowing how to orient a square processor is easy: Just mark a corner. You'd have to get the rotation of a round processor EXACTLY right, which would probably be way trickier (there might be some fancy way of mounting a round processor, like how lenses are "twisted" in to the socket, at which point the pins line up, except I don't think that'd scale well to over 1,000 tiny pins - I think mechanical breakage would be a significant concern).

    Oh, and also: As it stands, motherboard makers don't seem to have to remove a ton of features to accommodate a larger processor, for example, how an ATX Ryzen and Threadripper motherboard might have the same # of PCI-X, RAM, SATA, etc. slots.
  • willis936 - Thursday, October 11, 2018 - link

    Hexagon is 2D closest packed. Circles have many, many other issues with them. Hexagon only has 50% more sides than a square and could potentially net quite a few more chips from a die. Also the maximum variance in packaging trace length would go way down as well.
  • 0ldman79 - Thursday, October 11, 2018 - link

    True, but you're talking die and he's talking package.

    Also they have a bit of dead space in the wafers and the die. It's not like there are transistors all the way to the edge of the chip.

    Having multiple edges to avoid would probably actually cost them space vs a rectangle or square die.
  • willis936 - Thursday, October 11, 2018 - link

    I was talking about both die and package. Hexagonal die would increase yields. Having both a hexagonal die and package would make package trace length shorter and have less variance between pins. A mathematic proof of this would be fun to do but it's easiest to convey the idea by imagining the worst case traces: the corners. The lower the outer angle of the corner, the shorter the trace has to be. A circle would be best where the outer corner angle is 180 degrees, but a hexagon's 240 degrees is much better than a rectangle's 270 degrees.

    I don't know if traces are the limiting factor in packaging size but it would help signal integrity at least.
  • thetuna - Thursday, October 11, 2018 - link

    A hexagonal layout would make "dicing" the chips literally impossible :)
  • woggs - Thursday, October 11, 2018 - link

    Bingo.
  • rahvin - Thursday, October 11, 2018 - link

    So according to the wiki article on wafer dicing says that dicing other shapes is possible as long as they are using full cut laser which the article implies isn't as common as other methods. It doesn't say a hexgon is possible but if they are using computer controlled full depth laser cuts it would seem possible to get pretty much any shape.

    So accordingly rectangular dies aren't just because of dicing, there are lots of factors I'm sure that drive using rectangular dies. Things like simply history driving the tooling to all be geared towards rectangular dies because the early dicing techniques (physically cutting) were incapable of anything but straight lines all the way across the wafer. And simple things like the robotic picking and packaging is simpler with rectangular layouts.

    I bet when you pile up all these reasons anything but rectangular just doesn't make sense.
  • 0ldman79 - Thursday, October 11, 2018 - link

    They also can't just put stuff where ever they want. The CPU are still somewhat "modular" in design in that the ALU is here and this shape, the FPU is here, the L1, L2, L3, GPU, etc...

    To change the shape of that, which is currently rectangular, would require a full redesign of the chip. They're actually designing these things accounting for the time it takes for the electrons to make the trace to the next transistor/gate/whatever.
  • willis936 - Friday, October 12, 2018 - link

    Layout isn't an easy problem even when you're assuming a rectangular die. You can make all of your islands rectangles but there are some good numberphile videos about math games that highlight just how few combinations of rectangles can be made into a larger rectangle. Designers still have to squish dimensions and make compromises to fit everything in. So a hexagon would create a different set of challenges but just because the valid layouts are less intuitive to a human I'm not sure their would be inherently fewer good layout choices. The biggest challenge I can think of is how to make a rectangular transistor make a non-90 degree mating point with another transistor. I'm imagining something like a nearest neighbor interpolation of a slanted line (think a graduated step response like unfilitered PCM).
  • woggs - Thursday, October 11, 2018 - link

    Wafers are cut into chips using a saw or laser scribe and then snapped along straight lines, like cutting glass. Straight lines across the wafer... No other option makes any sense at all.
  • willis936 - Thursday, October 11, 2018 - link

    Lasers and water jers do not need to cut in a straight line. Water jets are not used for obvious reasons. Hexagonal cutting patterns would certainly add manufacturing time because of the extra cutting time but I think if design and manufacture concerns were ignored there would be mild performance gains when we simply don’t have any more bandwidth left in existing electrical channels.
  • Qiasfah - Thursday, October 11, 2018 - link

    You don't roll for initiative with a d12, plz
  • CaedenV - Thursday, October 11, 2018 - link

    lol
    I think they are rolling for damage
  • PeachNCream - Thursday, October 11, 2018 - link

    Was gonna say that someone hasn't played a D20 based RPG. Though there are other systems that don't put the D20 to use like Savage Worlds where the largest die you toss is a D12 and combat initiative is determined by drawing cards from a poker deck. Then there's Shadowrun which is infamous for requiring a bucket of D6s. Eh, it was still fun joke even if the D12 isn't commonly used for initiative.
  • mode_13h - Thursday, October 11, 2018 - link

    Yeah, but it was a good turn of phrase. I think we can probably overlook that detail.
  • eastcoast_pete - Thursday, October 11, 2018 - link

    @Ian: Thanks! Any updates/news/rumours on availability of the i9 processors in the next 3 months? Nice that they ship review samples, but we all know that those are from early low number/yield runs. Intel's fabs are currently overwhelmed, after their planned move to a 10 nm process was postponed or cancelled, leaving their working 14 nm FF++ fabs unable to cope with demand. Knowing how soon i9s will ship in quantity and thus sell at sane prices might make a difference if building a system. Thanks!
  • Ian Cutress - Friday, October 12, 2018 - link

    Review CPUs are from the final batch of ES. Intel has said that they'll meet demand for 2018 as per their financial report, and that going forward their focus is on Core and Xeon over other segments.
  • CaedenV - Thursday, October 11, 2018 - link

    Where was this 5-sided packaging when the Pentium chip was a thing?
  • 0ldman79 - Thursday, October 11, 2018 - link

    lol

    Seriously.

    Intel missed the boat on that one.
  • Mr Perfect - Thursday, October 11, 2018 - link

    If Intel is only sampling the i9, then will there be no review of the i7? I'm curios to see if 8c8t is faster then 6c12t.
  • Ian Cutress - Friday, October 12, 2018 - link

    We're trying to get one.
  • brunis.dk - Thursday, October 11, 2018 - link

    Look at Intel sucking up to you, while they can't deliver an i5 K from the last 3 generations. They are like a teenage boy still wet behind the ears and AMD just blew them a kiss.
  • mode_13h - Thursday, October 11, 2018 - link

    What's up with all the pentagons?

    And why does the IHS say 3.60 GHz?
  • Ian Cutress - Friday, October 12, 2018 - link

    3.60 GHz is base frequency and where 95W TDP comes from.
  • tamalero - Monday, October 15, 2018 - link

    Do these chips that were customized for reviewers.. could they be binned?

Log in

Don't have an account? Sign up now